IJDIWC

IJDIWC
International Journal of
ISSN 2225-658X (Online)
DIGITAL INFORMATION AND WIRELESS COMMUNICATIONS
Volume 4 Issue 4
2014
TABLE OF CONTENTS
Original Articles
PAPER TITLE
AUTHORS
Imran Jattala, Junaid Farooqi,
WIRELESS SENSOR NETWORK (WSN) BASED AUTOMATIC FIRING
Shakeel Durrani, Nassar Ikram,
PRACTICE SYSTEM (AFPS) FOR TRAINING OF LAW ENFORCEMENT
Irfan Hussian, Muazzam A.
AGENCIES (LEAS)
Goraya, Obaid B. Zakria
INFORMATION TRANSPARENCY IN ADMISSION SELECTION OF
STATE ISLAMIC HIGHER EDUCATION IN INDONESIA USING Linda Salma Angreani
INFORMATION TECHNOLOGY APPROACH
HYBRID BOUNDARY DETECTION METHOD FOR IMAGE WITH Syaiful Anam,
APPLICATION TO CORONARY PLAQUE
Noriaki Suetake
Eiji
Atef
408
418
Uchino,
A MULTI TREE BASED APPROACH FOR PERFORMANCE ANALYSIS IN

Itu Snigdh, Swarn Saurabh
HIERARCHICAL WIRELESS SENSOR NETWORKS
RELIABILITY AND ACCURACY OF NEURAL NETWORKS FOR David Semaan,
EXCHANGE RATE
Abdallah Kassem
PAGES
428
438
Harb,
447
ANALYTICAL APPROACH OF COST-EFFECTIVE AND SECURE SIPChulhee Cho, Jae-Young Choi,

BASED MOBILITY MANAGEMENT SCHEME FOR NEMO
Jun-Dong Cho, Jongpil Jeong
ENVIRONMENTS
458
FALL DETECTION SYSTEM DESIGN BY SMART PHONE
474
Yung-Gi Wu, Sheng-Lun Tsai
PERSIAN/ARABIC HANDWRITTEN DIGIT RECOGNITION USING Mohsen Biglari, Faezeh Mirzaei,

LOCAL BINARY PATTERN
Jalil Ghavidel Neycharan
486
Shibarchi Majumder, Rahul

Shankar, Mani Shankar Prasad
493
NAVIGATION ERROR REDUCTION IN SWARM OF UAVS
International Journal of Digital Information and Wireless Communications (IJDIWC) 4(4): 408-417
The Society of Digital Information and Wireless Communications, 2014 (ISSN: 2225-658X)
Wireless Sensor Network (WSN) based Automatic Firing Practice System (AFPS) for
Training of Law Enforcement Agencies (LEAs)
Imran Jattala, Junaid Farooqi, Shakeel Durrani, Nassar Ikram
Horizon Technologies
Islamabad, Pakistan
imran.jattala@gmail.com, shakeel.durranii@gmail.com, junid.farooqi@gmail.com,
dr_nassar_ikram@yahoo.com
Irfan Hussian, Muazzam A. Goraya, Obaid B. Zakria
National University of Science & Technology (NUST)
Islamabad, Pakistan
irfanmechatronics@gmail.com, dmagoraya@hotmail.com, obaid2002@yahoo.com
ABSTRACT
The critical importance of an efficient infantryman in
special operations force, tactical paramilitary and Law
Enforcement Agencies (LEAs) is insurmountable. One
of the many vital aspects of an effective solider is
excellent marksmanship which requires extensive
training at sophisticated firing ranges. Modern firing
ranges are supported by Automatic Firing Practice
Systems (AFPS) and this paper presents the design and
development of such a system based on WSN. AFPS
provide an automatic bullet-impact count during firing
training session and is modular scalable in design for
multiple of eight concurrent shooters. The system is
versatile and flexible allowing for different small-arms
and firing training modes and supports night firing
exercise. AFPS comprises of two major components,
the automatic target-box and a commander console.
Automatic target-box has a motor & gear assembly,
target sheet, bullet-impact sensor, control board, and
WiFi communication module. Commander console is a
ruggedized sunlight readable 10.4 Tablet PC, which
with a built-in WiFi acts an access point. The
automatic target-boxes equipped with embedded WiFi
modules form sensor nodes of a WSN. The paper
presents the complete System Development Life Cycle
(SDLC) of the firing practice system and associated
WSN. The AFPS and bullet-impact sensor was
extensively tested on Firing Ranges for accuracy of
bullet-impact count. The results showed a bulletimpact count accuracy of over 97 percent.
KEYWORDS
WSN, WiFi, Bullet-impact Sensor,
Console, Firing Practice System
INTRODUCTION
The infantryman still plays a critical role in the
outcome of battles in modern warfare especially
with the advent of weapons of mass destruction
(WMD) where wars between equally adversaries
are high improbably. Proficiency in effective
shooting and excellent marksmanship is one of the
many vital characters of a well-disciplined and
efficient solider. Instilling this character into new
recruits require the need of dynamic shooting
ranges for training of small-arms Law
Enforcement Agencies (LEAs), paramilitary and
military [1]. Firing practice systems used in
developing countries are predominantly manual or
semi-automatic systems [2]. These semi-automatic
systems have inaccurate bullet-impact count and
had to be manually and tediously counted by the
on-duty staff. Automated shooting practice
systems developed by leading defense companies
are cost intensive and are beyond financial
budgets estimate of most of LEAs in the
developing world [3]. The proposed automatic
firing practice system has been developed to
achieve the following objectives:
Commander
WSN based AFPS provides instant bullet

impact count update for each target plate,
Provisioning of different firing training &
exercise modes,
Comprehensive training scenarios for LEAs,
Para-military and military organization,
Firing results database (FRD) for data analysis
and provision of trend and statistical analysis.
408
Commander Console acts as an access point (AP)
while 24 Field Boxes (FBs) form WiFi sensor
nodes for AFPS-WSN.
This paper presents the complete System
Development Life Cycle (SDLC) of the
Automated Firing Practice System (AFPS) from
concept to development and manufacturing. The
paper is divided into four parts, description of
AFPS, development of Target box assembly,
commander console (CCon) & WSN, and finally
the on-site test results of AFPS.
10
11 12
DESCRIPTION OF AFPS
Automatic Firing Practice System (AFPS)
provides recruits of LEAs with better training
environment and achieves a higher degree of skill
more rapidly. Since AFPS is reconfigurable to
different shooting training module that can be
incorporated via software from the CCon, AFPS
has a robust structure that allows it to adapt to
different terrains and environments. The
architecture of AFPS is shown in Fig. 1.
13
14 15
16
17
18
19
20 21
22
23 24
Target Boxes with

WiFi Sensor Nodes
WiFi Antenna
Hi-Gain 14 dBi Outdoor

Directional Antenna
Commander Console
Figure 1: Architecture of AFPS
The main features of AFPS are accurate bulletimpact count, rejection of false-impact (a ricochet
or stone), wireless connectivity and a ruggedized
CCon. The main modules of AFPS are defined as:
capability. Target boxes were Powder Coated with

automotive grade paint for all-weather operational
requirements. The target boxes were made IP67
compliant [4].
1.1 Target Box Assembly

Target boxes were designed for a ruggedized
deployment. Aluminum boxes were casted &
machined to develop boxes that housed motor,
gear assembly, target-sheet frame, counterweights, limit switches and linkages. LED light
was installed on the box for night firing mode
1.2 Commander Console

Commander console for AFPS was designed based
on a ruggedized 10.4 Tablet PC, with a WiFi Link
to 24 Target box units to form a WSN. A WiFi
local area network was established at the firing
range, between CCon and target units within a
range of 1000 meters with clear line-of-sight [5].
409
The target units connect to the commander and

provide update on their current status. All the
target units are synchronized to same starting
position before firing practice.
DESIGN OF TARGET BOX ASSEMBLY
The Target box assembly comprised of a targetsheet & frame with bullet-impact sensor attached
and a target-box housing motor gear assembly & a
control board. The WiFi communication module is
also placed inside is target-box.
1.3 Mechanical Designing of Target Box
The target plate and enclosure was designed and
modeled using Pro-E, market leader CAD/CAM
software [6]. The designs were evaluated and
revised/updated recursively to meet the design
criteria. Specialized gears were used that had
heat-treated long-life capability, having less
wear and tear prolonging its service life for a
specific duty cycle. Lithium grease was used to
ensure smooth operation of gears [7]. The AFPS
enclosure was properly sealed to ensure
ruggedness. The final Pro-E design is shown in
Fig. 2.
1.4 Target Box Assembly

Target box assembly was designed & developed
considering the harsh environmental deployments.
AFPS is to be operated in open-air firing range
under all weather conditions. The components
selected or developed are:
Figure 2: Automatic Target Box
Bullet-Impact Sensor: The most critical

module within the AFPS is the Bullet Impact
Sensor (BIS), based on an piezoelectric
Vibration Sensor PKS1-4A1 from OEM
Murata was used in AFPS [8].
DC Motor: +24V High Torque Motor was
selected from a local source [9].
Gear Assembly: Worm-gear assembly was
designed to provide high torque [10].
Switches & Connectors: Limit-switches were
installed to provide motor over-shoot
protection. Ruggedized mil specs circular
connectors were selected for the WiFi Antenna
& Power cable connections [11].
LED Light: +12V LED light was installed to
enable night firing mode operation of AFPS
[12].
Control Board: Piezoelectric Vibration Sensor
(PVS) PKS1-4A1 from Murata was interfaced
to a microcontroller AT89c2051 based control
board [13]. BJTs & FETs based motor driver
was designed to drive the DC motor [14].
Limit switches were interfaced to the
microcontroller to provide safety for
motor/target-sheet frame over-drive. The
control board has its own SDLC which
included developing the microcontroller
firmware, design and fabrication of the PCB,
procurement of Bill-of-Materials (BOMs),
stuffing of the PCB, its quality assurance and
finally its functionality testing.
Target Sheet: Different materials of Target
Sheet were assessed and tested for hardness
and elasticity at firing ranges for endurance
testing. The repairing methods were also
evaluated for ease of maintenance and care.
Fiber glass was finally selected as the material
for target sheets, other advantages also include
cost effectives and easy availability thought
out Pakistan [15].
410
The finalized environmental tested and the final
target box assembly is shown in Fig. 3.
Figure 3: Target Box Assembly
Target box with mil specs circular connectors is

shown in Fig. 4.
Figure 4: Circular Connectors for AFPS Target Box
1.5 Bullet-Impact Sensor

A special enclosure was designed for Piezoelectric
Vibration Sensor (PVS) PKS1-4A1. The bottom of
the target sheet where the PVS is located was

covered with additional armor plate, to ensure the
safety of the sensor. The sensor was placed
longitudinal to axis of shock wave propagation.
The placement of Bullet-Impact sensor is shown in
Fig. 3.
Figure 5: Bullet-Impact Sensor Mechanical Placement
Target enclosure has a base-box and a target sheet.

The shape used for Target Plate (TP) is a standard
humanoid figure used commonly at Fire Ranges
[16]. The height of the TP used for the AFPS is 42
inches and was developed using Fiber-Glass. The
TP fitted with Piezoelectric Vibration Sensor
PKS1-4A1 was then subjected to rigorous firings
in order to record the bullet-impact responses.
Initial results were subjected to detailed
statistical signal analysis and a subsequently an
algorithm was developed in Matlab to distinguish
between impact signals of the bullet-impact, other
objects impact and bullet-ricochet signals [17]. The
PVS interface circuit was revised for improvement
and with the addition of with an installation of
noise filter, and with an implementation of
hardware and software de-bouncers, better results
were achieved, thus PVS sensor was finalized for
the AFPS project.
The sensor installed on the AFPS is
Piezoelectric Vibration Sensor (PVS) and was
extensively tested for various types of shocks and
vibrations [18]. An amplifier circuit was designed
for PVS to provide a suitable gain to its signal.
PVS was attached to a target sheet and live rounds
411
were used for testing. Amplified PVS signals were

recorded on a digital oscilloscope for various
ammunition types and provide calibration
benchmarks, for e.g. an amplified bullet-impact
response is shown in Fig. 6.
Figure 6: Amplified Response of Bullet-Impact Sensor
The finalized system was installed at 10 different

places across Pakistan and bullet-impact results for
10,000 bullets were collected. The responses were
of five different types of ammunitions, fired from a
distance of 200 to 500 meters. The results were
once again subjected to detailed statistical analysis
and minor adjustments were made in the
algorithm.
The bullet-impact detection algorithm was finetuned and was tested on the recorded 10, 000
bullet-impacts. The results of the Matlab bulletimpact sensing algorithm showed accuracy over of
97 percent.
1.6 Control Board
An 8051 microcontroller based control board was
designed for AFPS. The control board had a Bullet
Impact Sensor Interface, Motor Driver Circuit, and
Light Control Circuits. The Proteus ISIS [19]
design of Control Board is shown in Fig. 7.
Figure 7: Control Board Circuit Design
The main parts of Control board circuit are:

Bullet Impact Sensor Interface: Piezoelectric
Vibration Sensor PKS1-4A1 was interfaced
with the Control Board
Bullet Impact Counter: Analog comparator

based bullet-impact counter circuit was
designed. The design of Bullet Impact Counter
was revised over recursive design iterations.
412
Motor Driver: IRF540 Power MOSFETs were

used to drive the DC motor [20].
Light Control: IRF540 Power MOSFET was
used for LED light control.
1.7 WiFi Communication Module

Serial to WiFi module NPort W2150A by Moxa
with WiFi antenna was used for Target side
communication [21]. The W2150A effectively
forms sensor node part of WSN. W2150A provides
a RS232-to-WiFi interface enabling the Control
Board microcontroller to CCon over WiFi shown
in Fig. 8.
Figure 8: Moxa WiFi Communication Module
DESIGN OF COMMANDER CONSOLE

Ruggedized Tablet PC MARS-3100R from
Advantech was used with sunlight readable LCD
and back-up battery [22]. CCon Firmware was
developed in Labview with Microsoft SQL Server
database. Wireless capability is introduced with a
WiFi Antenna attached to an Antenna-mast also
acting as the Receiver Station [23]. Customized
and proprietary communication Protocol for
commander and target communication was
designed & developed. The WSN formed with
CCon & 24 Target units provides instant bulletimpact count and provides a dynamic deployment
advantage to the AFPS.
1.8 Graphical User Interface of Commander
Console
CCon software is primarily used for conducting
firing practice and monitoring shooter score. The
GUI is designed in Labview [24]. The software is
equipped with a Microsoft SQL Database to
provide the capability of detailed firing practices
results storage and analysis. The layout of CCon
GUI is shown in Fig. 9.
Figure 9: Commander Console Labview GUI
413
The main features of CCon software are:
The software/GUI is divided into eight subareas. Each sub-area represents each target unit
A bullet-impact is immediately reported to the
CCon and shooter score is update on the GUI
in real-time
GUI shows individual score of each target plate
AFPS CCon supports night firing mode.
The CCon is 10.4 ruggedized tablet MARS3100R by Advantech and provides portability to
the commander. CCon has a supporting stand to
place it on table with a USB keyboard and mouse
for conventional Desktop like usage. The
commander console is shown in Fig. 10.
Figure 10: AFPS Commander Console
1.9
WiFi Communication Protocol
AFPS WiFi Communication Protocol is defined as:
Configuration: CCon assign IDs to address of
targets a, b, c w, for 24 Targets
Get Target Status: Commander reads the status
of Target
Set Target Status: Commander sets the current
status of Target to Up or Down with LED
Light On or Off
Get Bullet-Impact: Target immediately reports
to Commander the Bullet Impact
Bullet-Impact
Acknowledgement:
Commanders Acknowledgement of the
reception of Bullet-Impact
Hit Down Mode: Command to put Target in
Hit Down Firing Practice Mode. Target will go
down on every hit and will remain in such
position until command for popup is issued
Pop Up Mode: Command to put Target in Popup Firing Practice Mode. The target will go
down on each bullet hit but popup again

automatically and immediately
Still Mode: Command to put Target in Still
Firing Practice Mode. Target will remain static
even when hit
Time Mode: Command to put Target in Time
Firing Practice Mode. The target should
appear 5 times for 6 seconds at intervals of 5
or 6 seconds
Command Acknowledgement: Commands
reception acknowledgement sent by Targets to
Commander Console
Communication between CCon and Targets is
secured through built-in security features of WiFi
like WEP, WPA, and WPA2.
1.10 +24V Regulated Power Supply
Power requirement for 24 AFPS units was
calculated. +24 Volt Regulated Power Supply was
designed & developed for AFPS units. Power
requirement for AFPS Target Boxes (with WiFi
Module) was calculated to be 5A at +24V. A +24
Volt, 150A regulated DC Power Supply was
designed and developed to power-up 24 AFPS
Target Boxes.
AFPS Testing & Results
The preliminary field trial of AFPS showed overcounting of the bullet-impacts. The reason for
over-count was identified and the bullet-impact
circuit was critically analyzed. The bullet-impact
signal was producing sustained vibrations which
were causing a false-count. The design of bulletimpact counter circuit was reviewed and debouncer was added both in hardware and software.
The bullet-impact over-count & Corrected signals
are shown in Fig. 11.
Figure 11: Bullet-Impact Over-Count & Corrected Signal
A testing methodology was established in which

the performance of the AFPS would be critically
analyzed during the live testing phase and error
corrections will be incorporated for the
improvement of the final system. A complete setup
414
of AFPS was assembled at the firing range with a

CCon and three target units.
In the first phase of the live firing test, on a clear
day, 100 bullets were fired on each of the three
target units, from a distance of 200 meters. First
firing test produced an average accuracy of 82.5%.
First firing results are shown in Fig. 12.
To further improve the efficiency of the AFPS,

temperature compensation was added to the bulletimpact sensing circuit. Third firing test showed a
greater improvement with an average accuracy of
94.7%. The results of third field firing are shown
in Fig. 14.
Figure 14: AFPS - Third Field Testing Results
Figure 12: AFPS - First Field Testing Results
Bullet-Impact sensing circuit was improved

after the first trail. During the second phase of the
live firing test, on a clear day, 100 bullets were
fired again on each of the three target units, from a
distance of 200 meters. Second firing test produced
an average accuracy of 92.1%. Second firing
results are shown in Fig. 13.
A software de-bouncer was added in the bulletimpact detection logic. Software de-bouncer
brought a significant improvement further to the
system. Fourth firing test showed even greater
improvement with an average accuracy of 97.6%.
Results of fourth field firing of AFPS are shown in
Fig. 15.
Figure 15: AFPS - Fourth Field Testing Results
Figure 13: AFPS - Second Field Testing Results
The four phases of AFPS field testings and

improvements made on the basis of the results
obtained from these field testings brought the
415
average Software Count Accuracy to a astonishing
figure of 97.6%. The overall achievement from a
82.5% to 97.6% mark shows the success of the
development procedure and cycle. The graph of
average Software Count Accuracy percentage is
shown in Fig. 16.
Figure 16: AFPS Average Software Count Accuracy

Percentage
CONCLUSIONS
A firing practice system has been designed and
developed to automate the existing manual or
semi-automatic firing ranges. WSN based AFPS
provided real-time bullet-impact count wirelessly
adding a greater level of flexibility for the
Commander for conduct of dynamic firing exercise
scenarios. The complete SDLC of the system has
been presented including mechanical and electrical
design of different parts of AFPS with revision and
improvement through different development
stages. The improvement process continued during
the field trials phase as well, till a 97.6% Bulletimpact count accuracy was achieved. The system is
controlled through a wireless (WiFi) commander
console. The power to the target units is powered
by a DC power supply. The next version of the
AFPS will be battery-powered, wirelessly
controlled, completely portable system for versatile
shooting practice scenarios.
REFERENCES
[1] S. Li, "The Design and Implementation of Shooting
System
Simulation
Platform
for
Police
College," International
Conference
on Scalable
Computing and Communications; Eighth International
Conference on Embedded Computing, 2009.
SCALCOM-EMBEDDEDCOM'09., pp.566-570, 25-27

Sep. 2009.
[2] F. Ali, and A.B. Mansoor, "Computer Vision based
Automatic Scoring of shooting targets," IEEE
International Multitopic Conference, 2008. INMIC
2008, pp.515-519, 23-24 Dec. 2008.
[3] J. Sun, F. Zhang, Y. Li, and S. Wang, "Research on key
technologies
in
automatic
targetreporting," International Conference on Automatic
Control and Artificial Intelligence (ACAI 2012),
pp.1732-1735, 3-5 Mar. 2012.
[4] R. Ahiska, H. Mamur, and M. Ulis, "The design and
application of a LED driver and a LED armature with a
thermoelectric
generator," Fourth
International
Conference on Power Engineering, Energy and
Electrical Drives (POWERENG), 2013, pp.902-905, 1317 May 2013.
[5] A.K.M.M. Islam, A. Zeb, and K. Wada,
"Communication protocols on dynamic cluster-based
wireless sensor network," International Conference
on Informatics, Electronics & Vision (ICIEV), 2013,
pp.1-6, 17-18 May 2013.
[6] L. Yongfeng, T. Hongsen, Q. Tao, and L. Lianda,
"Research on 3-D computer aided design of rapid
manufacture for the movement system in diesel engine
based on PRO/E," International Conference on
Mechanic Automation and Control Engineering
(MACE), 2010, pp.589-592, 26-28 Jun. 2010.
[7] J. Bryan, "Reducing downtime by proper motor
lubrication," Conference Record of 2013 Annual
IEEE Pulp and Paper Industry Technical Conference
(PPIC), pp.50-62, 23-27 Jun. 2013.
[8] M. Dubois, M. Farhat, E. Bossy, S. Enoch, S.
Guenneau, and P. Sebbah, "Flat lens for pulse focusing
of elastic waves in thin plates," Applied Physics Letters ,
vol.103, no.7, pp.071915-071915-4, Aug 2013.
[9] L. Zi-Yi, W. Sew-Kin, P. Wai-Leong, and O. Chee-Pun,
"The design of DC motor driver for solar tracking
applications," 10th IEEE International Conference
on Semiconductor Electronics (ICSE), 2012, pp.556559, 19-21 Sep. 2012.
[10] L. D. Xu, C. Wang, Z. Bi, and J. Yu, "AutoAssem: An
Automated Assembly Planning System for Complex
Products," IEEE
Transactions
on
Industrial
Informatics, vol.8, no.3, pp.669-678, Aug. 2012.
[11] N. Greenough, and C.C. Kung, "A new high-efficiency
stepper motor driver for old technology stepper
motors," IEEE 25th Symposium on Fusion Engineering
(SOFE), 2013, pp.1-4, 10-14 Jun. 2013.
[12] A.W. Louw, and C. Neethling, "Digital LED lighting
solutions," Proceedings of the 21st Domestic Use of
Energy Conference (DUE), 2013, pp.1-4, 3-4 Apr.
2013.
[13] X. Zhang and Z. Zhang, "Design on Sampling Circuit of
EEG Signal Based on AT89C2051 SingleChip," Fourth International Conference on Innovative
416
Computing, Information and Control (ICICIC), 2009,
pp.454-457, 7-9 Dec. 2009.
[14] J. Xi, G. Liao, and W. Yang, "Study of Stepping Motor
Subdivision Driver," International Conference on
Intelligent Computation Technology and Automation
(ICICTA), 2010, vol.3, pp.443-446, 11-12 May 2010.
[15] F. Yuan, L. Tsai, V. Prakash, D. P. Dandekar, and A.M.
Rajendran, "Dynamic shear strength of S2 glass fiber
reinforced
polymer
composites
under
shock
compression," Journal of Applied Physics , vol.103,
no.10, pp.103537,103537-8, May 2008.
[16] C. Huang, S. Wang, C. Chang, C. Huang, Y. Cheng and
L. Fu, "An air combat simulator in the virtual reality
with the visual tracking system and force-feedback
components," IEEE International Conference on
Control Applications, 2004. Proceedings of the 2004,
vol.1, pp.515-520, 2-4 Sep. 2004.
[17] R. Mirshak, "Collateral damage risk: Quantifying the
water
surface
bullet
ricochet
problem," International Waterside Security Conference
(WSS), 2010, pp.1-7, 3-5 Nov. 2010.
[18] T. Matsuno, Y. Shima, T. Motoyoshi, K. Koyanagi, T.
Oshima, and T. Fukuda, "Basic characteristics analysis
of tuning fork structured piezoelectric vibration sensor
and application for robot control," International
Conference on Mechatronics and Automation (ICMA),
2011, pp.1532-1537, 7-10 Aug. 2011.
[19] X. Xiumei, and P. Jinfeng, "The simulation of
temperature and humidity control system based on
PROTEUS," International Conference on Mechatronic
Science, Electric Engineering and Computer (MEC),
2011, pp.1896-1898, 19-22 Aug. 2011.
[20] M. S. Lee, and H. C. Lee, "Dummy Gate-Assisted nMOSFET Layout for a Radiation-Tolerant Integrated
Circuit," IEEE Transactions on Nuclear Science,
vol.60, no.4, pp.3084-3091, Aug. 2013.
[21] R.D. Snow, J.R. Garner, and J.M. Whitaker.
"Developing Load Cell Monitoring on the Mock UF6
Feed and Withdrawal System." ORNL/TM-2013, no.34,
2013.
[22] www.ruggedpcreview.com/3_slates_advantech_3100r.h
tml
[23] M. Radek, and J. Zidek. "Application of synthetic
instrumentation that applies the trend of software-based
approach for measuring on the field of modern wireless
transfer systems." International Journal of Digital
Information
and
Wireless
Communications
(IJDIWC), vol.2, no.3, pp.208-221, 2012.
[24] E. Wang, S. Zhang, and Z. Zhang, "A Study on GPS
Signal Performance Analysis Platform Based on
LabVIEW," 8th International Conference on Wireless
Communications, Networking and Mobile Computing
(WiCOM), 2012, pp.1-4, 21-23 Sep. 2012.
417
Information Transparency in Admission Selection of State Islamic Higher Education

in Indonesia Using Information Technology Approach
Linda Salma Angreani
Faculty of Science and Technology
Maulana Malik Ibrahim State Islamic University
Jalan Gajayana No.50, Malang
Indonesia
lindasalma@gmail.com
ABSTRACT
The right to information is one of fundamental
human right essential. The transparency, availability
and accessibility of information enable people to hold
policy-makers to account and participate in the
decision making process, promoting better
governance, and reducing inefficiency and
corruption. In Indonesia, there is an integrated system
of student admission to State Islamic Higher
Education, named SPMB-PTAIN. This kind of
integrated admission selection model to enter State
Islamic Higher Education in Indonesia through
written examination conducted simultaneously
nationwide involving 53 Islamic Higher Education
Institution and coordinated by the Ministry of
Religious Affairs of Republic of Indonesia. In 2013,
this model are used by for about 200 thousands
applicants. However, there is no such model that
provides information about exams result tracking for
users. These condition lead to some critical issues
related to the transparency in SPMB-PTAIN exams
results. The subject of this paper is proposing the
monitoring system as part of system improvements of
SPMB-PTAIN (named SMUPDA) in terms of
monitoring and transparency. SMUPDA has been
tested in SPMB-PTAIN 2013 and was made such a
clear form of continuous improvement in data
processing of SPMB-PTAIN. SMUPDA can also
provide transparency of data processing on the
implementation of SPMB-PTAIN. Therefore the
similar model may be applied to other systems
related to public services to provide transparency and
gain the public trust of public services institutions
itself.
KEYWORDS
Right to information, Transparency, Continuous

Improvement, Data Processing, Work Flow,
Admission Process, Monitoring System
1. BACKGROUND
The right to information is recognized as a
fundamental human right essential for
meaningful democracy and the advancement of
other rights [1]. The transparency, availability
and accessibility of information enable people to
hold policy-makers to account and participate in
the decision making process, promoting better
governance, and reducing inefficiency and
corruption [2]. The Freedom of Information Act
in Indonesia adopted in 2008 named Undangundang No.14 Tahun 2008 tentang Keterbukaan
Informasi Publik, and it came into effect on 1
May 2010 [3].
In Indonesia, there is an integrated system of
student admission to State Islamic Higher
Education, named SPMB-PTAIN. This kind of
integrated admission selection model to enter
State Islamic Higher Education in Indonesia
through written examination. SPMB-PTAIN
conducted simultaneously nationwide involving
53 Islamic Higher Education Institution and
coordinated by the Ministry of Religious Affairs
of Republic of Indonesia. In 2013, this model are
used by for about 200 thousands applicants.
Figure 1 shown the new student admission
process published by SPMB-PTAIN Committee.
418
However, there is no such model that provides
information about exams result tracking for
users. Therefore, users only have information
about exams final result whether their pass or
fail the exams - as published by the admission
committee, without knowing the detail results of
the exams, especially the cause that make them
fail the exams. These condition lead to some
critical issues related to the transparency in
SPMB-PTAIN student admission exams results.
The subject of this paper is proposing the
monitoring system as part of system
improvements of SPMB-PTAIN system in terms
of monitoring and transparency. This monitoring
system has been tested in SPMB-PTAIN 2013
and was made such a clear form of continuous
improvement in data processing of SPMBPTAIN.
3. Marking process (Flagging) any of unfair or

cheating that was reported as well as the
absence of exams participants and finishing
the data validation from the scanning
process of LJU based on participant online
registration data.
4. Scoring and assess process on sampling
participant exam answers for answer key
analysis purposes.
5. Develop a statistical report of the appraisal
value of the sampling participants for
answer key analysis purposes.
6. Scoring process on all participant exam
answers using the answer key that has been
declared as valid answer key.
7. Setting up the all participant data forms that
have been equipped with test scores with the
format specified and sent it to the ICT
coordinator.
8. Creating statistical reports for analysis of
scoring results.
9. Creating an attendance report of all
participants.
Figure 2 shows the workflow of tasks to be
completed by data processing team based on
Data Processing job description in the SPMBPTAIN SOP book.
Figure 1. SPMB-PTAIN Admission Process

Source: [4] with modification (steps translation)
2. DATA PROCESSING FUNCTION IN

SPMB-PTAIN
As stated in SOP book of SPMB-PTAIN, the job
description of Data Processing Team are [5]:
1. Scanning all of exam answer sheet (called
LJU Lembar Jawaban Ujian)
2. Receive all participant list from online
registration (the data were provided by ICT
coordinator team).
To run the workflow described in figure 2, the

data processing work team was formed in small
team consist of:
1. Verification and Validation Team.
2. Document Archiving Team.
3. Scanning Team.
4. Scoring Team.
5. Scoring Team.
When users finishing their written exams in
SPMB-PTAIN (step 7 in figure 1), they will
immediately receive the results within 2 (two)
weeks after (step 8 in figure 1). However, there
are several processes performed by admission
committee start from step 7 to step 8 in figure 1.
419
In SPMB-PTAIN, there are at least 10 (ten)
steps, as shown in figure 2, that are managed by
data processing team of SPMB-PTAIN [6].
scoring and passing team for scoring and rank

them in accordance with the applicable
regulations (steps 9&10). The results obtained
from the scoring team set as baseline
information in national graduation meeting
agenda (step 11). The results of this national
meeting will end with announcements of exams
result to the public.
3. ISSUES FROM CURRENT STATE
From the results of evaluation of SPMB-PTAIN
2012, some issues are found in the
implementation of data processing in 2012 [7]
and also reported by data processing team [8].
Table 1 shows details of that issues.
Table 1. Issues in Data Processing of SPMB-PTAIN 2012
No.
1
Figure 2. SPMB-PTAIN data processing
The processes that starts from LJU delivery

along with supplementary files (letter of
statement from exams implementer, attendance
form, etc.) by the Local Committee (called
Panlok Panitia lokal) to the National Data
Processing (steps 1 &2 of figure 2). Next
process followed by the handover of LJU and
proceeded with the verification & validation
process (steps 3&4). After validation step, LJU
will store and waiting for scanning process (step
5). After that LJU will followed the scanning
process (step 6). The results of data captured by
scanning team are revalidated (a quality
assurance process) step 8, while at the same
time the physical of LJU then stored as archives
(step 7). Captured data will submitted to the
Issues
There is a delay in delivery of LJU and disrupting
the flow or process and queue of the next process of
data processing.
There are some errors in the recording report by
exam supervisors in field, such as:
a. The number of participants presence in the
exams location was not match with the
actual situation.
b. Exams supervisors are fill biographical
data of exams participants that did not
attend in the Exam Answer Sheet (LJU)
There is some discrepancy between the physical
LJU and number of LJU reported that make the
matching control of LJU is not running
There are some participants who make a mistake to
fulfill biographical data, including a crucial data
such as:
a.
Participant Test Number
b.
Exams Code
4. DATA MATCHING CONTROL IN DATA

PROCESSING
From several issues that occur as shown in Table

1, the data processing team propose steps of
recommendations for improvement of that
issues. Recommendations are made based on the
factors causing that issues. Presumptions of the
420
main factors that causes some issues in table 1 is
came from the large number of documents that
handled by exams implementer team in field
(such as number of LJU, statement of purpose,
participant attendance list, etc.), and large people
involved in the data processing. This make a data
matching control in data processing are needed.
Here are the recommendations and tools needed
to control data processing issues [7]:
1. Recommendation: Better planning
coordination are needed in delivery of LJU to
reduce the possibility shipping delays.
Tools: LJU delivery schedules and time
arrival estimation with delivery reporting
features are needed.
2. Recommendations: for evaluation No. 2 and
3 (table 1) is Develop and Implement Exams
Reporting Applications with benefits:
a. Find out more detailed information early
about exams participant attendance list
b. Attendance list report can be used as a
data control of physical LJU received
and processed
c. Simplify the process of verification and
validation of LJU.
d. Detection and reporting of unfairly
information.
e. Perform grouping of unfairly acts as a
justification for graduation meeting.
f. Find out more information early about
mismatches amount of LJU and perform
data tracking.
g. Avoid the occurrence of errors and
irregularities in LJU information, such
as no/incompatibility examinee number,
etc.
Tools needed: Develop and Implement
Reporting Exam applications, with features:
a. Online apps, and fulfill as soon as
possible after the exam ends
b. Have notification function when it was
not fulfill yet after the exams done and
sent it to the relevant parties.
c. Attendance list of examinees

d. Provide comparison list of examinees
with database
e. Have notification function when there
are discrepancies on the comparison
results
f. Printing results of attendance report list
for LJU admission control tool.
g. Scan attendance list report evidence
h. Unfairly report and Grouping levels of
unfairly acts
3. Recommendation: Scanning team does
conduct intervene in the LJU who has been
filled by exam participants. LJU are left as its
original condition.
4. Recommendation: Do a socialize campaign
about the importance of the accuracy of LJU
by exam participants.
Tools needed:
a. Socialize campaign can be done using
several medias such as Poster, Email,
Web announcement, etc.
b. The announcement shall submit the
statistics data that there are many
students who fails the exam due to an
error of filling exams numbers, etc. The
campaign can also be done by
submitting LJU fulfillment tips.
All of these recommendations then leads to
development of application of Exams
Monitoring and Data Processing System named
SMUPDA (Sistem Monitoring dan Pengolahan
Data). Some important things that need to be
maintained the compliance in information and
data control including:
1. Number of Applicants.
2. LJU Delivery date.
3. Number of Examinees.
4. Number of LJU.
5. Number of LJU scanned
6. Number of data capture in scanning process
7. Number of data for scoring process
8. Number of data scoring result
421
9. Number of participants who pass the exams
5. EXAMS
MONITORING
AND
DATA
PROCESSING SYSTEM (SMUPDA)

There are 2 (two) main objectives that leads to
SMUPDA development:
1. Developed a system for SPMB-PTAIN
committee and national exams data
processing that provide tracing function in
LJU processing. Thus the possibility of lost,
scattered, or misplaced of LJU can be
minimized.
2. Provide transparency of information related
to data processing. This information
transparency aims to improve public
confidence of student admissions process of
SPMB-PTAIN. In the other side, for each
examinees who did not pass the exams, they
will know exactly the cause of their failure
and can make improvements in the future.
Panlok The results of that initial data will be set

as baseline in the database. For data processing
team, that information then will be matched with
the physical documents sent by Panlok. After
verification process, that data will also be a
baseline information for next processes. Each
steps in data processing then followed by
examination of suitability at the time of
handover between every steps within data
processing.
Information that was stored in database are also
can be extracted for appropriate information to
examinees, i.e. detail information of exams
scoring. With this information, examinees will
receive transparent information regarding
determination of exams result or admission
graduation for them.
There are three main area that improved while

developing SMUPDA system: process, people,
and technology.
5.1. Process
Based on SPMB-PTAIN data processing steps
(figure 2), SMUPDA create new additional
process through that steps. Basically, that new
additional process are designed to provide
information for monitoring process function of
exams results (LJU), including number of LJU
and its comparison through each steps, detail
exams result, scoring, etc. After all, aggregate of
these information can be show to examinee as an
executive summary of their own data. Figure 3
show the workflow of information entry and
tracking in SMUPDA.
At the beginning of the process (step 1), all
related information of exams, such as log book,
attendance list, number of LJU, and so forth, are
enter in SMUPDA database performed by
Figure 3. SMUPDA process
5.2. People
One of successful key of SMUPDA
implementation is the discipline of people who
run and manage the system to follow their role
and responsibilities. Based on SMUPDA
workflow (figure 3), the role and responsibilities
of every actor in the system are:
1. Actor: Exams Local Committee (Panlok)
Role and responsibilities:
a. Filling log book, attendance list, number

of LJU, etc.
422
2.
3.
4.
5.
6.
b. Monitor the exams processing results

conducted by data processing team
(Lahta).
Actor: Lahta Retrieval and Verification
a. Accepting LJU submissions and other
supplementary documents from Panlok
b. Conformity checks of number of LJU
between data in the sistem and physical
LJU that sent by Panlok
Actor: Lahta Archive and Documentation
a. LJU conformance checks with the
received physical LJU
b. Manage storing process of LJU before
scanning
c. Manage and archive LJU after scanning
Actor: Lahta Scanning
a. Number of LJU conformance checks
between data in the system with physical
LJU LJU to be scanned
b. Perform LJU scanning
c. Perform results inspection and validation
of scanned LJU
d. Submit scanned result data to the Lahta
scoring team
Actor: Lahta Scoring
a. LJU conformance checks between data
in the system with LJU scanned result.
b. Perform scoring process.
c. Submit scoring results to Lahta
Graduation team.
Actor: Lahta Graduation
a. Perform LJU conformance check with
LJU scoring result.
b. Perform ranking process.
c. Set up data and information of exams
result for admission graduation meeting
agenda.
d. Perform control and monitoring process

of
data
processing
through
announcement exam result.
e. Coordinating the admission graduation
meeting agenda.
7. Actor: Information Management
a. Organize and manage the flow of
information on the system
b. Manage information for Panlok, Lahta,
and users, as well as other stakeholders.
5.3. Technology
In general, technology that used in SMUPDA
divided into 4 (four) areas:
1. Panlok area; used by Panlok to entry first
data as baseline data and communicate and
perform
information
transaction
electronically.
2. Lahta area; used by Lahta to be able to
perform information tracking related to
processing exams result.
3. Service and Management area; used by
information management people to manage
all information and service to be used by all
stakeholder.
4. User area; used by examinee to get
transparency of information from SPMBPTAIN committee.
Figure 4 shown the composition scheme of that
areas.
423
Figure 4. SMUPDA technology implementation areas.
Techlogy in Panlok area is facilitated Panlok to

perform
information
transaction
with
SMUPDA. Information transaction may include
data entry of exams supplementary files or
perform process tracking of data processing.
Information tracking or process tracking are
limited to their own exams region. All
transaction are provided in secure web mode.
For those Panlok who cannot access SMUPDA
from web (still happen in some area in
Indonesia due to lack of internet access) some
information transaction can be done using short
message service (SMS).
Figure 5. SMUPDA looks at panlok area [9].
Each Panlok activity will always be associated

with SMUPDA in form of information, including
log book entry, attendance list entry, exams sheet
collection process, etc. Each of these activities
will always be recorded including PIC that
responsible to that activities, so each activity that
performed in Panlok area will have a specific
person in charge that recorded in the system.
Thus, the expected security and suitability of
LJU can be further improved.
In Lahta area, Lahta operator can track all
information related to data processing of exams
result. Information tracking and reporting
process that perform in this area are also based
on Lahta team personal account. Track changes
function in the system are activated to log every
activities of data processing, including person
who done that activities and the time when the
activities was done.
424
With track shanges and tracking function, each
sub section of data processing team can perform
physical check of LJU number with the data that
stored in the system. If there is a shortage of LJU
that was sent to the next team than fixing process
can be done immediately. This process can also
be followed by the other parties related to
SMUPDA system. Figure 6 show SMUPDA
looks in Lahta area [10].
4.
Figure 6. SMUPDA looks at Lahta area [10].
5.4. SMUPDA Modules and Features

Overall features contained in SMUPDA are as
follows:
1. Login, user login/logout to SMUPDA
2. Panlok Module, contains:
a. Attendance list, entry of examinee
presence/absence
b. Recap, including
i. Start, exams rooms related to
examinee list
ii. Rooms, attendance list by exams
room and examinees
iii. Examinee, detail data of examinee
list per location and exams room
iv. Notes, notes or log book of exams
per room and examinees.
c. History, tracking function of LJU in
data processing
d. Send, entry and sending number of LJU
data as baseline information for SPMBPTAIN committee
3. Verification Module, data entry of physical
verification of LJU equipped with
5.
6.
7.
comparison function with data submitted by

SPMB-PTAIN Committee
a. Rooms and student, which show
attendance list for verification
b. Notes, which contains notes and
information per exams room or student
c. Exams, which contains recap exams
information per location, exams rooms,
or student
d. Receive, which contains list of PIC that
responsible for LJU retrieval and
submission.
Scanning Module, scanning result upload
function and comparison function with
verification information.
a. Rooms, Student, and Notes; contains
attendance list per location, exams
rooms and students with detail notes for
each (if available).
b. Exams, contains recap of number of
examinees per exams location
c. Import, contains utility to import
scanning result.
Scoring Module, scoring result upload
function
Graduation Module, admission graduation
data upload function
Admin Module, contains admin and
monitoring function of each process in
SMUPDA, including panlok, verification,
scanning, scoring, track changing, statistical
dashboad, etc.
6. SUMMARY
1. Data processing monitoring system
(SMUPDA) that implemented in SPMBPTAIN 2013 have been reviewed and
can be implement in next SPMB-PTAIN
process.
2. The implementation of SMUPDA as
Data Processing improvement in SPMBPTAIN 2013 is part of governance
activities.
425
3. SMUPDA are also part of transparency
in public service activities.
4. SMUPDA provide detail information of
examinees and any obstacles to be
reported in field in real time.
5. SMUPDA are also act as tracking model
in each phase of data processing process.
6. SMUPDA have functionality of early
detection of discrepancies.
7. SMUPDA
act
as
information
recapitulation report in real time on each
phase of data processing.
8. SMUPDA made the control and
monitoring function are based on
system, rather than on personal.
9. SMUPDA is a model in realization of
continues improvement of the previous
year activities.
7. DISCUSSION AND NEXT RESEARCH
Although the improvements model has been
implemented in the SPMB-PTAIN 2013, in fact
the system is still in the testing phase. The full
implementation of such a system is also needs to
be improved especially with the human
resources that involved in data processing
activities.
Therefore, as a continuation of this study, the
evaluation process are needed on the successful
implementation of the system in at least 2 things.
First, from SMUPDA application, whether it is
ready meet the data processing needs of SPMBPTAIN, in this case is the control of information
and data. Secondly, the evaluation of any factors
that related to successful implementation of the
system due to capability of people who using the
system.
Finally, SMUPDA is intended as a solution to
some issues that happened earlier as part of
improvement process. Moreover this system can
also be used as a model that provides
transparency of information. Similar systems can
be developed to provide transparency of
information and data processing on other public

services.
REFERENCES
[1 Article 19; Tifa Foundation;, "Indonesia: Fulfilling

] the Right to Information," 20 December 2010.
[Online]. Available:
http://www.article19.org/resources.php/resource/1
676/en/indonesia:-fulfilling-the-right-toinformation. [Accessed 20 February 2014].
[2 A. Faisol, T. B. Suryani, S. K. E. Amy and W. T.
] Hanggoro, "Fulfilling the Right to Information Baseline Assessment on Access to Information in
East Nusa Tenggara, Indonesia," ARTICLE19 and
Tifa Foundation, Kupang, 2010.
[3 Ministry of Justice and Human Rights of Republic
] of Indonesia, Undang-undang Keterbukaan
Informasi Publik, Jakarta: Ministry of Justice and
Human Rights of Republic of Indonesia, 2008.
[4 SPMB-PTAIN Commitee, "SPMB-PTAIN
] Registration and Exams Process Flow," State
Islamic University Maulana Malik Ibrahim,
Malang, Indonesia, 10 January 2013. [Online].
Available: http://www.spmb-ptain.ac.id.
[Accessed 20 July 2013].
[5 Panitia Pelaksana SPMB-PTAIN 2013 (SPMB] PTAIN 2013 Executive Commitee),, Standard
Operational Procedure: Seleksi Penerimaan
Mahasiswa Baru Perguruan Tinggi Agama Islam
Negeri (SPMB-PTAIN) Tahun 2013., Malang:
UIN Malang Press, 2013.
[6 Angreani, "Fasilitas Penelusuran LJU pada Ujian
] Nasional (Tracking Facilities of Exams Result in
Nationwide Higher Education Admission
Examination)," Malang Post, p. 15, 7 July 2013.
[7 Panitia Pelaksana SPMB-PTAIN 2012, SPMB] PTAIN 2012 Evaluation Book, Malang: Panitia
Pelaksana SPMB-PTAIN 2012, 2012.
[8 L. S. Angreani, "Laporan Kegiatan Proses
] Pengolahan Data SPMB-PTAIN 2012 (SPMBPTAIN 2012 Data Processing Report)," SPMBPTAIN Data Processing, Malang, 2012.
[9 SPMB-PTAIN 2013 Data Processing, SUMPDA
426
] Manual Book for Panlok, Malang: SPM-PTAIN
Committee, 2013.
[10SPMB-PTAIN 2013 Data Processing, SMUPDA
] Manual Book for Lahta, Malang: SPMB-PTAIN
Committee, 2013.
427
Hybrid Boundary Detection Method for Image with Application to Coronary Plaque
Syaiful Anama,b, Eiji Uchinoa,c and Noriaki Suetakea
a
Graduate School of Science and Engineering, Yamaguchi University

1677-1 Yoshida, Yamaguchi 753-8512, Japan
{r501wa, uchino, nsuetake}@yamaguchi-u.ac.jp
b
Department of Mathematics, University of Brawijaya
Jl. Veteran Malang 65145, Indonesia
c
Fuzzy Logic Systems Institute
680-41 Kawazu, Iizuka 820-0067, Japan
ABSTRACT
This paper proposes a hybrid boundary detection
method for image based on a new modified level set
method and a fuzzy model. It is applied to a boundary
detection problem of coronary plaque. Level set
method has been applied widely in image processing. It
however does not work well for an intravascular
ultrasound (IVUS) image because an image gradient,
commonly used for calculating a speed function in the
level set method, cannot detect an image boundary
well. The level set method and the weighted image
separability proposed by the authors in the past were
applied for a coronary plaque boundary detection
problem. The level set method could not however
detect the plaque boundary in several regions. One
problem was that the candidates of the plaque
boundary detected by the weighted separability were
unclear in several regions. The other problem was that
the IVUS image often becomes shadowed and it
contains no texture information there, which is caused
by the presence of the guide wire. To overcome this
problem, we propose a new modified level set, and we
further propose a hybrid boundary detection method
based on the new modified level set and the Takagi
Sugeno (T-S) fuzzy model for detecting a coronary
plaque boundary. The boundary detection accuracy of
the proposed method was significantly better than
those of the previous methods we proposed in the past.
KEYWORDS
Coronary plaque, intravascular ultrasound image,
hybrid boundary detection method, new modified level
set method, fuzzy model.
1 INTRODUCTION
Acute coronary syndrome (ACS) happens when
the heart is not getting enough blood. If the
coronary arteries are narrowed or blocked by a

rupture of vulnerable plaque, which is built up
inside the coronary arteries, the heart does not get
enough oxygen. This can cause a heart attack.
Intravascular ultrasound (IVUS) method is a
medical imaging technique which allows to see the
inside of the blood vessel, visualizing the coronary
plaque in living individual [1].
IVUS images are employed for a diagnosis of
ACS. In the quantitative assessment of the
compositions of coronary plaque for a diagnosis of
ACS, firstly the luminal boundary (LB) and the
adventitial boundary (AB) of the coronary plaque
are to be detected and evaluated precisely.
Medical doctors manually detect the plaque
boundary and evaluate the area of plaque. Plaque
boundary detection is however a very hard and
time consuming work for medical doctors. This is
not only because the plaque boundary of the IVUS
image is difficult to be recognized, but also
because the number of IVUS images to be
processed by a medical doctor is very large. For
those reasons, a method to detect automatically the
plaque boundary with high accuracy is strongly
required.
The authors have proposed some coronary
plaque boundary extraction methods in [2] and [3]
to solve this problem. In those methods, the
Takagi-Sugeno (T-S) fuzzy model [4] was
employed to detect the plaque boundary.
Those methods could significantly reduce the
workload of medical doctors. However, those
methods needed the seed points to detect the
plaque boundary. If the seed points were placed in
the wrong position the accuracy of the methods
would be reduced drastically.
428
Boundary detection is one of the tasks in
computer vision which has wide applications such
as feature extraction, object recognition and image
segmentation [5]. The boundary detection is to find
the lines separating homogeneous regions.
Active contour models have been applied for
detecting image boundaries [6-11]. The active
contour models have several advantages over the
classical image segmentation methods, e.g., edge
detection, thresholding, and region growth.
The first advantage of the active contour models
is that they can achieve sub-pixel accuracy of
object boundaries [9]. The second advantage is that
the active contour models can be easily formulated
under a principled energy minimization
framework, and allow incorporation of various
prior knowledge, such as shape and intensity
distribution for robust segmentation [10]. The third
advantage is that they can give smooth and closed
contours as a segmentation result, which is
necessary for many applications.
Active contour models can be classified into
two categories, i.e., the classical snakes and the
level set method. In this paper, we use the level set
method, introduced by Osher and Sethian [12],
because it is a highly robust and accurate method
for tracking interfaces moving under complex
motions, which meets our requirements.
The level set method has been applied
successfully in many cases in image segmentation.
It has several advantages over other segmentation
methods such as the snake method, region growth
and thresholding. The advantages of the level set
method over the snake method are that the curve
may break or merge naturally during an evolution,
and its topological changes are thus automatically
handled.
The authors have proposed a method for
detecting a plaque boundary by using the weighted
image separability method and the level set method
[13]. An image gradient is commonly used for
calculating the speed function in the level set
method, but it cannot work well in the IVUS
image. Therefore, in [13], the image gradient in the
speed function was substituted by the weighted
image separability. However it failed to detect the
plaque boundary in some regions of the image as
shown in Fig 1.
The plaque boundary detection in IVUS image

is very difficult because a region of the IVUS
image backward the guide wire often becomes
shadowed, and then it contains no texture
information in the shadow region as shown in Fig
2. Thus it fails to detect the plaque boundary. The
other problem is that the candidates of the plaque
boundary detected by the weighted image
separability are unclear in some regions, which is
caused by noise.
The authors also proposed a method for
detecting an image boundary by using the
modified level set method [14]. In [14], the speed
function is modified to make the level set method
more successful in boundary detection regardless
of noise.
In this paper, we propose a farther modified
level set method. We at the same time propose a
hybrid boundary detection method based on its
modified level set method and the T-S fuzzy
model. The speed function in the level set method
is further modified based on [14] and [13] to take
advantages of each method.
We could successfully detect the plaque
boundary in the guide wire shadow region by using
the newly proposed hybrid boundary detection
method. The effectiveness of the present method
was also evaluated by the experiments using the
real IVUS images.
2 CORONARY PLAQUE BOUNDARY
CALCULATIONS IN IVUS IMAGE
The IVUS method is one of the medical imaging
techniques. It is an application of the ultrasound
technology to a medical area. In the IVUS method,
the catheter with an ultrasound probe attached to
its end is inserted and rotated in the coronary artery
Figure 1. The plaque boundary detected by the method [13]
429
proposed by the authors in the past.
Those are a luminal boundary (LB) and an

adventitial boundary (AB) shown in Fig. 2.
2.1 Anisotropic Diffusion Filter
The Perona Malik diffusion (PMD) filter [7] is one
of the methods for noise filtering. It has two
advantages, i.e., one is to preserve the edges of an
image, and the other is to reduce speckle noise.
The PMD filter is defined by:
It
(a)
I
div (c( x, y, t )I )
t
c( x, y, t )I c( x, y, t )I ,
(1)
c( x, y, t ) g ( I ( x, y, t ) )
(2)
where
denotes a diffusion coefficient. I represents a

gradient of an image. g () refers to an edge
stopping function, which is defined by:
1
g ( z)
,
2
(3)
z
1
K
(b)
Figure 2. The guide wire shadow problem of the IVUS
image. (a) The guide wire shadow of the IVUS Image in the
cartesian coordinate. (b) The guide wire shadow of the IVUS
image in the polar coordinate.
where K is a parameter which controls the strength

of diffusion. The initial condition is given by:
I ( x, y,0) I 0 ( x, y).
(4)
The PMD in discrete version is given by:

to observe the inside of the blood vessel,
visualizing the coronary plaque in vivo.
The ultrasound signal is transmitted from the
ultrasound probe, and the radio frequency (RF)
signal reflected from the tissue is also received by
the ultrasound probe. An IVUS B-mode image
shown in Fig. 2 is obtained by analyzing the
received RF signal.
The IVUS B-mode image is constructed of the
amplitude information of the received ultrasound
RF signals. That is, the RF signals are transformed
into intensities, and the intensities in all radial
directions are used to form a tomographic crosssectional image of a coronary artery to get the Bmode image of Fig. 2.
The plaque boundaries are necessary to be
detected for characterizing the coronary plaque.
I s( n 1) I s( n )
g (I
(n)
s, p
) I s(,np) ,
(5)
where s ( x, y) is the coordinates of the pixel of

concern. p and I s(n ) represent the neighboring
pixels of s and an intensity at I s(n ) with an
iteration count n , respectively. The discretization
of image gradient is given by:
I s(,np) I p( n ) I s( n ) ,
(6)
where s and s represent the diffusion directions

and the number of pixels in the neighboring area,
respectively. is a parameter.
g () takes a large value at the regions where
the intensity gradients are low. It takes a small
430
value at the regions where the intensity gradients
are high.
2.2 Image Separability
The candidates of the plaque boundaries in IVUS
image are given by using a statistical discriminant
measure of the image separability [16]. The
advantages of the image separability are:
1) Insensitive to noisy and blurred edges,
2) Able to differentiate the edges between
texture regions.
The weighted separability is modified by
considering the conditions peculiar of IVUS
image. The weighted image separability for pixel
h in Fig. 3 is calculated by the difference of the
two regions as follows:
I I
I
h max A B
I max
I max
w
h
Figure 3. Calculation of the image separability.
(7)
where I A and I B are the averages of intensities in

the regions of A and B. I max is the maximum
intensity of the whole IVUS image.
h represents the original image separability
which is defined by:
n A ( I A I ) 2 nB ( I B I ) 2
S
(I
k 1
I)
(8)
where n A and nB are the numbers of the pixels in

the regions of A and B, respectively. I is the
average of the intensities in the combined regions
of A and B. S and I k are the number of the pixels
and the level of intensity of the k-th pixel in the
combined region of A and B.
hw satisfies 0 hw 1 , and it takes a large
value when two regions are separated from each
other [17].
3 PROPOSED METHOD
We present in this paper a hybrid boundary
detection method based on the new modified level
set and the T-S fuzzy model for detecting a plaque
boundary in the coronary artery.
Figure 4. Guide wire shadow detection.
This chapter is divided into three sections.

Those are: (i) guide wire shadow detection, (ii)
plaque boundary detection by using the new
modified level set method, and (iii) inference of
plaque boundary in the guide wire shadow region
by using the T-S fuzzy model.
3.1 Guide Wire Shadow Detection
IVUS image has a guide wire shadow and it does
not contain any texture information there. Thus the
level set method fails to detect the plaque
boundary in that region.
To overcome this problem, the plaque
boundary in the guide wire shadow region is
inferred using the plaque boundary information on
431
the left and on the right hand side of the guide
wire shadow region.
In the first step of the present hybrid boundary
detection method, the guide wire shadow region is
detected. Its procedure is as follows:
1) Convert the B-mode image of Fig. 2(a) in the
cartesian coordinate system to the polar
coordinate system of Fig. 2(b).
2) Detect the position of the guide wire shadow.
From Fig. 2 it can be seen that the guide wire
shadow is located in the small gray-level area.
The mean gray-level of every column of the
polar coordinate system is calculated to
determine the guide wire shadow. The mean
gray-level of every column is given by:
contour C as the zero level set of a time dependent

level set function ( x, y, t ).
We assume that the level set function gets
the positive values outside the zero level contour,
and the negative values inside. The inward normal
vector can be expressed as N / , where
is a gradient operator.
The curve evolution of (10) is converted to:
/ t F ,
(11)
which is referred to as a level set evolution
equation. In this paper, we use the level set (x)
Y ( x)
I ( x, y )
y 1
, x 0,1, 2,..., 255,
(9)
where N is the number of pixels in one

column. The mean gray-level of (9) for Fig.
2(b) is shown in Fig. 4.
It is observed that the shadow area around the
guide wire has a small value of the mean graylevel as shown in Fig. 4. We can predict roughly
the location of the guide wire shadow region by
this mean gray-level of (9).
3.2 Plaque Boundary Detection by Using
Newly Modified Level Set Method
(10)
where t and s are a set point in time and a curve

parameter, respectively. N and F denote the
inward normal vector to the curve C , and the
speed function. The speed function F controls the
motion of the contour.
The curve evolution of (10) can be converted to
a level set formulation by embedding the dynamic
1023
255
Intensity
C / t F N ,
Figure 5. IVUS image after ignoring the guide wire

shadow.
Distance from Probe [pixels]
The level set method has been applied in many

areas, especially for detecting the image boundary.
In the level set method, the contour is represented
by the zero level set of a higher dimensional
function. This is called a level set function, and it
formulates the motion of the contour based on the
evolution of the level set function.
The curve evolution of a parametric contour
C ( x(s, t ), y(s, t )), is given by:
511
0
Angle Index
Figure 6. IVUS image after applying the PMD filter to Fig. 5.
432
1
511
Intensity
1023
2)
3)
0
Angle Index
Figure 7. Weighted image separability of Fig. 6.

255
4)
5)
Intensity
1023
511
0
Angle Index
Figure 8. Initial contour of the new modified level set.
( )div ( g / ) g ( ),
(12)
where is a dirac delta function, div is a

divergence operator, and g is a speed function,
which is given by:
g 1/(1 (G * I ),
7)
8)
9)
The plaque boundary detected by using the

present modified level set is shown in Fig. 9.
defined by:
/ t div (d p ( ) )
6)
analyzing the real IVUS images which are

used for the experiments. Examining the
many IVUS images, a and c are set to be b-14
and b+8, respectively. b is a point where the
mean-gray level becomes minimum.
Merge areas c to 255 (region B) and 0 to b
(region A) by placing the latter on the right
side as shown in Fig. 5.
Reduce the speckle noise by applying the
PMD filter to Fig. 5. The filtering result is
shown in Fig. 6.
Calculate the weighted image separability to
obtain Fig. 7.
Calculate the new modified speed function of
(14).
Give the initial contour of the level set, e.g.,
as shown in Fig. 8.
Calculate the contour evolution by using (12).
Calculate the new contour.
Repeat steps 7) and 8) until it converges or
the maximum number of iterations is reached.
(13)
3.3 Inference of Plaque Boundary in the Guide

Wire Shadow Region by Using T-S Fuzzy
Model
The plaque boundary in the guide wire shadow
region is inferred by the T-S fuzzy model [4]. The
plaque boundary is interpolated by the information
which is taken from the plaque boundary on the
where G is the Gaussian filter and I is the image

to be processed[8].
In this paper, we propose a new speed function
defined by:
m
1
.
g
w
1
(14)
The following steps are the procedure of the

proposed plaque boundary detection:
1) Detect the guide wire shadow region (angle
index between a and c) shown in Fig. 4. The
positions of a and c are determined by
Figure 9. Plaque boundary detection results by using the

present modified level set method.
433
left and on the right of hand side the guide wire
shadow region as shown in Fig. 10.
The plaque boundary is inferred by the series of
the following fuzzy rules:
If xi is Au then f u ( xi ) au xi bu ,
(15)
where Au is a fuzzy set with the membership

function (MSF) u ( xi ), xi corresponds to the
angle index, and f u ( xi ) is a linear function.
Fig. 10 shows the complementary linear MSFs

which are allocated to infer the plaque boundary.
The u-th rule is used for approximating the plaque
boundary by a linear function in the interval
[b 50, b 50] . The plaque boundary is inferred
by:
y ( xi ) u ( xi ) fu ( xi ) u 1 ( xi ) fu 1 ( xi ).
(16)
The optimum coefficients in the consequent

part of the fuzzy rule are determined by using the
least square method. It minimizes the following
error criterion:
E ( yi y i ( xi )) 2 ,
(17)
where yi is a plaque boundary that is detected by

the newly modified level set method on angle
index i.
The black dotted lines on the angle interval [a,
c] in Fig. 11 show the plaque boundary inferred by
the T-S fuzzy model.
4 EXPERIMENTAL RESULTS
Three IVUS images were used for evaluating the
performance of the proposed method. The
proposed hybrid boundary detection method was
compared with the methods by using only the T-S
Table 1. RMSEs of boundary detection results for image 1.
Figure 10. Membership functions of the T-S fuzzy model
in the guide wire shadow region.
Figure 11. Plaque boundary detection results by a hybrid

of the newly modified level set method and the T-S fuzzy
model.
( m)
Method
Hybrid Boundary Detection
Method (1st Experiment)
Method (2nd Experiment)
Method (3rd Experiment)
Method (4th Experiment)
Method (Average)
T-S Fuzzy Model [2]
T-S Fuzzy Model Optimized by
Particle Swarm Optimization [3]
LB
AB
9.1
20.4
9.3
20.0
9.2
20.1
9.3
19.9
9.3
20.1
9.2
20.1
13.7
28.4
12.2
28.5
434
fuzzy model [2], and the T-S fuzzy model
optimized by particle swarm optimization (PSO)
[3].
If the initial contour of the level set is different,
the proposed method will produce a different
plaque boundary. Therefore the experiment for
each image was repeated 5 times with different
initial contours. The desired boundaries (correct
boundaries) were decided by expert by using the
difference between the image brightness.
(a)
( m)
Method
Method (Average)
T-S Fuzzy Model [2]
LB
AB
11.8
24.9
12.8
17.3
12.7
17.6
12.6
17.2
12.7
17.5
12.5
18.9
28.0
35.4
23.8
33.2
(b)
Figure 12. Comparisons of the plaque boundary detection
methods. (a) The IVUS image to be processed. (b) The
plaque boundary detection results. The solid white lines
show the desired (correct) boundaries and the solid black
lines show the boundaries detected by the proposed
method. The dashed black lines and the dashed white lines
show the boundaries detected by the T-S Fuzzy model [2]
and the T-S Fuzzy model optimized by particle swarm
optimization [3], respectively.
( m)
Method
Method (Average)
T-S Fuzzy Model [2]
LB
AB
15.3
12.0
15.6
13.1
15.5
21.8
15.9
12.5
15.7
12.3
15.6
14.3
20.0
30.2
19.9
40.1
The root mean square errors (RMSEs) between

the desired (correct) boundaries and the
boundaries detected by the proposed hybrid
boundary detection method are shown in Tables 1,
2 and 3. The RMSEs of the proposed method are
significantly better than those of the previous
methods [2], [3] for all images.
Fig. 12(a) shows the IVUS image to be
processed, and Fig. 12(b) shows one of the plaque
boundary detection results by the proposed
method. The solid white lines show the desired
(correct) boundaries and the solid black lines show
the boundaries detected by the proposed method.
The dashed black lines and the dashed white lines
435
show the boundaries detected by the T-S fuzzy
model [2] and the T-S fuzzy model optimized by
particle swarm optimization (PSO) [3],
respectively.
It can be observed from Fig. 12(b) that the
boundaries detected by the proposed method are
closer to the desired (correct) boundaries than
those by the previous methods [2], [3].
5 CONCLUSIONS
We have proposed a hybrid boundary detection
method for detecting a coronary plaque in an
IVUS image. It incorporates a newly modified
level set method with the Takagi-Sugeno fuzzy
model. The present method was compared with
the previous methods [2], [3], and the boundary
detection accuracy of the present method was
significantly better.
Acknowledgments: This work was supported by
the Grant-in-Aid for Scientific Research (B) of the
Japan Society for Promotion of Science (JSPS)
under the Contract No. 23300086. Syaiful Anam
would also like to thank DGHE postgraduate
scholarship of Indonesia for supporting his stay at
Yamaguchi University, Japan.
6 REFERENCES
[1] Potkin, B. N., Bartorelli, A. L., Gessert, J. M., Neville,
R. F., Almagor, Y., Roberts, W. C., Leon, M. B.:
Coronary Artery Imaging with Intravascular HighFrequency Ultrasound. Circulation, vol.81, pp.1575
1585 (1990).
[2] Koga, T., Ichiyama, S., Uchino, E., Suetake, N., Hiro,
T., Matsuzaki, M.: Fully Automatic Boundary
Extraction of Coronary Plaque in IVUS Image by
Anisotropic Diffusion and T-S Type Fuzzy Inference.
Advances in Intelligent and Soft Computing, SpringerVerlag, AISC 75, pp.139147 (2010).
[3] Anam, S., Misawa, H., Uchino, E., Suetake, N.:
Parameter Tuning by PSO for Fuzzy Inference-Based
Coronary Plaque Extraction in IVUS Image. In: Proc.
of the Joint 6th International Conference on Soft
Computing and Intelligent Systems and 13th
International Symposium on Advanced Intelligent
Systems (SCIS & ISIS 2012), pp.1426-1429, Japan,
(2012).
[4] Takagi, T., Sugeno, M.: Fuzzy Identification of

Systems and Its Applications to Modeling and Control.
IEEE Transactions on Systems, Man, and Cybernetics,
vol.SMC15, pp.116132 (1985).
[5] Leordeanu, M., Sukthankar, R., Smichisescu, C.:
Efficient Closed-Form Solution to Generalized
Boundary Detection. European Conference on
Computer Vision, LCNS, vol.7575, pp.516-529 (2012).
[6] Horbert, K., Rematas, K., Leibe, B.: Level-Set Person
Segmentation and Tracking with Multi-Region
Appearance Models and Top-Down Shape Information.
In: Proc. of International Conference on Computer
Vision, pp.1871-1878, Spain, November 2011.
[7] Li, C., Huang, R., Ding, Z., Gatenby, J. C., Metaxas, D.
N., Gore, J. C.: A Level Set Method for Image
Segmentation in the Presence of Intensity
Inhomogeneities with Application to MRI. IEEE
Transactions on Image Processing, vol.20, pp.20072016 (2011).
[8] Li, C., Xu, C., Gui, C., Fox, M. D.: Distance
Regularized Level Set Evolution and Its Application to
Image Segmentation. IEEE Transactions on Image
Processing, vol.19, pp.3243-3254 (2010).
[9] He, L., Zheng, S., Wang, L.: Integrating Local
Distribution Information with Level Set for Boundary
Detection. Journal of Visual Communication and
Image Representation, vol.21, pp.343-354 (2010).
[10] Panchal, P. U., Jondhale, K. C.: Image Object
Detection Using Active Contours via Level Set
Evolution for Segmentation. Journal of Signal and
Image Processing, vol.3, pp.97-101 (2012).
[11] Osher, S., Sethian, J., Fronts Propagating with
Curvature-Dependent Speed: Algorithms Based on
Hamilton-Jacobi
Formulation.
Journal
of
Computational Physics, vol.79, pp.12-49 (1988).
[12] Li, M., He, C., Zhan, Y.: Adaptive Regularized Level
Set Method for Weak Boundary Object Segmentation.
Mathematical Problems in Engineering, vol.2012, pp.116 (2012).
[13] Anam, S., Misawa, H., Suetake, N., Uchino, E.:
Coronary Plaque Boundary Extraction in IVUS Image
by Using Weighted Image Separability and Level Set
Method. In: Proc. of the 14th SOFT Kyushu Chapter
Annual Conference, pp.37-38, Japan (2012).
436
[14] Anam, S., Uchino, E., Suetake, N.: Image Boundary
Detection Using the Modified Level Set Method and a
Diffusion Filter. In: Proc. of the 17th International
Conference on Knowledge-Based and Intelligent
Information & Engineering (KES 2013), pp.161-169,
Japan (2013).
[15] Perona, P., Malik, J.: Scale-Space and Edge Detection
Using Anisotropic Diffusion. IEEE Transactions on
Pattern Analysis and Machine Intelligence, vol.12, pp.
629-639 (1990).
[16] Fukui, K.: Edge Extraction Method Based on
Separability of Image Features. IEICE Transactions on
Information System, vol. E78-D, pp.1533-1538 (1995).
[17] Koga, T., Uchino, E., Suetake, N.: Automated
Boundary Extraction and Visualization System for
Coronary Plaque in IVUS Image by Using Fuzzy
Inference-based Method. In: Proc. of the International
Conference on Fuzzy Systems, pp.1966 1973,
Taiwan (2011).
437
A Multi Tree based Approach for Performance Analysis in Hierarchical Wireless

Sensor Networks
1,2
Itu Snigdh1, Swarn Saurabh2

Dept. of Computer Science & Engg., B.I.T Mesra, India
1
itusnigdh@bitmesra.ac.in, 2 swarn3107@gmail.com
ABSTRACT
A wireless sensor network consisting of heterogeneous
sensors is capable of dealing with more than one type
of information about any region at a time. Sensor
applications that cater to disaster relief require faster
data delivery and effective connectivity than the other
monitoring applications. Moreover, usually, sensors
collect a single type of data. Our approach focuses on
generating a tree based communication architecture
that is made up of three different types of sensors and
ensures that at least one sensor of each type is
connected to the communication backbone; thereby
forming a MULTI-TREE. By responding to query
initiated by the sink for one type of data at a time, the
heterogeneous sensors ensure reduced amount of
carried load and latency in data transmission. Our
strategy implemented for structural health monitoring
application shows that when compared to a minimum
spanning tree under the same simulation constraints,
performs better as it reduces the carried load and delay
to one third under the existing set of ad hoc protocols.
domains. These vary from Type I applications (eg.

real time applications like intrusion detection
systems, Alarm or event detection, structural
health monitoring etc.) to Type II applications (eg.
Continuous monitoring like environmental
monitoring) or Type III (eg. hybrid applications
like structural health monitoring that requires
continuous monitoring as well as immediate alarm
system in case of an anomaly or disaster). The
basic areas of importance in a WSN can be
summarized as follows in figure 1.
KEYWORDS
Wireless sensor network, Minimum Spanning Tree,
Multi tree, Carried load, Energy consumption, Delay.
1 INTRODUCTION
The main aim of a wireless sensor network is to be
able to deliver the required functionality with
unattended operation for the longest possible time
without sacrificing the major constraints.
Irrespective of the type of application scenario, the
main requirement for nodes is to be self-powered.
This implies they will either have to contain
enough stored energy to last for years, or they will
have to be able to scavenge energy from the
environment. Till date, a wireless sensor network
has been subject to a vast number of application
Figure 1. Wireless sensor network features
As obvious, the communication quality of a

network can vary according to the varied spatial
displacements in addition to the associated
438
connections. This are true to a greater degree for
ad-hoc (or infrastructure-less) communication.
Once sensors have been deployed, they need to
form some kind of network for communication.
Since the sensors have limited energy, the network
used should be efficient enough to minimize the
energy usage.
In a clustered network, many sensors connect to a
single sensor, known as cluster head, forming a
cluster, which joins to the sink to complete the
communication backbone network. All sensors
send their data to their cluster head, from where
these data are transferred to the sink. The
communication between the member and cluster
heads and the corresponding inter cluster head
communication usually happens in multi tiers. So,
this architecture resembles a hierarchy similar to a
tree which we analyze in our approach. Network
energy consumption is reduced as the number of
nodes involved in long distance transmission gets
reduced. But in this type of architecture, when
cluster head goes out of order, the whole cluster
becomes useless rendering reconfiguration. This
situation is analogous to a tree, where, to use the
sensors of such cluster, tree restructuring can be
done for that part of the network. Hence, we
advocate the use of a tree-based network for the
aforementioned purpose. In a tree based network,
we free the node connections from any loop or
cycle, which ensures the non redundancy and
transmission of data in one direction only. Here,
each sensor node is connected to sink directly or
via other nodes depending on their distance from
the sink and all the sensors can send their data to
the sink via single or multi-hops. The issue that
arises is to minimize these transmission distances
by using different available tree routing algorithm
and protocol. Some of the available tree based
architecture in wireless sensor networks are:
clustered tree network [1], Minimum Steiner tree
[2], Minimum spanning tree, Minimum-hot-spot
query trees [3], Query routing tree [4]. MinimumHot-Spot Query Trees are query routing trees with
three properties; decreased collisions during data
transmission, decreased query response time, and
increased system lifetime and coverage. [4] uses a
query routing tree, which is formed in two phases
1) node discovery phase and 2) node balancing
phase. It tries to make a near_balanced tree on the
basis of work-load. Similar research as in [1] deals

with a Clustered hierarchical tree, in which
sensors join to cluster middle head, which further
join to cluster head, and finally joins the sink. It
uses Simulated Annealing method to optimize the
energy usage. [5] uses a Greedy incremental tree,
in which first trunk is formed adding the branches
latter. It focuses on data aggregation techniques
for minimizing energy consumption by using
some threshold value. [6] Deals with multi-tree
formation in four steps 1) Initialization phase, 2)
Tree selection phase 3) Normal phase 4) Recovery
phase. This procedure first forms the multi-trees
and uses the minimum cost tree for
communication. When minimum cost tree fails,
next minimum cost tree is used.
The above researches focus either on the energy
conservation or on managing the workload, delay
or coverage of the network. Moreover they deal
with homogeneous sensors, which can be used for
only one type of data. If we need to monitor some
area for more than one type of data at the same
time, we need different sensors. There is no
literature available as far as our knowledge that
takes into account collection of multiple data types
via the same communication backbone. This
problem motivates us to search for a method
where a single backbone is utilized for collecting
different types of data.
Ongoing current research focuses on in-network
data processing or data aggregation. There are
ample works that advocate the use of tree based
structure for achieving the quality of service
metrics for wireless sensor networks. In most
cases we can infer that use of single paths to
connect each node to the base station has been
preferred as opposed to using the multiple paths at
the same time due to the duplication of the same
packet [12]. Moreover, on analysis of the
performance of cluster-based and tree-based
topologies, it is found that cluster-based topology
is more energy efficient for aggregation than treebased topology; but in the case of acquisition,
tree-based topology is more energy efficient than
cluster-based topology. Other similar works like
Tree-based routing protocols are aimed to
construct the best route from a node to base station
[13]. Protocols like HTECRP [14] claim to
manage congestion and perform fairness on the
439
network by assigning privileges to the traffic.
ViTAMin [15] offers a hierarchical backbone tree
algorithm for energy efficiency and sufficient
network lifetime. While Localized area spanning
tree (LAST) protocols for wireless short range
sensor networks optimizes the energy cost and the
interference imposed by the structure [16], a BFS
based tree rooted at the base station offers shortest
path traversal for each data message which utilizes
the sensor resources efficiently by employing a
local repairing approach for the crashing nodes
thereby increasing the lifetime. [17]. CTP is a
routing protocol implemented in TinyOS-2. x,
offers 90-99.9% packet delivery in highly
dynamic environments while sending up to 73%
fewer control packets than existing approaches.
[18], [19], [20]. Tree based strategies reduce the
burden of retransmissions and hence can be used
for congestion management. [14]. Thus, it can be
believed that a tree structure is very popular as in
wireless sensor network structure, in most of
applications when we have one sink and too many
sender nodes.
To address this problem we discuss a query
routing multi-tree approach on a set of
heterogeneous sensors. By using different sensors
for different purpose we aim to reduce the burden
of the nodes in terms of packets generated,
forwarded and aggregated by one third. This also
affects the overall delay and the final throughput
of the network. We implement our algorithm for
structural
health
monitoring
applications
specifically as it does not require a large number
of sensor nodes or a dense deployment scheme
and needs three different sensors to aptly monitor
the structure. The Requirements of Structural
Health Monitoring are enlisted as follows: SHM is based on monitoring buildings and
bridges which requires different classes of
data to be collected. Our multi-tree
approach
tries
to
facilitate
this
requirement.
Data redundancy is acceptable.
Data security is not that important.
Doesnt require dense deployment.
Saving energy is not the main constraint as
redeployment is possible.
These conditions make it ideal for analyzing the

performance of our algorithm.
The remainder of this paper is structured as
follows. Section 2 describes the MULTI-TREE
algorithm. The results are analysed in section 3
where we compare the performance of the
proposed algorithm with respect to a Minimum
spanning tree under a given set of protocols.
Concluding remarks and directions for future
research are provided in Section 4.
2 SYSTEM MODEL
We assume that the initial deployment of the
sensors in the region of interest is an undirected
graph G = (V, E), where V = VA VS VT and
VA, VS and VT are the set of nodes for
Accelerometer, Stress/Strain and Temperature
type sensors respectively and E is the set of edges
connecting these vertices. For the necessary
connectivity a sub-graph T = (Vt, Et) is
constructed which is a spanning tree of G and ET
= (u, v) be the set of edges such that ET T, u Vt
and v (V-Vt) which is selected according to the
following condition.
Sink calculates its distance from VA, VS and VT
and connects to min(E) for each set of VA, VS and
VT putting the connected vertex in Vt and edge in
Et. Next VA, VS and VT from the set Vt calculates
its distance from nodes in the set (V-Vt) and
connects to min(E) for each set of VA, VS and VT
from (V-Vt) putting the connected vertex in Vt and
edge in Et. This process repeats (VA, VS and VT)
V.
2.1 Assumptions
Once the sensors are connected, they start
communicating in order to receive and send data.
We assume that the sensor network has the
following properties:
Sensor nodes are heterogeneous (three
types of sensors) and location aware; and
Sensor
nodes
have
limited
and
irreplaceable battery power.
The tree is formed under a constraint that
each sensor will connect to only one sensor
440
of each type within their communication

range.
After tree construction, sink sends query of
desired type to receive data from the
network.
Data aggregation and fusion takes place to
minimize the total energy usage.
Sensor nodes are randomly deployed.
Sensor nodes and sink are stationary.
2.2 Procedure
Step 1: Choose any element r and a set S =
{r} where S and E = (Taking r as the root
node of the spanning tree.)
Step 2: Find the lightest edge (here the weight is
chosen on the basis of the shortest distance and the
associated sensor class type) such that one end
point is in S and the other point is in V\S. Add
the edge to E and the other point to S.
Step 3: If V\S = then stop and output the
spanning tree (S, E) otherwise go to step 1.
number of receptions in case of the parent node

and also the cost of data fusion or aggregation .
Since the majority of the energy consumption is in
transmitting and receiving data, this architecture
ensures that none of the sensors will deplete its
energy sooner than the other.
Table 1. NOTATIONS
Notations
N
Node_List[]
Array_T
Array_A
Array_S
Dist[]
S
i
n
k
EDi,j
Description
total number of nodes
deployed of each
type
Array to store all
deployed nodes
Array
to
store
Temperature/Pressure
sensor nodes
Array
to
store
Accelerometer sensor
nodes
Array
to
store
Stress/Strain sensor
nodes
Array
to
store
Euclidean distance
between i and j
Euclidean distance
between i and j
2.3 Algorithm
/*Tree Generation*/
Figure 2. Multi-Tree
A: Accelerometer
S: Stress/Strain,
T: Temperature/Pressure
In our architecture, sink links to one sensor of
each type, as shown in the figure 2, restricting its
children to be a maximum of three. This results in
a balanced tree construction, in which none of the
nodes are overburdened. When the sink requests a
data type, only those sensor nodes that belong to
its class respond to the query, this reduces the
1. Begin
2. Node_List[0] = Sink;
3. K=1;
4. For(i=0 to N-1)
5. {
6. For(j=i to N-1)
7. {
8. Dist[j] = Calculate EDi,j;
9. }
10. Sort Dist[i,n];
11. Connect Node_List[i] and Array_T[i];
12. Node_List[k]=Array_T[i];
13. K++;
14. Repeat STEP 5 to STEP 13 for Array_A[];
15. Repeat STEP 5 to STEP 13 for Array_S[];
441
16. }
17. End
For query processing, whenever sink needs to get
any type of data, it sends a query to all its
children, which is further forwarded to rest of
sensor nodes by next level nodes in the tree till it
reaches to all the sensor nodes. Sensors after
receiving the query responds by transmitting its
collected data to the upper level nodes in the
network, which is further forward till it reaches
sink. All the sensors respond to the query of their
type only and discard other queries, which results
in maximum of one-third sensors of the entire
network to send its data. This reduces the load on
the entire network by significant amount, thus
increasing the lifetime of the network. As the sink
node starts the tree construction, for which it
calculates the distances of all the neighboring
nodes the complexity is O(n). However, for
sorting the distances the complexity is O(n 2). So,
the overall complexity of Multi-Tree is O(n3).
The performance has been evaluated for Zigbee
application between the nodes in Qualnet
simulator as shown in figures 3 and 4. Figure 3
depicts the snapshot of backbone using a multi
tree with three different node types (represented
by different colors) while figure 4 depicts the
MST backbone with homogeneous nodes. The
results presented are obtained by Qualnet and
stand alone C packages as per the following
parameters:
9.
Energy model
10.
11.
12.
13.
Node Type
Items to Send
Item Size
Application Type
Linear gradient
model
MICAZ motes
100
127
Zigbee
Figure 3. Multi tree using Heterogeneous sensors
Table 2. Simulation Parameters
S.No
1.
2.
3.
4.
5.
6.
7.
8.
Parameters
Radio type
Transmission power
Node classes
Values
802.15.4
3.0 dbm
Stress/strain,
pressure
and
temperature,
Accelerometer
type
Number of nodes
13 nodes of each
type
Packet
reception PHY 802.15.4
model
Reception model
Modulation scheme O-QPSK
CCA Mode
Carrier sense
Noise factor
10.0
Figure 4. Minimum Spanning Tree
3 RESULTS and DISCUSSION

The multi tree is compared and analyzed with
respect to a corresponding minimum spanning tree
442
for the same set of simulation parameters and for
four routing protocols namely; AODV, Bellman
Ford, Dymo and Fisheye. The following figures
show the throughput of the proposed algorithm,
the offered load at the transport layer ,the average
carried load, the number of packet drops, the
average energy consumed in transmitting and
receiving, the delay and the jitter. The results are
analyzed for the Network and MAC Layer as the
Mac Layer protocols that traditionally manage
power saving are designed to be application aware
to some degree, for example they provide service
differentiation for data, query and management
packets. The factors considered for validating the
performance of our approach over the minimum
spanning tree is as follows:
a) Throughput at the application layer:
Throughput can be calculated as: Total bytes sent
or received*8 / (Simulation time - time first packet
is sent or received).
b) The Carried load: Indicates the total workload
that an individual node needs to process in terms
of data communicated.
c) The Average energy consumed by sensor
nodes in transmitting and receiving as any
wireless sensor network has to take energy
constraints into consideration. Energy consumed is
calculated using radio energy model [7].
The energy consumed to send a k-bit message over
distance d is:
fs
d0
Es (k , d ) k.Eelec k. fs .d 2 d d 0
(1)
Es (k , d ) k.Eelec k. mp .d 4 d d 0
(2)
Where Eelec denotes the electronics energy;

and mp are the amplifier energy;
fs mp is a constant.
To receive a k-bit message, the consumed energy

is:
(3)
Er (k ) k.Eelec
To fuse m messages with k-bit, the consumed
energy is:
(4)
E fs (m, k ) m.k.E fusion
To gather and transmit a frame of data, the total

energy dissipation of a node is:
En m.k.Eelec m.k.E fusion k.Eelec k. fs .d 2 (5)
d) The Average delay at the network layer which
is crucial in case of real time applications, the
average delay is computed as follows:
Avg. End to end delay = Total transmission delays
of all the received packets/ No of packets received
Where,
The transmission delay of a packet = Total time
when a packet is received at the server - Time
when a packet is transmitted at the client.
e) Average Jitter is calculated as the total packet
jitter for all received packets divided by number of
packets received minus one where Packet jitter is
the Transmission delay of the current packet
minus Transmission delay of the previous packet
[8, 9, 10, 11].
f) Packet drops: The number of packets dropped
at the Mac layer is due to link failure or
congestion or collisions. The throughput is
directly affected by the number of packets
dropped in the MAC layer and hence those
statistics are also included.
Figure 5 shows the throughput of the entire
network, which is better for multi-tree than MST
for all the protocols. We know that in case of a
tree based communication the chances of
contention and collision increases due to the single
available path to transmit. Hence we see that in an
MST the throughput at the application layer is
lesser due to non availability of the channel.
However, in case of multi-tree we have different
routes for different types of data thereby
decreasing the contention for singe route to the
base station or sink. Figure 6 shows load at
transport layer due to unicast, which is less for
multi-tree than MST for all the used protocols,
indicating that our approach observes less traffic
load. We include this graph to confirm that our
multi-tree reduces the individual traffic load that
needs to be processed by each node in contrast to
the total load processed in case of a uniform MST.
443
Figure 8 and 9 shows that multi-tree consumes
less energy as compared to the corresponding
MST in transmission as well as in receiving for
AODV, DYMO and Fisheye while for Bellman
Ford protocol both tree consumes almost same
energy.
Figure 5. Throughput (bits/second)
Figure 8. Energy consumed (in mWh) in Receive mode
Figure 6. Unicast Offered load at the transport layer

(bits/second)
Figure 7 shows that average carried load per node

is lesser in case of multi-tree for AODV, DYMO
and Fisheye protocols for the reasons already
mentioned. However, for Bellman Ford it is more
for multi-tree owing to the static nature of protocol
while assignment of the routes. This implies that
the total energy consumption per packet in case of
multi-tree is lesser in comparison to the similar
MST backbone.
Figure 9. Energy consumed (in mWh) in Transmit mode
Figure 10 and 11 show that average delay and

jitter is lesser in case of multi-tree for AODV,
DYMO and negligible for Bellman Ford and
Fisheye for both the trees. The reduction in delay
is noticed due to the fact that multi tree reduces
the contention for single routes and hence the
packets are sent without prolonged wait times.
Since we use the ZIGBEE protocol the packets
sent in CAP is more hence a lesser delay and
corresponding jitter is observed.
Figure 7. Average Carried load per node (bits/second)
444
4 CONCLUSIONS
Figure 10. Average Delay (in seconds) per node for the
protocols
Figure 11. Average Jitter (in seconds) per node for the
protocols
Figure 12 shows that the total number of packets

dropped at MAC layer are lesser in case of multitree for AODV, DYMO and Fisheye and almost
same for Bellman Ford for both trees.
From the simulation results we validate our claim

as to the reduction in the carried load which
affects the delay and the corresponding energy
consumption. The load criteria is not a very
important
parameter
for
analyzing
the
performability of a sensor network as the data
communicated in structural health monitoring
applications is usually redundant. But we see that
since the congestion and collisions are lesser due
to reduced traffic, we observe lesser number of
packet losses at the MAC layer. The results show
the performance of the multi-tree approach in all
the layers of the communication namely;
throughput at the application layer, total unicast
load at network layer in IP UDP , delay and jitter.
We also include the results of the MAC layer in
terms of the number of packet losses. Also the
energy comparison depicts that the multi tree
shows significant saving under the same network
scenario in comparison to the MST backbone
based communication architecture. Hence it can
be used for real time sensor network applications
that require faster communication. The following
results depict the prospects of using multiple
sensors for reduced data latency. It can further be
addressed to improve the reliability of networks.
7 REFERENCES
1.
2.
3.
4.
Figure 12. Number of packets dropped at MAC layer

5.
Hung-Chin, J., Lee, H., Huang, J.: Optimal Energy

Consumption for Wireless Sensor Networks, JCIS,
(2006).
Dengpan, Z., Gao, J.: Maintaining approximate
minimum steiner tree and k-center for mobile agents in
a sensor network, In INFOCOM Proceedings IEEE, 1-5,
(2010).
Georgios, C., Zeinalipour-Yazti, D., Gunopulos, D.:
Minimum-hot-spot query trees for wireless sensor
networks, Proceedings of the Ninth ACM International
Workshop on Data Engineering for Wireless and
Mobile Access, 33-40, (2010).
Yingchi, M., Xiaofang, L., Yi, L.: Workload-based
Query Routing Tree Algorithm in Wireless Sensor
Networks, International Conference on Computational
Intelligence and Software Engineering (CiSE),1-4,
(2010).
Kai, L., Zhao, H., Yin, Z., Luo, D.: An Adaptive
Classified Data Aggregation Arithmetic for Wireless
Sensor Networks, International Conference on Wireless
445
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
Communications, Networking and Mobile Computing,

WiCom., IEEE, 2739-2742, (2007).
Fariborzi, H., Moghavvemi, M.: EAMTR: energy aware
multi-tree routing for wireless sensor networks, IET
communications 3, no. 5, 733-739, (2009).
Bencan, G., Jiang, T.: A tree-based routing protocol in
wireless sensor networks, International Conference on
Electrical and Control Engineering (ICECE), IEEE,
5729-5732, (2011).
Qiang, T. L., Yu, Q.F.: A Distributed Slot Assignment
Algorithm with Minimum Jitter and Delay Guarantee
for Real Time Applications for Wireless Sensor
Networks, 12th IEEE International Conference on High
Performance Computing and Communications (HPCC),
383-390, (2010).
Kan, Y., Gidlund, M., Akerbergy, J., Bjorkman, M.:
Low Jitter Scheduling for Industrial Wireless Sensor
and Actuator Networks, 39th Annual Conference of
Industrial Electronics Society, IECON, IEEE, 55945599, (2013).
Nghia, L. H., Zalyubovskiy, V., Choo, H.: Delayminimized Energy-efficient Data Aggregation in
Wireless Sensor Networks, International Conference on
Cyber-Enabled Distributed Computing and Knowledge
Discovery (CyberC), IEEE, 401-407, (2012).
Joohwan, K., Lin, X., Shroff, N. B., Sinha, P.:
Minimizing delay and maximizing lifetime for wireless
sensor networks with anycast, IEEE/ACM Transactions
on Networking (TON) 18, no. 2, 515-528 (2010).
Daniele, M., Ortolani, M., Re, G. L.: A Network
Protocol to Enhance Robustness in Tree-Based WSNs
Using Data Aggregation, Internatonal Conference on
Mobile Adhoc and Sensor Systems, (MASS 2007),
IEEE, 1-4, (2007).
Zusheng, Z., Yu, F.: Performance analysis of clusterbased and tree-based routing protocols for wireless
sensor networks, International Conference on
Communications and Mobile Computing (CMC), vol. 1,
IEEE, 418-422, (2010).
Mohajerzadeh, A. H., Mohammad, H. Y., Zahra E.:
Tree based energy efficient and congestion aware
routing protocol for wireless sensor networks, 11th
IEEE Singapore International Conference on
Communication Systems, (ICCS), 1707-1711, (2008).
Jaekwang, K., Lee, J.: ViTAMin: A Virtual Backbone
Tree Algorithm for Minimal energy consumption in
wireless sensor network routing, International
Conference on Information Networking (ICOIN) IEEE,
144-149, (2012).
Johansson, T., Evgeny O., Lenka C.: Interference aware
construction of multi-and convergecast trees in wireless
sensor networks, Next Generation Teletraffic and
Wired/Wireless Advanced Networking, Springer Berlin
Heidelberg, 72-87, (2008).
Chakraborty, S., Nandi, S., Karmakar, S.: A Tree-Based
Local Repairing Approach for Increasing Lifetime of
Query Driven WSN, 14th International Conference on
Computational Science and Engineering (CSE), IEEE,

475-482, (2011).
18. Patel, D., Chawla, B., Parekh, C.: Energy Aware and
Link Quality Based Routing in Wireless Sensor
Networks under TinyOS-2. x., International Journal of
Modern Engineering Research Vol.3, Issue.3, 13571365, (2013).
19. Colesanti, U., Santini, S., The Collection Tree Protocol
for the Castalia wireless sensor networks simulator,
ETH Zurich, Zurich, Switzerland (2011).
20. Gnawali, O., Rodrigo F., Kyle J., David M., Philip L.:
Collection tree protocol, Proceedings of the 7th ACM
Conference on Embedded Networked Sensor Systems,
1-14, (2009).
446
Reliability and Accuracy of Neural Networks for Exchange Rate

David Semaan1, Atef Harb1 and Abdallah Kassem2 (IEEE Member)
1
Faculty of Business Administration & Economics, Graduate Division
2
Department of Electrical and Computer and Communication Engineering
Notre Dame University Louaize
Zouk Mosbeh
david.semaan@hotmail.com, aharb| akassem@ndu.edu.lb
ABSTRACT
Most exchange rates are volatile and mainly rely on
the principle of supply and demand. Millions of people
around the world are influenced, one way or another,
by the variation in exchange rates. In this research we
demonstrate that the Artificial Intelligence, specifically
Artificial Neural Networks (ANN), can improve the
accuracy of forecasting exchange rates compared to
statistical techniques such as regression. When we
compared the results from regression and artificial
neural network, it was clear that the ANN
outperformed regression in forecasting exchange rates.
Moreover, it became clear that using ANNs instead of
regression for forecasting exchange rates is rewarding
and necessary because the average error given by an
ANN is smaller than the average error given by
regression. Accuracy in forecasting became a major
issue and not a minor detail. It was the combination
between Artificial Intelligence and Macro Economics
that made these two models come into reality, making
it possible to use computer sciences and engineering
fields in the service of an economical problem.
The results of our research showed that using ANN
with the right parameters and variables rather than
using a regression model will yield a result with a
lower error margin.
KEYWORDS
Artificial Neural Network, Regression, Exchange Rate,
Forecasting, Inflation, Interest Rate, Multilayer
Perceptron;
I. INTRODUCTION
Most exchange rates are volatile and mainly rely
on the principle of supply and demand. Millions of
people around the world are influenced, one way
or another, by the variation in exchange rates. All
international companies rely on exchange rates to

make everyday business decisions such as buying
goods, opening worldwide branches, investing in
foreign countries, etc Forecasting is an essential
discipline in planning and running a business.
Success hugely depends on getting those forecasts
right. As we know, however, that the future is
highly uncertain.
The purpose of this study is to demonstrate the
power of Artificial Intelligence, specifically ANN,
in forecasting exchange rates, and comparing the
results to the statistical techniques results.
For several decades, many techniques were used to
forecast exchanges rates; some of them were
proven to be better than others [1]. Specifically,
regression was a very good technique to be
considered [2]. But recently, artificial intelligence,
and especially artificial neural networks, became a
powerful tool for forecasting exchange rates [3].
At this stage, it is clear that the main research
question about which technique is better in
forecasting exchange rates.
Section II shows the importance of risk
management for IT projects, where a good risk
management plan can ensure the success of a
project. In section III we can find an overview of
ANNs and an overview about statistical techniques
and regression. Section IV presents an overview of
the state of knowledge concerning forecasting,
ANNs, and regression. The procedures and the
methodology of the system implementation are
described in Section V. Section VI highlights the
system specifications and presents preliminary
results. Finally a conclusion is given in section
VII.
447
II.
IT PROJECTS RISK MANAGEMENT
Forecasting exchange rate can fail without good

managing IT system, it may have some risks. IT
can help to manage with a minimum risk and error
for forecasting exchange rate. [4]
The Risk Management process is divided into six
steps:
1) Goal Definition: review the stated goals for
the project, refine them and define implicit
goals and constraints explicitly.
2) Identification: identify potential threats to
the project using multiple approaches.
3) Analysis: classify risks, complete risk
scenarios, estimate risk effects and estimate
probabilities and utility losses for risk
scenarios.
4) Control and Planning: select the most
important risks, propose controlling actions
and the actions for implementation.
5) Control: implement the risk controlling
actions.
6) Monitoring: monitor the risk situation.
Organizations need to keep IT projects on
schedule and costs under control. However,
organizations must also encourage managers to
respond to changing business needs and exploit
technological
opportunities
before
their
competitors do so.
III. OVERVIEW OF ARTIFICIAL NEURAL
NETWORKS AND REGRESSION
A) Artificial Neural Networks (ANN)
An Artificial Neural Networks purpose is to
recognize patterns in the available data by
learning. After training the network with existing
data samples, it becomes able to predict or forecast
similar patterns or behavior in future data. In
general, ANNs are more attractive and useful than
traditional techniques because of their unique
properties [5]. In many cases, ANNs offer a better
alternative than algorithmic functions and expert
systems [6]. In every situation where we cannot
define the exact relation between inputs and
outputs, or when the input-output relation is
complex, ANNs can give an accurate
approximation for this relation [7]. We can
imagine an artificial neural network as a black box
that can predict an output pattern when it

recognizes a given input pattern. Inside ANNs, we
can find a bunch of processing elements or PEs,
connected between them with adjustable weights.
Figure 1 shows an artificial neural networks
structure.
The neural network must first be "trained" by
having it process a large number of input patterns
and showing it what output resulted from each
input pattern. Once trained, the neural network is
able to recognize similarities when presented with
a new input pattern, resulting in a predicted output
pattern [7].
Fig. 1. Neural Network Structure
There are many types of artificial neural

networks, and the most popular one is a Multilayer
Perceptron or MLP. MLP always needs a target
output in order to learn, thats why it is called a
supervised network [7]. An MLP set a relation
between inputs and the desired outputs, and it does
so by using data that is previously stored. For an
MLP to learn, it uses an algorithm called back
propagation algorithm. By using this algorithm,
the previously stored input data is fed into the
neural network many times, and at each time the
actual output is compared to the stored output, and
an error is calculated based on the difference
between these outputs. This error is then used to
adjust the weights of the connections between
neurons. After each iteration, the error magnitude
should decrease (refer to figure 2) until it is less
than a predefined number called error threshold.
At the end, the calculated output is closer to the
desired output.
The power of ANNs is shown in the fact that
any complex pattern in a huge amount of data can
be extracted by the Artificial Neural Network
without the need to know the exact mathematical
relationship between the dependent and
independent variables [8][9]. This is exactly the
case in forecasting exchange rates, where huge
448
amounts of data are to be treated and analyzed in
order to reach a final number that represents a
forecasted exchange rate.
trained neural network is used as a Fed Forward

Controller.
C) Statistical Regression Models
Fig. 2. Neural Network Learning Model
B) Useful Applications of ANNs

Besides being used as a forecasting method,
ANN is widely used for many applications. One of
these methods is Automatic Traffic Signs
Recognition [10]. The steps of traffic signs
recognition are:
1- Image processing: algorithms that are applied
on the pixels values of the image to perform a
specific process, such as enhancement and
restoration.
2- Image features extraction: There are mainly two
types of features; low level and high level
features. Low level features can be extracted
directly from the image, like edges, lines,
textures, etc. A set of high level features can be
extracted from low level features, which are
useful in image understanding and content
recognition.
3- Image Segmentation: Dividing the image into
non-overlapping, homogenous, and connected
regions.
4- Traffic Sign Image Recognition: The resultant
image will pass through the recognition process,
in which the image is entered into a trained
neural network and the decision is made
according to the result of the NN.
Another ANN application is a Controller for
Robotic Manipulator [11], where ANN is used for
inverse neural controller which is used to control
the operation of six Degree Of Freedom (6DOF)
robotic manipulator. A neural network is trained to
identify (learn) the inverse of the plant dynamic
with the aid of a set of training data then the
Regression can be considered as a statistical tool

that investigates the relation between one
dependent variable and one or more independent
variables. The relation, if it exists, proves the
cause/effect relation between the variables. For
example, the effect of interest rate on exchange
rates, or the effect of fiscal policy on goods prices
[13]. The primary focus is on the analysis of the
relation between one or more independent
variables and one dependent variable. Regression
tries to show how a dependent variable changes in
response to the change of one independent
variable, while keeping the other independent
variables are kept unchanged. Regression analysis
tries to set the relation between independent
variables and the dependent variable through a
function that we call the regression function. In
practice, regression is used for forecasting and
prediction. It can also be used to show which of the
independent variables is greatly related to the
dependent variable, and by which amount. In
addition, the performance of regression depends on
the methods form that generates data. In most of
the cases, there are some assumptions that should
be made to use regression (e.g. linearity of data,
normal distribution, etc). To obtain better
results, these assumptions should be respected.
IV. STATE OF KNOWLEDGE
Since the widespread introduction of floating
exchange rate regimes amongst the major
currencies in the early 1970s, the problem of
correctly anticipating exchange rate fluctuations is
one, and corporate treasurers of companies having
any international dealings have had to face it in
order to manage successfully the exchange risk
inherent in international contracts.
A lot of entities are interested in forecasting at
least the direction of some exchange rates. Using
an exchange rate forecasting model that can guide
businesses and traders in their decision making
can be effective and essential in order to maximize
profit and minimize the risk in any business.
449
Because all known forecasting exchange rate
models have approximately the same efficiency,
we can find a lot of them. This shows the amount
of difficulty and complexity of generating an
efficient and reliable exchange rate forecasting
model.
In his book entitled Exchange Rate
Determination: Models and Strategies for
Exchange Rate Forecasting, Michael Rosenberg
explained that globalization and cross-border
interaction contributed largely to redefining the
worldwide business arena. He mentioned as well
that success for the companies dealing with
business in the arena is the result of forecasting
and determining accurately and thoroughly the
exchange rates.
Authors in [14] worked on estimating and
selecting feedforward and recurrent networks in a
careful way, and they also wanted to evaluate the
forecasting performance of selected networks in
different periods, thats why they have proposed a
two-step procedure. Different forecasting results
were gathered, and they were not similar. Among
five series which were evaluated, in only two out
of them, networks with significant market timing
ability (sign predictions) and/or significantly lower
out-of-sample MSPE (relative to the random walk
model) were found. The forecasting performance
in the other remaining series is not as satisfying as
in these two. According to their results, PSC is
seen to be sensible in selecting networks and the
two-step procedure that had been used could be as
a standard network construction procedure in other
applications. The results reveal that nonlinearity in
exchange rates may be exploited in order to
improve both point and sign forecasts. Although
some of the results reported are quite motivating,
they provide only limited evidence supporting the
usefulness of neural network models.
Walczak [15] speaks about neural networks
saying that they have been shown to be a
promising tool for forecasting financial time
series. Several design factors influenced
significantly the accuracy of neural network
forecasts. These factors include selection of input
variables, architecture of the network, and quantity
of training data. The issues about input variable
selection and system architecture design have been
widely researched, but regarding the information
use in producing high-quality neural network

models the issue remains unclear since it has not
been adequately addressed. In this paper, the
effects of different sizes of training sample sets on
forecasting currency exchange rates are dealt with.
Future currency exchange rates can be forecasted
with 60% accuracy due to those neural networks
which are given an appropriate amount of
historical knowledge, while a worse forecasting
performance is shown due to those neural
networks trained on a larger training set.
Khashei and al. [16] proposed a hybrid model
that gives better results when there are incomplete
data sets. This hybrid model combines ANN with
fuzzy regression. This model was empirically
proven to give more accurate results in financial
forecasting. More, Yu and Huarng [17] applied
neural networks by implementing fuzzy time
series model. This method improved forecasting
accuracy.
Now that we had an overview about forecasting
exchange rates, statistical techniques, regression,
and ANNs, we are ready to investigate the
efficiency and reliability of ANNs in forecasting
exchange rates. This is accomplished by setting
two models for forecasting, one is using
regression, and the other is using an artificial
neural network. After that we need to test the
accuracy of each method by comparing error
margins from each method.
V. PROCEDURES AND METHODOLOGY
Forecasting exchange rates is not anymore a
theoretical field, but in fact it is based on empirical
experiments and findings. Therefore; our empirical
model that will be used needs a definition of the
model and the parameters or factors of the model:
Dependent Variable (model output) and
independent variables (model inputs).
A. Dependent Variables
The only dependent variable that we have is the
Euro/Dollar Exchange rate. When using
Regression, the exchange rate is supposed to be
the output of the equation generated by this
technique. When using Artificial Neural Network,
the exchange rate is the output of the MLP
(Multilayer Perceptron) network. Theoretically,
the exchange rate is the result of all the factors
450
previously mentioned (e.g. interest rate, oil price,
etc).
B. Independent Variables
The first, and most important independent
variable, is the interest rate. This variable is by far
the most influential factor that affects the
exchange rate. We are using to interest rates,
representing two independent variables: One Year
USD LIBOR, and One Year Euribor.
The second one is the inflation. It is well known
that inflation rates affect a countrys currency
value. We are using to inflation rates, representing
two independent variables: US Core Inflation Rate
and Euro Area Core Inflation Rate. When a
countrys inflation rate rises relatively to that of
another country, decreased exports and increased
imports depress the high-inflation countrys
currency. High inflation rates increases the foreign
exchange rates and hence weakens the local
currency. Notice here that we used core inflation
rates which excludes the prices of energy, to avoid
double counting because oil price is counted is an
independent variable.
Oil price is the third independent variable. In
general, oil prices have noticeable influence on
USD, and hence it has influence on the Euro/Dollar
exchange rate. Figure 3 shows the main factors
affecting the exchange rates used in this model.
Fig. 3. Factors affecting the exchange rates.
VI. SPECIFICATIONS AND PRIMARY

RESULTS
Two models are used: The first one uses
regression techniques, and the second one uses an
Artificial Neural Network. Both models are set

using SPSS. The models will serve many purposes:
To show the accuracy of each technique
concerning forecasting exchange rates.
To show the actual relationship between the
independent and dependent variables used, and
to show the influence of the independent
variables on our dependent variable (Exchange
rate).
To be able to make a comparison between the
two models concerning accuracy and fitness.
To test the ability of an Artificial Neural
Network to approximate the actual relationship
between the dependent and independent
variables. This relationship is in reality very
complex and practically impossible to be
represented in a clear mathematical equation.
A)
Model 1: Using Regression
The first model that we are using is based on

regression technique.
To use regression, first we need to define our
variables:
One dependent variable: Exchange Rate
Five independent variables: US Inflation,
EU Inflation, Oil Price, One Year
EURIBOR, and One Year USD LIBOR
After filling up all these variables, the result of
using regression can be shown and analyzed.
Tables I till IV show the Coefficients, Frequencies,
Model Summary and ANOVA tables, as taken
from SPSS.
From the table III, we notice that Adjusted R
Square=0.754, meaning that the independent
variables used in our model explain 75.4% of the
variability of the dependent variable (Exchange
Rate). Moreover, we notice that R= 0.873 or
87.3%, which means that the independent variables
correlate well with the dependent variable. In
addition, from the ANOVA table IV, we notice
that the Significance variable Sig. = 0, meaning
that the probability that the results are by random
chance is zero, so we can conclude that the model
is significant. By looking at the bell shaped
histogram in figure 4, we can conclude that the
data is normally distributed.
Table I. Coefficients table.
451
Coefficients
ExchangeRate = 1.111 + 0.060 * USInflation - 0.166 * EUInfaltion +
(1)
0.004 * OilPrice + 0.049 * Euribor1Year -
Model
1 (Constant)
USInflation
Unstandardized
Standard.
Std. Error
Beta
1.111
.025
.060
.010
0.020 * USDLibor1Year
.563
Sig.
44.071
.000
6.291
.000
EUInfaltion
-.166
.017
-.890
-9.579
.000
OilPrice
.004
.000
.882
18.227
.000
Euribor1Year
.049
.009
.401
5.514
.000
.007
-.221
-3.047
.003
USDLibor1Year -.020
Table IV. ANOVA table.

ANOVAa
Model
Sum of
Squares
Df
Mean
Square F
Sig.
Regression
2.220
.444
.000b
Residual
.690
126
.005
Total
2.910
131
81.104
a. Dependent Variable: ExchangeRate

a. Dependent Variable: ExchangeRate
b. Predictors: (Constant), USDLibor1Year, OilPrice, EUInfaltion,
Table II. Frequencies table.
Euribor1Year, USInflation
Statistics
USInfl
EUInfl
Oil
Euribor
USDLibor
Exchg.
ation
ation
Price
1Year
1Year
Rate
Valid
132
132
132
132
132
132
Missing
Mean
2.4000 2.1265
66.11
2.65656 2.50564
1.28
Std. Dev.
1.3993 .80113
29.37
1.2194
1.618903
.1490
Skewness
-.677
.230
.553
.681
-.696
-.892
Std. Error of
.211
.211
.211
.211
.211
.211
1.101
2.297
-.992
-.617
-.987
.693
.419
.419
.419
.419
.419
.419
Skewness
Kurtosis
Std. Error of
Kurtosis
Table III. Model Summary table.
Fig. 4. Histogram of the unstandardized residual.
Model Summary
Model
1
R
a
.873
Adjusted R
Std. Error of the
Square
Square
Estimate
.763
.754
.073988
a. Predictors: (Constant), USDLibor1Year, OilPrice, EUInfaltion,

Euribor1Year, USInflation
b. Dependent Variable: ExchangeRate
Now that we are sure that data are normally

distributed, we can proceed by testing the model.
From table I, we can build the equation that
relates the independent variables to the dependent
variable, as follows:
Note here that, in order to come up with this result,

we used only 132 sets of data, between the years
2002 and 2012.
From Equation (1), we can notice the following:
The parameter for USInflation is positive
(+0.06), meaning that when the US core
inflation increases, the US Dollar depreciates,
and the Euro Dollar Exchange rate increases.
This is a logical result.
The parameter for EUInflation is negative (0.166), meaning that when the EU core
inflation increases, the Euro depreciates, and
the Euro Dollar Exchange rate decreases. This
is also a logical result.
452
The parameter for OilPrice is positive
(+0.004), meaning that when the oil price
increases, the Euro Dollar Exchange rate
increases. This is a bit confusing but it can be
explained as follows: Since the United States
is the most important oil importer around the
world, and since its economy is highly related
to energy; it could be clearly said that any
rising in oil prices, would first affect or even
damage the U.S. economy.
The parameter for Euribor1Year is positive
(+0.049), meaning that when the Euro Zones
interest rates increase, the Euro appreciates,
and the Euro Dollar Exchange rate increases.
This is also a logical result.
The parameter for USDLibor1Year is
negative (-0.020), meaning that when the US
Dollars interest rates increase, the US Dollar
appreciates, and the Euro Dollar Exchange
rate decreases. This is also a logical result.
Table V shows the results of plugging the data of
2013 (first eight months) into our regression
equation.
neuron output layer as shown in figure 5. Under the

input layer, each input variable is assigned a
neuron. As well, under the input layer, each input
variable is fed to a neuron after a standardization
procedure that transforms the input values to the
[-1,1] interval. In addition to the input values, we
can find a constant input with a value of 1, this
input is called the bias. Under the hidden layer, the
value from the input neurons are multiplied by a
weight and added to yield to a combined value this
value is fed to a transfer function that outputs
another value which in turn is fed to the output
layer. Under the output layer, each value arriving
from the hidden layer is multiplied by a different
weight, and the resulting values are added to
produce a new value. This value is fed to a transfer
function that outputs a new value transferred to the
output of the network. In the MLP model, the same
set of data as in the first model, i.e. Oil price, US
core inflation, Euro Zone core inflation, One Year
EURIBOR, and One Year USDLIBOR as
independent variables, and the Euro/Dollar
Exchange Rate as the dependent variable are used.
Table V. Eight Months Regression results

Date
Oil
Price
Eur
ibor
1Y
USD
Libor
1Y
Exch.
Rate
Reg.
Rate
Regr.
%
Err.
USInf.
EUInf.
Jan-13
1.6
1.97
105.04
0.5753
0.814
1.33
1.31
1.38
Feb-13
1.84
107.66
0.5942
0.762
1.34
1.37
2.57
Mar-13
1.5
1.73
102.61
0.545
0.735
1.30
1.34
3.09
Apr-13
May13
1.1
1.18
98.85
0.5284
0.717
1.30
1.39
6.57
1.4
1.42
99.35
0.4838
0.694
1.30
1.37
5.30
Jun-13
1.8
1.6
99.74
0.5071
0.684
1.32
1.36
3.49
Jul-13
1.61
105.21
0.5254
0.684
1.31
1.40
6.80
Fig. 5. MLP-ANN Architecture.
Aug-13
1.5
1.34
108.06
0.5423
0.668
1.33
1.42
6.95
After filling up the Dependent variable

(ExchangeRate), the Covariates or independent
variables, and all the other required parameters; we
choose the partitions tab to add the following
parameters:
Training: 100; Test: 32; Holdout: 8
The holdout option is very important. It means
that one or more set of data will not be included in
the training of the network. Instead, it will be used
after training to test the efficiency of the network
[18].
Av.%
Error
4.52
Para
meter
0.06
B)
Para
meter
-0.17
Para
meter
0.004
Para
meter
0.049
Para
meter
-0.02
Cons
tant
1.11
Model 2: Using Artificial Neural Network
The second model uses one famous type of

Artificial Neural Networks, known as a Multilayer
Perceptron (MLP-ANN). It has a five neurons
input layer, a three neurons hidden layer, and a one
453
There are many different activation functions,
some of the most commonly used are:
Threshold:
1 if x t 0
g(x) =
(2)
0 if x t 0
Sigmoid:
g(x) =
(3)
1 e 2s( x t )
Hyperbolic tangent:
g(x) = tanh(s(x t))
sinh( s ( x t ))
cosh(s ( x t ))
e s( x t ) e s( x t )
e s( x t ) e s( x t )
e 2( s ( x t )) 1
(4)
e 2( s ( x t )) 1
Where is the value that pushes the center of

the activation function away from zero and is
a steepness parameter. Sigmoid and hyperbolic
tangent are both smooth differentiable
functions, with very similar graphs, the only
major difference is that hyperbolic tangent has
output that ranges from -1 to 1 and sigmoid has
output that ranges from 0 to 1. A graph of a
sigmoid function is given in figure 6, to
illustrate how the activation function looks like.
Fig. 6. A graph of a sigmoid function with s= 0.5 and t=0
The t parameter in an artificial neuron can be seen

as the amount of incoming pulses needed to
activate a real neuron. This parameter, together
with the weights, is the parameter adjusted when
the neuron learns.
In addition to the activation function, there are
many parameters that should be defined for ANNs,
some of them are:
Dependent variables
Nominal
Ordinal
Scale
The method for rescaling covariates
Standardized
Normalized
Adjusted Normalized
None
Partition
Training
Test
Holdout
Architecture
Number of Hidden Layers
Output layer activation function
Identity
Softmax
Hyperbolic tangent
Sigmoid
Activation Function: The activation function
"links" the weighted sums of units in a layer to
the values of units in the succeeding layer
Hyperbolic tangent
Sigmoid
Type of Training: The training type
determines how the network processes the
records
Batch
Online
Mini-batch
Optimization Algorithm: This is the method
used to estimate the synaptic weights
Scaled conjugate gradient
Gradient descent
Training options for the scaled conjugate
gradient algorithm
Initial Lambda
Initial Sigma
Interval Center and Interval Offset
Training options for the gradient descent
algorithm
Initial Learning Rate
Lower Boundary of Learning Rate
Momentum
Learning rate reduction, in Epochs
Stopping Rules: These are the rules that
determine when to stop training the neural
network
Maximum steps without a decrease in
error
Maximum training time
454
Maximum Training Epochs
Minimum relative change in training error
Minimum relative change in training error
ratio
When using an MLP Neural Network, there is an
important option called Independent variable
importance analysis. This option allows us to
evaluate each independent variable according to its
importance and influence on the dependent
variable. It performs a sensitivity analysis, which
computes the importance of each predictor in
determining the neural network. The analysis is
based on the combined training and testing samples
or only on the training sample if there is no testing
sample. This creates a table and a chart displaying
the importance and normalized importance for each
predictor. Note that sensitivity analysis is
computationally expensive and time-consuming if
there are large numbers of predictors or cases.
Figure 7 represents the normalized importance of
the independent variables.
The oil price has the least importance, its

influence on exchange rate is not that important,
although it still has some influence.
What is the significance of the importance analysis
of the independent variables?
The importance analysis shows the importance
and the influence of each independent variable on
the dependent variable. As mentioned in chapters
two and three the most important and influential
factor on exchange rate is the interest rate. The
result from the importance analysis confirmed that
fact and it became clear that the leader variable in
determining exchange rate is the interest rate.
Moreover, inflation came second in the importance
analysis, which also confirms the theories of
macro-economic that clearly state that inflation
plays a major role in the exchange rate
determination.
Above all this, the importance analysis showed that
our Artificial Neural Network model was actually
able to approximate the real complex relationship
between the independent and dependent variables.
So comparing the results of the regression model
and the ANN model, we can notice that the ANN
model was able to give a better approximation
model than the regression model according to the
importance analysis of the independent variables
for both models.
Table VI shows the average error resulting from
using the two models, using data sets for eight
months (January 2013 to August 2013).
Table VI. Resulting average error
Date
Fig. 7. Normalized importance of the independent variables
From this table and chart, we can notice the

following:
It is clear that the interest rates are the most
important factors in determining the exchange
rate. This conclusion is conformant with our
previous assumptions, and also with the
economic theories.
Inflation rate comes second in importance after
the interest rate, but with less importance and
influence.
Reg. Rate
Regr. % Err.
ANN Rate
ANN % Err.
Jan-13
1.31
1.38
1.29
3.30
Feb-13
1.37
2.57
1.31
2.21
Mar-13
1.34
3.09
1.30
0.04
Apr-13
1.39
6.57
1.32
1.57
May-13
1.37
5.30
1.31
0.76
Jun-13
1.36
3.49
1.31
0.94
Jul-13
1.40
6.80
1.32
0.85
Aug-13
Aver% Error
1.42
6.95
4.52
1.34
0.69
1.29
Note that the same parameters and constant from

table V are used in table VI. From this table, we
can notice the following:
455
Concerning the Regression Model, the
average percentage error over the 8 sets of
data is 4.52%.
Concerning the ANN Model, the average
percentage error over the 8 sets of data is
1.29%. The ANN model is very accurate, and
it was able to approximate the relation
between the dependent and independent
variable within a 1.29% average error margin.
This error margin is considered very low and
the system is hence reliable for forecasting.
Now lets calculate the average error for both
methods. Referring to the tables, we can calculate
the error as follows:
For the regression model, we can calculate the
percent error for each month between January
2002 and December 2012 using the following
formula:
Regression Monthly Error =
ABS(ExchangeRate - Regression)
ExchangeRate
(5)
The average error of all 132 data sets between

January 2002 and December 2012, is:
Average Regression Error = 4.69%
For the ANN model, we can calculate the

percent error for each month between January
2002 and December 2012 using the following
formula:
ANN Monthy Error =
ABS(ExchangeRate - ANN)
ExchangeRa te
(6)
And the average error for the same data set is:
Average ANN Error = 2.17%
In addition, according to the eight data sets used

for testing both shown in table VI:
Average Regression Error from testing samples = 4.52%
Average ANN Error from testing samples = 1.29%
So the ANN model gave lower error margin, and

hence more accurate results!
C)
Whats new in this research
Forecasting exchange rates has been the target
of many researches and financial analysts since the
70s of the last century. Thousands of papers and
researches were written, and researches were more
and more advanced. On the other hand, Artificial
Intelligence is still considered a fresh field, with
promising results in many areas. Forecasting

exchange rates is still considered a challenge
nowadays. Combining artificial intelligence with
financial issues such as forecasting exchange rates
requires enough knowledge in both fields.
The originality of this research is the actual
combination of a financial issue (i.e. forecasting
exchange rates) with a methodology that can be
considered as a software mimic of the human
brain. Many points can be considered as new or
original in this research, as follows:
The uniqueness of the comparison: To the
best of our knowledge, it is the first time that
an actual comparison is made between the
two techniques (regression versus artificial
neural network). Many researches were done
before; discussing one of these two
techniques, but it is the first time that an
actual, empirical, and realistic comparison
was made based on a discrete model for both
techniques.
The uniqueness of the model: To the best of
our knowledge, the two models set in this
research are the result of our own work and
research; the two models were not partly or
entirely used before.
The uniqueness of the parameters used in the
models: Using an artificial neural network in
an effective way requires an exhaustive fine
tuning that should be done by a professional
who knows what to change and how to
change. The parameters used in an artificial
neural network include, but are not limited
to, the following:
o The type of the neural network that best
fits the problem at hand (e.g. Multi
Layer Perceptron versus Radial Basis
Function).
o The number and type of the input layers.
o The number of the hidden layers.
o The number of neuron in each hidden
layer.
o The type of training.
o The initial values of the inner
parameters.
o The activation function.
o The optimization algorithm.
o The network structure.
456
o The network performance.
All these parameters, in addition to many others,
have to be carefully selected and fine-tuned in
order to get an optimized result. The fine tuning
effectively consumes a lot of time, using trial and
error, and experimenting many combinations.
VII. CONCLUSION
A comparison between two methods the
artificial neural network and the regression method
in forecasting exchange rates is detailed in this
paper. It was a serious effort toward finding an
adequate system that can forecast the
Euro/Dollar exchange rate with an acceptable error
margin. Both methods were set according to our
thorough research in the fields of forecasting,
macro-economics, statistical techniques, and
artificial intelligence.
The obtained result shows that exchange rates
can be forecasted with a minimum margin of error,
and therefore our ANN model can be used in
forecasting. It can have an impact on many
economical aspects. International companies can
no longer pretend that they have no idea about the
future rate of currencies.
This implementation also has many advantages.
First, future international transactions will become
safer and less prone to exchange rate fluctuations
since it is now feasible to know the trend of
currencies. Second, international companies will
suffer from fewer losses when doing international
transactions or when signing future contracts in
different international currencies. This all leads to
more profit for these companies.
On the other hand, this method has a slight
disadvantage because companies and individuals
might completely rely on such methods in order to
forecast exchange rates, disregarding other
external factors that may affect currency rates. To
prevent such inconvenience, fine tuning and
enhancements to this method are essential to
minimize the error margin and maximize
accuracy.
3. Pacelli V., Bevilacqua V. and Azzollini M., An Artificial

Neural Network Model to Forecast Exchange Rates", in Journal
of Intelligent Learning Systems and Applications, Vol. 3, n.
2/2011, pp. 57-69, 2011.
4. Amine NEHARI TALET, Razali MAT-ZIN and Maaradj
HOUARI, "Risk Management and Information Technology
Projects", International Journal of Digital Information and
Wireless Communications (IJDIWC), Vol. 4, No. 1, pp. 1-9,
2014.
5. Andreou, A.S., Georgopoulos, E.F., and Likothanassis, S.D.,
"Exchange rate forecasting: A hybrid algorithm based on
genetically optimized adaptive neural network", Computational
Economics, vol. 20, no. 3, pp. 191-210, 2002.
6. Mohamad Hassoun, "Fundamentals of Artificial Neural
Networks, Massachusetts Institute of Technology, MIT Press,
1995.
7. Delashmit W.H. and Manry M.T., Recent developments in
multilayer perceptron neural networks, Proceedings of the 7th
Annual Memphis Area Engineering and Science Conference
(MAESC 2005), 2005.
8. BingXiang Liu; Hua Wang; Xiang Cheng, "Exchange Rate
Forecasting Method Based on Particle Swarm Optimization and
Probabilistic Neural Network Model," International Conference
on Network Computing and Information Security (NCIS), ,
vol.1, pp.288-292, 2011.
9. Gill, S.S., Gill, A.K. and Goel, N., Indian currency exchange
rate forecasting using neural networks, IEEE, International
Conference on Advanced Management Science, pp. 196-199,
2010.
10. Mohammad A. N. Al-Azawi, "Neural Network Based
Automatic Traffic Signs Recognition", International Journal of
Digital Information and Wireless Communications (IJDIWC),
Vol. 1, No. 4, pp. 753-766, 2011.
11. Baker Al-Bahri and Zainab Al-Qurashi, "Practical Neural
Controller for Robotic Manipulator", International Journal of
Digital Information and Wireless Communications (IJDIWC),
Vol. 2, No. 1, pp. 33-39, 2012.
12. Levich, R.M. International Financial Markets: Prices and
Policies, second edition. McGraw-Hill/Irwin, 2001.
13. Chung-Ming, K., & Liu, T., Forecasting Exchange Rates
Using Feedforward and recurrent Networks, The Journal of
Applied Econometrics, Vol. 10, pp. 347-364, 1995.
14. Walczak, S., An Empirical Analysis of Data Requirements for
Financial Forecasting with Neural Networks, Journal of
Management Information Systems, Vol. 17, No. 4, pp. 203-222,
2001.
15. Khashei, M., Hejazi S.R., Bijari, M."A new hybrid artificial
neural networks and fuzzy regression model for time series
forecasting", Journal of Fuzzy Sets and Systems, Vol. 159,
Iss.7, pp. 769-786, 2008.
16. Yu T.H. and Huarng K., "A neural network-based fuzzy time
series model to improve forecasting", An International Journal
of Expert Systems with Applications, Vol. 37, Iss.4, pp. 33663372, 2010.
17. De Grauwe, P., Exchange Rate Economics, Where Do We
Stand? CESifo seminar series. MIT Press, Massachusetts,
USA., 2005.
18. David Semaan, Atef Harb and Abdallah Kassem, Forecasting
Exchange Rates: Artificial Neural Networks Vs Regression,
will appear in the Third International Conference on eTechnologies and Networks for Development (ICeND2014),
2014.
REFERENCES
1. Pacelli, V., "Forecasting Exchange Rates: a Comparative
Analysis", International Journal of Business and Social Science,
vol. 3, no. 10, 2012.
2. Sandler, Ralph, Forecasting Exchange Rates Within the
Context of an International Financial Management Course,
Advances in Financial Education. Vol. 6, pp. 119-135, 2008.
457
Analytical Approach of Cost-Effective and Secure SIP-based Mobility Management

Scheme for NEMO Environments
Chulhee Cho1, Jae-Young Choi1, Jun-Dong Cho2, and Jongpil Jeong2,*
1
College of Information and Communication Engineering
Sungkyunkwan University, Suwon, Korea
2
Department of Human ICT Convergence (*Corresponding Author)
Sungkyunkwan University, Suwon, Korea
tgb017@nate.com, {jaeychoi,jdcho,jpjeong}@skku.edu
ABSTRACT
The mobile Virtual Private Network (MVPN) of the
Internet Engineering Task Force (IETF) is not designed
to support NEtwork MObility (NEMO) and is not
suitable for real-time applications. Therefore,
architecture and protocols to support VPN in NEMO
are needed. In this paper, we propose a cost-effective
and secure mobility management scheme (SeSIP) that
is based on session initiation protocol (SIP) and
designed for real-time applications with VPN. Our
scheme to support MVPN in NEMO enables the
session to be well maintained during movement of the
entire network. Further, in order to reduce the
authentication delay time in handoff operations, the
signaling time which occurs to maintain the session is
shortened through our proposed handoff scheme which
adopts authentication using HMAC-based one-time
password (HOTP). Our performance analysis results
show our proposed scheme provides improvement in
average handoff performance time relative to existing
schemes.
KEYWORDS
NEMO, MVPN, SIP, HOTP, Mobility Management.
1 INTRODUCTION
As the coverage area of wireless LAN (WLAN)
expands, the demand from users is growing for
access to the Internet anytime and anywhere. To
satisfy this requirement, technologies that enable
access to the Internet on trains, busses, ships, and
other modes of transportations have come into the
limelight. One such technology is NEMO, an IP
network mobility technology [1-3]. The acronyms
used in this paper is listed in Table 1. NEMO
enables Internet connection service to be provided

from the mobile router (MR) with all the nodes
inside the network not recognizing the mobility, a
standardization that is making progress in IETF
based on IPv6. NEMO provides mobility service
through direct links to the Internet network
without passing through other networks. Thus, it
can be applied to telematics, Personal Area
Networks (PAN), Ad-hoc networks, etc., as well
as providing various means of mobility. The IETF
NEMO working group has completed several
RFCs to enable a network to move from one
location to another location while still maintaining
its local nodes ongoing sessions. For example, a
NEMO VPN can be used in public safety, where
wireless devices in a police patrol car can access
to the criminal databases, driver license and
vehicle registration databases, or other services in
the dispatch center as the car travels between
different subnets. Similar type of services can also
be used in ambulance or mobile medical car,
where various wireless devices or sensors are
deployed inside the car.
Security has recently emerged as an important
issue for the Internet, and virtual private network
(VPN) was developed to ensure stability in user
communications between the Internet and the
intranet. VPN service in NEMO has wide-ranging
applications, providing stable access to the
intranet for mobile networks. For example, NEMO
VPN enables access to the criminal, drivers
license, and car registration databases from a
police patrol car via the mobile device, thereby
helping to increase public safety. However, a
method for providing VPN service has yet to be
identified in the NEMO working group of IETF.
458
Although IETF proposed a VPN architecture that
supports mobility, this solution did not consider
mobile equipment groups and is only applicable
for a single node, and, furthermore, it is based on
MIP, which is not suited for real-time applications.
MVPN of IETF uses one IPSec [4] tunnels and
two MIP tunnels. These three tunnels are major
contributors to overhead during the real-time
packet transfer. Thus, a new architecture and
protocol are required to support the MVPN in safe
NEMO. In addition, the complexity of the
authentication procedure and multiple signaling
messages that may occur in various nodes due to
the movement of the mobile equipment group are
also major contributors to overhead.
This paper proposes a Cost Effective and
Secure Mobility Management Scheme (SeSIP)
based on the SIP (Session Initiation Protocol)
which is suitable for real-time application on
MVPN and which shortens the signaling time.
This design maintains the session continuously as
the overall network moves. It integrates SIP-based
MVPN and NEMO to provide efficient group
mobility for high security and real-time services.
Additionally, all SIP clients can directly
communicate with each other, bypassing the
mobile agent such as the Home Agent (HA) in
MIP. Thus, the path is optimized. This is useful
for real-time applications such as IP-based voice
communications (VoIP) and video streaming, and
it does not require an IPSec tunnel or MIP tunnel.
Hence, a single NEMO VPN gateway can support
an entire mobile network upon the address request
of a mobile network that has changed its
connection location address, resulting in
considerable reduction of signaling overhead.
Moreover, this approach reduces the signaling
numbers since all CNs connection addresses are
combined in a URL list and integrated in a single
INVITE message for transfer. Further, this design
adopts an authentication method based on HMACbased One Time Password (HOTP) [5] to shorten
the authentication time, a significant element of
delay during hand-off, thereby improving the
ongoing signaling time to maintain the session.
Moreover, this approach integrates the generation
signals of multiple nodes inside the mobile
network to reduce signaling time.
Table 1. Parameters for Handoff Signalling Cost.

Acronym
Description
ALG
Application Level Gateway
AVP
Attribute Value Pair
CN
Correspondent Node
CoA
Care of Address
HA
Home Agent
HMAC
Hash-based Message Authentication Code
HOTP
HMAC based One Time Password
IKE
Internet Key Exchange
i-HA
Internal HA
i-MIP
Internal MIP
MAA
Multimedia-Auth-Answer
MAR
Multimedia-Auth-Request
MIDCOM
Middlebox Communication
MIKEY
Multimedia Internet Keying
MIP
Mobile IP
MN
Mobile Node
MR
Mobile Router
NEMO
Network Mobility
OTP
One Time Passwords
RTCP
RTP Control Protocol
RTP
Real-time Transport Protocol
SA
Security Association
SDP
Session Description Protocol
SIP
Session Initiation Protocol
SIP-NVG
SIP NEMO VPN Gateway
SRTP
Secure Real-time Transport Protocol
TEK
Traffic Encryption Key
TGK
TEK Generation Key
UAA
User-Authorization-Answer
UAR
User-Authorization-Request
VPN
Virtual Private Network
VPN-TIA
VPN Tunnel Inner Address
X-HA
external HA
X-MIP
external MIP
This paper consists of the following sections:

Section 2 examines the problems of architecture
for MVPN proposed in the existing IETF, and it
looks into the need for a SIP-based MVPN.
Section 3 describes the proposed SIP-based
mobility management scheme which is cost
459
effective and secure. Section 4 discusses the
analytical model to evaluate the functioning of the
proposed scheme. Section 5 describes the
numerical results for the analysis presented in
Section 4. Finally, conclusions are drawn in
Section 6.
2 RELATED WORK
IETF has previously defined the architecture and
protocol for MVPN [6]; it is shown in Fig. 1. Here,
the internal HA (i-HA) and external HA (x-HA)
are present in the intranet and Internet and the two
HAs. A new care-of address (CoA) is first
obtained from the dynamic host configuration
protocol (DHCP) server or foreign agent (FA)
when the MN moves out of the intranet. This CoA
is registered in x-HA. Then, MN creates a VPN
gateway and IPSec tunnel using its external home
address (x-HoA). An IPSec tunnel is created by
using internet key exchange (IKE) [7]. Fig. 1
shows the three tunnels (x-MIP, IPSec, and i-MIP).
transfer delay for host and session mobility

[12,13]. However, if SIP is applied to NEMO, it
may increase the handoff signaling cost, when
many re-INVITE messages are transferred among
the sessions in progress. HTTP digest is the basic
user authentication realized in SIP. This
authentication uses a secret key and is based on
the challenge-response paradigm. Most protocols
in Internet applications use this mechanism for
client authentication before providing services;
however, SIP authentication using HTTP digest
increases the signaling exchange in the protocol
design, requiring two handshakes to occur. To
simplify the authentication procedure, we can
instead adopt HOTP-based authentication,
shortening processing time and so reducing
signaling cost. Therefore, in this paper, we
consider these methods for supporting the MVPN
in NEMO and for shortening authentication time
and signaling time in mobile networks, suitable for
real-time applications.
MN
Mobile Gateway
DHCP
X-HA
AAAF
AAAH
Ipsec-VPN
i-HA
CN
MN gets a new CoA
DHCP
Reg-Req & MN-AAA (x-MIP)
802.11b
AP
AAAH
802.11b
AP
AMR
AMR
i-HA
HAR
MN
x-HA
HAR
VPN Gateway & AAAF
HAA
Subnet
Subnet
HAA
AMA
DHCP/i-FA
DMZ
Reg-Reply (x-MIP)
AMA
Register to x-HA
(CoA mode)
Intranet
Internet
x-FA
HDR, SA
Subnet
1. Register to x-HA by using CoA
HDR, SA
HDR, KE, Ni
Subnet
2. Establish Ipsec tunnel with VPN by using x-HoA
HDR, KE, Nr
HDR*, KE, Idii, HASH_I
3. Register to i-HoA by using VPN-TIA

802.11b
AP
HDR*, KE, Idii, HASH_R
Establish Ipsec tunnel
Three Tunnels
802.11b
AP
MN
x-MIP
(x-HA to MN Coa)
IPsec
(VPN GW to x-HoA)
Reg-Req(i-MIP)
i-MIP
(i-HA to VPN-TIA)
Original Packet
Reg-Reply(i-MIP)
HDR*, N/D (Information Exchange)
Register to i-HA
Terminate Ipsec tunnel
Figure 1. MVPN proposed by IETF.
Fig. 2 shows the signaling message flows of

IETF MVPN. Because the mobility of a mobile
equipment group was not considered in IETF
MVPN, it cannot be applied to NEMO, since it
would cause long handoff latency and end-to-end
latency [8,9]. These tunnels significantly increase
the overhead due to packet length and processing
time, and this can degrade the performance in realtime applications. Although SIP-based MVPN was
proposed, only the mobility of a single node was
considered [10,11].
In this paper, we propose SIP-based NEMO
because it is easily distributed and reduces the data
Figure 2. Signalling flows of IETF MVPN.
3 COST-EFFECTIVE AND SECURE

MOBILITY MANAGEMENT SCHEME
3.1 System Architecture
IETF has defined architecture and protocols for
mobile VPN. However, The IETF mobile VPN
cannot be applied to NEMO because it does not
consider the mobility of a group of mobile devices.
Besides, the IETF solution is based on IPsec and
MIPv4, so it will incur long handoff latency and
460
end-to-end latency. On the other hand, SIP has
been proposed to provide host mobility and
session continuity. However, by adopting SIP into
NEMO, it may increase signaling cost during
network handoff. So, I propose architecture and
protocols to support VPN in NEMO, which is
called Cost Effective and Secure Mobility
Management Scheme [40]. The proposed SeSIP
comprises SIP, secure real-time transport protocol
(SRTP) [14], multimedia internet keying
(MIKEY) [15], and a Diameter server [16] to
provide VPN services in NEMO. Fig. 3 depicts the
architecture of the proposed SeSIP.
Foreign Network 2
Diameter Server
AP/BS
VPN Gateway
MN
SIP Proxy 1
Intranet
Firewall
SIP Proxy 2
(SIP registrar)
ALG
Foreign Network 1
CN
SIP-NVG
AP/BS
Mobile Network
MN
Signaling traffic
Data traffic
Figure 3. System architecture.
Fig. 3 shows a mobile network in a foreign

network (Internet) connecting to the CN in the
home network (intranet). The SIP NEMO VPN
gateway (SIP-NVG) shown in the mobile network
residing in Foreign Network 1 is the gateway of
the mobile network to other networks. It follows
the SIP standards and manages the traffic between
the mobile network and the outside world. The
VPN gateway consists of SIP Proxy 1 and an
application level gateway (ALG). There is a
firewall between the Internet and the intranet to
prevent external users from getting direct access to
the intranet. SIP Proxy 1 is a SIP proxy server,
which authenticates the incoming SIP messages
through the Diameter server. It also routes
messages to SIP Proxy 2 which is essentially a SIP
registrar. Meanwhile, MIKEY is used as the key
management protocol to provide security keys for
the ALG, which then oversees all data traffic.
In the proposed SeSIP, SIP is the main protocol
to manage the session between MN, SIP-NVG,
SIP Proxy 1, SIP Proxy 2, and CN. Diameter SIP

Application [17] is an adaptation of the Diameter
base protocol [16] that is used to authenticate and
authorize a user in the Diameter server while
resource allocation in ALG is achieved using
Middle box Communication (MIDCOM) [18]. In
addition, MIKEY messages are embedded inside
the messages of the Diameter base protocol and
the Session Description Protocol (SDP) [19] to
carry security information. For user plane, when
the mobile network resides in internet, SRTP is
used to secure the data transmission between MN
and ALG.
SIP is an application-layer signaling protocol. It
is used to create, modify, and terminate sessions in
the proposed SeSIP. SIP has defined its own
security and authentication schemes. In our
proposed CE-SeMMS, we use SIP to authenticate
and identify the mobile users. SIP also supports
user mobility and terminal mobility [12], [20].
Terminal mobility is achieved by sending new
INVITE (re-INVITE) to the CN using the same
call ID as that in the original session. The new
INVITE contains the new contact address the MN
has acquired in the new location. After receiving
the re-INVITE, the CN will redirect future traffic
to the MNs new location. SRTP defines a
framework to provide encryption and integrity for
Real-time Transport Protocol (RTP) [21] and RTP
Control Protocol (RTCP) [21] streams.
MIKEY is a key management protocol
developed for multimedia real-time applications
running over RTP/SRTP. In contrast to IKE,
which is widely used as key management protocol
for unicast, MIKEY is designed for peer-to-peer or
small interactive groups. MIKEY can fulfill the
requirements of different environments. For
example, a MIKEY message can be embedded
inside an SDP message. A new type k has been
defined in SDP to carry MIKEY message. The
main purpose of MIKEY is to transport the TEK2
Generation Key(TGK) and other related security
parameters or policies which are used in security
transport protocols.
The Diameter SIP Application allows a client of
a SIP server to be authenticated and authorized by
a Diameter server. There are six Diameter
commands in the Diameter SIP application. In the
proposed SeSIP, we use User-Authorization461
Request (UAR) / User-Authorization-Answer
(UAA) and Multimedia-Auth-Request (MAR) /
Multimedia-Auth-Answer (MAA) to process SIP
REGISTER and INVITE messages. The
authentication is done by the Diameter server
rather than by delegating to a SIP server. HOTPbased authentication is adopted in the proposed
SeSIP to reduce authentication time, an element of
delay time during handoff. HOTP is an OTP
creation algorithm based on event synchronization,
and the client and authentication server share the
secret key K. It uses C, the increasing counter
value, and HMAC-SHA-1 hash algorithm to
create the password. The increased value (C +1) is
used to create a new password (6 digits) during the
following authentication. The OTP mechanism
creates the single user password based on three
parameters: hash algorithm, secret key, and
challenge/counter. HOTP creates the password
using the authentication number (counter) - which
is remembered between the authentication server
and the user - as the input value of OTP, and the
authentication is performed only when the counter
value matches. The counter parameter has the
characteristics of synchronization OTP (HOTP
[5]), and the client creates a new password without
receiving the item beforehand from the
authentication server. HOTP performs client
authentication through one handshake, using the
OTP creation algorithm based on the event
synchronization method.
As discussed above, the SIP-NVG is the mobile
networks gateway to other networks. When a
mobile network roams among different IP subnets,
the SIP-NVG not only keeps ongoing sessions
unbroken, but also transmits data in a secure
manner. There are two types of interfaces owned
by SIP-NVG: egress interface and ingress
interface. A SIP-NVG attaches to the Internet
through an egress interface. Once a mobile
network moves to a new IP subnet, the egress
interface of the SIP-NVG will get a new IP
address. On the other hand, when an MN wants to
join a mobile network, it attaches to the ingress
interface of the SIP-NVG. In our design, each
mobile network has only one SIP-NVG which
essentially is an MR with SIP capability. The
proposed SIP-NVG is able to route SIP messages
and data traffic between its egress interface and
ingress interface by translating the corresponding

headers.
MN
SIP-NVG
Diameter
Server
SIP Proxy 1
SIP Proxy 2
(SIP registrar)
REGISTER
REGISTER
UAR
UAA
REGISTER
MAR
MAA
200 OK
200 OK
200 OK
REGISTER
REGISTER sip:hs.vpn.com SIP/2.0

From : <sip:mn_1@hs.vpn.com>
To : <sip:mn_1@hs.vpn.com>
Call-ID : 22018851@hs.vpn.com
Cseq : 1 REGISTER
Contact :
<sip:mn_1@nemo.vpn.com>
Content-Type : application/sdp
Content-Length : 0
REGISTER
REGISTER sip:hs.vpn.com SIP/2.0

From : <sip:mn_1@hs.vpn.com>
To : <sip:mn_1@hs.vpn.com>
Cseq : 1 REGISTER
Contact : <sip:sipnvg@nemo.vpn.com>
Content-Type : application/sdp
Content-Length : 0
Figure 4. Message flows and translation of REGISTER

when mobile network resides in foreign network.
Fig. 4 depicts the flow for registration when the

mobile network is in a foreign network. When an
MN enters a mobile network, the MN gets a new
IP address and registers it with the SIP-NVG. As
shown in Fig. 4, the MN updates its current
location with the SIP registrar residing in the
home network by sending the REGISTER with the
newly obtained contact address. In this example,
we assume the mobile network resides in a foreign
network, and the new address assigned for the MN
is mn-1@nemo.vpn.com. In our proposed
architecture, SIP Proxy 2 not only handles the
signaling messages but also acts as the SIP
registrar. As illustrated in Fig. 4, the SIP-NVG
translates the contact field in the REGISTER from
the MNs address into the SIP-NVGs URI address,
which is sip-nvg@hs.vpn.com. Also, the SIPNVG establishes a mapping table to record the
registration information for the MN. Hence, each
request targeted to the MN is redirected to the
SIP-NVG.
The proposed architecture depicted in Fig. 3
adopts an ALG which follows MIDCOM
architecture. We propose that the ALG only
accepts commands from SIP Proxy 2 and provides
462
responses for the corresponding commands. When
the ALG receives a special incoming RTP stream
from the home network to an MN in the Internet,
it replaces the whole IP/UDP/RTP header with a
new one, transforms the new RTP packet into
SRTP format, and delivers the SRTP stream to the
destination. In the reverse direction, the ALG
receives the SRTP stream from the Internet, and
the ALG decrypts it and verifies it to decide
whether the SRTP packet is valid. If the SRTP
packet is decrypted and verified successfully, the
RTP payload is carried by a new RTP header. The
new RTP packet is then transmitted to the home
network.
Each session in the ALG requires sufficient
external and internal resources. For example, the
external resource may include an external listening
address, external listening port, external
destination address, and external destination port.
Destination addresses and ports are provided by
SIP Proxy 2. Only when all resources are ready,
does the session in the ALG start. When either the
external or internal resource is reserved
successfully, the ALG will reply with the reserved
listening address and port to SIP Proxy 2.
3.2 Operation Procedures
In the architecture shown in Fig. 3, the entire
mobile network may move from one IP subnet to
another. This is called network handoff. It is also
possible that an MN moves into or moves out of
the mobile network. This is called node handoff.
Fig. 5 depicts the flow when an MN moves into
a mobile network which is located inside a foreign
network. First, the MN registers with the SIPNVG and the SIP registrar, as discussed in Section
3.2. Then, the MN must re-invite the CN, if there
are active sessions between them. For the INVITE
request, in addition to translating the CONTACT
field from the MNs address into the SIP-NVGs
URI address, the SIP-NVG also adds the
RECORD-ROUTE field where the SIP-NVGs
URI address is inserted. Therefore, subsequent
messages of the existing sessions will be routed by
the SIP-NVG. In this case, the signaling messages
will need to go through SIP Proxy Server 1 and be
authenticated by the Diameter Server before they
reach SIP Proxy Server 2. When the mobile
network and the CN are both located inside the

same realm, for example, the same intranet or the
same network domain in the Internet, the data
traffic between the MNs attaching to the mobile
network and the CN does not need to pass through
the ALG. Instead, the data traffic is transmitted
directly between the mobile network and CN.
Thus, we assume that the CN is inside the intranet
and that the mobile network moves between the
intranet and the Internet. Using Fig. 3 as an
example, we consider the cases when the mobile
network moves from Home Network to Foreign
Network 1, from Foreign Network 1 to Foreign
Network 2, then from Foreign Network 2 back to
the Home Network.
MN
SIP Proxy2
(SIP registar)
SIP-NVG
Diameter
Server
CN
Registration to SIP registrar
INVITE
INVITE
MAR
MAA
INVITE
200 OK
200 OK
200 OK
ACK
ACK
ACK
RTP
Figure 5. Message flows and translation of INVITE when

MN moves to a mobile network which is located inside the
home network.
Fig. 6 illustrates the flow when the mobile

network moves from the intranet to the Internet.
When a SIP-NVG moves to a foreign network, it
must register with the SIP registrar using its new
IP address. The SIP-NVG then checks whether
there are MNs with active sessions inside the
mobile network, according to the session table.
The SIP-NVG must re-INVITE all CNs to recover
all of the ongoing sessions. However, this process
may cause substantial amounts of signaling
463
messages in the wireless links. In order to reduce
the signaling overhead, the SIP-NVG combines all
contact addresses of CNs into a URI list. The URI
list is then conveyed by the SDP embedded in one
INVITE message. The re-INVITE contains the
new contact address of the SIP-NVG. This is sent
to SIP Proxy 1, which routes the message to SIP
Proxy 2, assuming that SIP Proxy 1 has been
informed that SIP Proxy 2 is responsible for
verification of ongoing sessions.
MN
SIP-NVG
SIP Proxy 1
Diameter Server
SIP Proxy 2
(SIP registrar)
ALG
CN

INVITE
with a URI list
INVITE
with a URI list
MAR
MAA
Allocate Resource
Allocate Reply
INVITE
:
200 OK
:
Allocate Resource
with URI list
200 OK
with a URI list
200 OK
with a URI list
Allocate Reply
with URI list
200 OK
:
ACK
:
ACK
with a URI list
ACK
with a URI list
ACK
:
RTP(SRTP)
RTP
Figure 6. Message flows when mobile network roams from

home network to foreign network.
SIP registrar, which received the INVITE

message that includes the URI list, transmits each
INVITE message to all CNs. CNs then transmit
the response message (200 OK) to the SIP
registrar if there are any active sessions. This
process may also become a major contributor to
the signaling message in the wireless link, and the
SIP registrar, which received multiple response
messages to reduce this signaling overhead, adds
all connection addresses of CNs to the URI list
and saves it. Later, the URI list is included in a
single Allocate Resource command and
transmitted to ALG. ALG then gives permission to
the external resource through the URI list and then
transmits the Allocate Reply response message to
the SIP registrar. Then, SIP registrar transmits a
response message, which includes the URI list in
SDP, to SIP Proxy 1. The proposed SeSIP method

adopts HOTP SIP authentication, for the
authentication of each active session. HOTP
performs client authentication through one
handshake, using the OTP creation algorithm
based on the event synchronization method. OTP
is created based on the number of authentications
which is the same as the time value synchronized
between the client and the authentication server.
For HTTP digest authentication, a one-time
password is created by using a certain random
number as the input value of the password
algorithm
based
on
non-synchronization
authentication. Therefore, if the authentication
server creates a certain random number and
transmits it to the user, the user creates the hash
value using this random number. The user
authentication is performed by responding to the
resultant output value using the one-time password.
This requires a handshake two times for
authentication. Consequently, the signaling cost
will increase. SIP-NVG saves the security
information which is necessary for authentication,
which shows MN when MN is connected to the
mobile network. Fig. 7 shows an example of the
password created with the HOTP algorithm [5] in
the Call-ID of the SIP message header.
INVITE sip : mn_2@hs.vpn.com SIP/2.0
From : <sip : mn_1@hs.vpn.com>
To : <sip : mn_2@hs.vpn.com>
Cseq : 43 INVITE
Content-Type : application/sip
Content-Length : 0
Figure 7. Example of a SIP Message including HOTP CallId
The Diameter command MAR/MAA in the

Diameter SIP application is used to process
authentication messages. To reduce the signaling
overhead, the security information for each
session is aggregated to one MAR/MAA message.
This can be done easily by setting the reserved bit
to M in the command flags within the
MAR/MAAs Diameter header to indicate that
there are multiple sessions to process. This causes
SIP Proxy 2 to send the MAR to request the
Diameter server to authenticate and authorize the
sessions. Then, the diameter server compares the
464
password included in the Call-ID to the calculated
password to determine authentication of the user
agent. If the two values are the same, the Diameter
server completes the user authentication through
one handshake. The Diameter server informs SIP
Proxy 2 of each sessions verification result, and it
generates one pair of TGK and MIKEY for each
verified session. Thereafter, the MAA has all the
session verification result codes. To establish
TGK and MIKEY message, pre-shared key, which
is one of the most efficient ways to handle key
transport, is used. The pairs of the TGK and
MIKEY messages are transmitted to the SIP Proxy
2 by MAA.
If the SIP-NVG is granted access to the intranet,
then SIP Proxy 2 must allocate enough resources
to guarantee that the sessions will be protected.
Thus, SIP Proxy 2 orders the ALG to reserve an
internal receiving address and receiving port for
each ongoing session. The command from SIP
Proxy 2 may include the Data SA and the TGK for
SRTP protection, and the CNs original listening
addresses and listening ports. The ALG responds
with the reserved addresses and ports. SIP Proxy 2
inserts the listening addresses and listening ports
in the SDP to reproduce INVITE requests based
on the URI list. It then routes them to each CN
individually. Each CN sends back a 200 OK if it
agrees with the SDP of the re-INVITE. After
receiving the 200 OK from each CN, SIP Proxy 2
again orders the ALG to reserve the external
receiving addresses and receiving ports. The ALG
responds with the reserved external receiving
addresses and receiving ports. If all allocated
resources are ready, SIP Proxy 2 replaces the
listening addresses and listening ports in the SDP
of each 200 OK, and it inserts a MIKEY Initiator
message into the SDP to transport the TGK. Then,
SIP Proxy 2 sends each 200 OK with the new SDP
to the SIP-NVG, which forwards it to the MN.
The ALG will then start to function when both
internal and external resources have been acquired.
When each MN attaching to the mobile network
receives the 200 OK, it must process the included
MIKEY Initiator message and extract the shared
TGK. The MN is then required to send an ACK
with SDP which includes the MIKEY responder
message. After the MN sends the ACK, it will
start all transport sessions.
MN
SIP-NVG
SIP Proxy 1
Diameter Server
SIP Proxy 2
(SIP registrar)
ALG
CN
INVITE
with a URI list
INVITE
with a URI list
MAR
MAA
Allocate Resource
with a URI list
Allocate Reply with a

URI list
200 OK
with a URI list
200 OK
with a URI list
200 OK
:
ACK
:
ACK
with a URI list
ACK
with a URI list
RTP(SRTP)
RTP
Figure 8. Message flow when the mobile network roams

from one foreign network to another
Fig. 8 illustrates the flow when the mobile

network moves again to another external network.
The SIP-NVG must also update its new location
with the Registrar and send re-INVITE to the CNs
with active sessions. However, SIP Proxy 2 will
function on behalf of the CNs to respond with the
200 OK. SIP Proxy 2 only orders the ALG to
modify the external listening addresses and ports.
The rest of the process is similar to that described
for Fig. 6.
MN
SIP-NVG
Diameter Server
SIP Proxy 2
(SIP registrar)
ALG
CN
INVITE
with a URI list
MAR
MAA
Deallocate Resource
with a URI list
Deallocate Reply
with a URI list
INVITE
:
200 OK
:
200 OK
with a URI list
200 OK
:
ACK
:
ACK
with a URI list
ACK
:
RTP
465
Figure 9. Message flow when the mobile network roams
from a foreign network back to its home network
Fig. 9 depicts the flow when the mobile

network moves back to the home network.
Because the data traffic is secure within the
intranet, the ALG is requested to deal locate both
the internal and external resources. Because the
proposed architecture is based on SIP rather than
MIP, the problems of MIP, such as the overhead
of three tunnels and potentially long end-to-end
latency, are not problems in our design.
transmission cost is proportional to the distance

between the two network nodes. The processing
cost includes the cost to process messages, verify
messages, and so on.
In our proposed SeSIP, the inter-realm roaming
of a mobile network includes three types of
handoff: 1) From the intranet (home network) to a
foreign network, 2) From a foreign network to
another foreign network, and 3) From a foreign
network back to the intranet.
They are represented as H hf , H ff and H fh ,
respectively.
3.3 Security Vulnerabilities

SIP authentication: This section presents the
qualitative analysis of security vulnerabilities in
the proposed SeSIP. Integrity and confidentiality
of SIP authentication messages are not protected.
Therefore, malicious users may sniff traffic to get
the plaintext or place a spam call. However, in the
proposed SeSIP, the transport of SIP messages can
be easily extended to incorporate Transport Layer
Security (TLS) [22] so the transmission of SIP
messages can be protected.
SIP parser attack: The free text format of SIP
message could make parsing difficult. Attackers
sending a very large messages with unnecessary
headers and bodies can exhaust the resource of
SIP server. The SeSIP may suffer from such attack
too. To solve this problem, the parser in the SeSIP
can be designed to check message size and discard
the one which exceeds the size limit. Also, a
practical implementation provided by [23], [24],
[25] can be adopted.
4 PERFORMANCE ANALYSYS
4.1 System Modeling
In order to support secure communication in VPN,
the proposed SeSIP sends signaling messages
carrying security information. It also sends
signaling messages to maintain session continuity
during handoff. To evaluate the performance of
the proposed SeSIP, it is important to quantify
signaling cost. Similar to that in [26], [27], [28],
[29], the signaling cost function comprises
transmission cost and processing cost. The
N subnets
Mobile network
Mobile network
Mobile network
k
A new session arrives or an existing session departs
Home Network
Foreign Network
H fh
H ff
H hf
Figure 10. Network topology for analysis
We assume that the network topology is

configured as shown in Fig. 10 such that the
mobile network returns to the intranet after it
moves across N1 foreign networks. Hence, when
N is larger, the mobile network travels far away
from its home network before it returns [40]. For
example, the handoff sequence can be shown as
H hf H ff H ff ...H ff H fh . Table 2 lists the parameters
used in the analysis.
For a mobile network, let f m (t ) be a general
density function for the network residence time tM
in a subnet.
Let E tM 1 . Its Laplace transform is
written:
f m* ( s) e st f m (t )dt.
t 0
Table 2. Parameter Definitions for Each Subsystem i .

Parameter
1
1
Definition
Number of networks a mobile network visits before it
goes back to the intranet.
Session arrival rate for a mobile network
Average session service time
Average network residence time
466
Maximum number of ongoing sessions
in a mobile network
Probability density function of

f ( s)
.
s
*
m
We assume that the arrival of SIP sessions to

the mobile network follows a Poisson process with
arrival rate . The service time of a session is
exponentially distributed with mean 1 . As
shown in the diagram in Fig. 9, similar to that in
[30], we define a(k ) as the probability that a
mobile network will move across k subnets
between two events while there are i ongoing
sessions in the mobile network. An event here is a
new session arrival or an ongoing session
departure from the mobile network. We denote
E tS as the interval between two consecutive
i
events. During E tS there are i ongoing sessions

in the mobile network. Based on the property of
sums of two independent Poisson process,
E tS can be considered as the inter arrival time
i
of a new Poisson process. Therefore,

1
i , 0 i c
E tSi
1
,
i c,
i
1
(1)
where c is the maximum number of ongoing

sessions allowed in a mobile network. According
to the above assumption, we have:
(1 f m* ( i ))
1
,
k 0,
ai (k )
1 f * ( ) 2 f * ( ) k 1 , k 0.
m
i m
i
i
(2)
For simplicity, we denote gi f m* ( i ) in the rest

of the paper. Let k jN q and 0 q N . Then,
ai ( jN q)
(1 gi )2 N j q
( gi ) gi yz j x q ,
i gi
(4)
Hence, we obtain:

gi f m* ( i )
.
i
(5)
In the proposed SeSIP, when a mobile network

moves across networks, it must perform
registration with the SIP Registrar to update its
location. It also must send re-INVITE messages to
the CNs if there are ongoing sessions with the
MNs in the mobile network. Hence, the cost
comprises two parts: the registration cost for
SIPNVG and the re-INVITE cost for maintaining
session continuity. The registration cost is
independent of the number of ongoing sessions in
the mobile network, because the SIP-NVG can
register with the SIP Registrar on behalf of the
whole mobile network. On the other hand, the reINVITE cost depends on the number of ongoing
sessions in the mobile network.
The cost increases when the number of ongoing
sessions increases. However, because we design a
URI list embedded in one re-INVITE message, the
cost to really send a re-INVITE message to each
individual CN is nearly constant, regardless of the
number of ongoing sessions in the mobile network.
Moreover, client authentication is performed
through one handshake by using HOTP-based
authentication [5] when the REGISTER or
INVITE is requested. This reduces the
request/response number when the signaling
message is processed, thereby reducing the
processing cost of the signaling. This can be seen
in Fig. 6, 8, and 9. Table 3 lists the parameters for
handoff signaling cost.
Therefore, we can denote the signaling cost for
handoff:
(3)
For demonstration purpose, we assume that the

network residence time follows a Gamma
distribution. The Laplace transform of a Gamma
random variable is expressed:
Shfi R f Lhf iI hf ,
S iff R f L ff iI ff ,
(6)
S Rh L fh iI fh .
i
fh
467
During E tS , we assume that the mobile
network crosses k subnets as shown in Fig. 10. We
define as the number of subnets the mobile
network moved across from the time it visited the
intranet until the time the last event occurred.
When 0 < < N, the total signaling cost in the
mobile network during E tS can be derived as :
i
Table 3. Parameters for Handoff Signalling Cost.

Param
eter
Definition
Average handoff cost when a mobile network moves from
the home network to a foreign network with i ongoing
sessions.
Average handoff cost when a mobile network moves from a
foreign network to another foreign network with i ongoing
sessions.
Average handoff cost when a mobile network moves from a
foreign network to the home network with i ongoing
sessions.
Average registration cost of a mobile network (sent by SIPNVG) when the mobile network enters its home network.
Average registration cost of a mobile network (sent by SIPNVG) when the mobile network enters a foreign network.
Average cost for the first part of re-INVITE when a mobile
network moves from its home network to a foreign network.
network moves from a foreign network to another foreign
network.
network moves from a foreign network to its home network.
Average cost for the second part of re-INVITE of a session
when a mobile network moves from its home network to a
foreign network.
when a mobile network moves from a foreign network to
another foreign network.
when a mobile network moves from a foreign network to its
home network.
S hfi
S iff
S ifh
Rh
Rf
Lhf
L ff
L fh
I hf
I ff
I fh
Ci ( N , i , , )
k 1
i k
N S fh N
k 0
k 1 k
S iff k
ai (k )
N N
i
hf
i
i
i
i
N 1
( Shf gi S ff gi S fh S ff )(1 gi ) gi
S iff
. (7)
i
(1 giN ) gi i
If is exponentially distributed, the p.d.f of

is:
1 2e ( N 1)/2 e1
e
,
1 e1

( N 1)/2
e1
e ( N ) 1 2e
1 e1
N 1
2
N 1
N 1.
2
(8)
Therefore,
Ci _ exponential ( N , i , )
N 1
C ( N , , , ) S
i
ff
(1 gi )
i i (1 giN )
1
i
i
i
i
N 1
(1 e )( Shf S ff ( S fh S ff ) gi )
1 2e ( N 1) 2 e1
(1 e1 )( Shfi gi S iff gi S ifh S iff ) giN 1
1 2e ( N 1) 2 e1
N 1
N 1 2 e
e N
g
g
i
1 i N 1 2
(9)
As discussed above, the arrival of sessions to a

mobile network follows a Poisson process, and the
session service time is exponentially distributed.
In addition, there is a limit c for the maximum
number of ongoing sessions allowed in the mobile
network. Therefore, we can model the number of
ongoing sessions in a mobile network as an M/M/
c/c queuing system. The steady state probability
that there are i ongoing sessions in the mobile
network is then given by [31]:
i c x
Pi
i ! i x 0 x ! x
(10)
As a result, the average handoff-signaling cost

per unit time can be derived as:
c
C
i 0
P i .
i _ exponential i
(11)
The variables i and Pi can be obtained from

(1) and (10). Further, we consider the effect of the
variance of the mobility pattern. We assume the
average residence time is gamma distributed [32],
[33]. Therefore, the variance is:
468
Var
(12)
M x 2d mg 2d xha 2Wmn , mg 2Wmg , xha 2 H ,
Table 4. Parameters for SeSIP Signalling Cost.

Parameter
Definition
ax
The processing cost for SIP registration at Node
bx
The processing cost for SIP INVITE message at Node
Ax , y
The transmission cost of SIP registration between Node

x and Node y
Bx , y
The transmission cost of a SIP INVITE message

between Node x and Node y
The total cost for SIP Proxy 1 to process and transmit

UAR/UAA messages to the Diameter server
The total cost for SIP Proxy 2 to process and transmit
MAR/MAA messages to the Diameter server
parameters are defined: Based on the signaling

message flow shown in Fig. 2, the above cost can
be calculated:
To evaluate the performance of the proposed

SeSIP, Table 4 lists the parameters used in SeSIP
Signaling Cost. where x and y can be mn; nvg;
pro; reg; alg; or cn which denote MN, SIP-NVG,
SIP Proxy 1, SIP Proxy 2 (SIP Registrar), ALG,
and CN, respectively. According to the signaling
message flow described in Section 3, the above
costs can be calculated:
Rh anvg areg 2 Anvg , reg M ,
R f anvg 2a pro areg 2 Anvg , pro 2 Apro ,reg U M ,
Lhf 2bnvg 3bpro 4breg 2bal g 3Bnvg , pro 3B pro , reg
4 Breg , alg Breg ,cn M ,
L ff 2bnvg 3bpro 2breg balg 3Bnvg , pro 3B pro ,reg
2 Breg , alg M ,
M i o 2d mg 2d xha 2d vpn d iha 2Wmn , mg

2Wmg , xha 2Wxha ,vpn 2Wvpn ,iha 2 H ,
M i i diha 2Wmn ,mg 2Wmg ,iha ,
Test 2emn 6emg 6exha 3evpn 6Z mn , mg
6Z mg , xha 6Z xha ,vpn ,
Tter Z mn ,mg Z mg ,vpn evpn .
The handoff cost of IETF MVPN when a

mobile network moves between networks is
derived as:
Dhf M x M i o Test ,
D ff M x ,
D fh M i i Tter .
In the architecture we propose, SIP-NVG

manages the overall network mobility, registering
the whole mobile network in the SIP Registrar
when it moves to a new subnet. If there is no SIPNVG, all MNs in the same mobile network must
update their locations separately. This increases
signaling cost. We can re-define the costs (6)
when there is no SIP-NVG as follows:
L fh 2bnvg 3breg balg 3Bnvg , reg 2 Breg , alg Breg ,cn
Shfi mR f iLhf iI hf ,
M ,
S iff mR f iL fh iI ff ,
I hf bmn bnvg breg bcn 2 Bmn , nvg 2 Breg ,cn ,

I ff bmn bnvg 2 Bmn ,nvg ,
I fh bmn bnvg 2 Bmn ,nvg breg bcn 2 Breg ,cn ;
To compare the signaling cost to that of IETF

MVPN, we assume the Diameter MIPv4
application [34] is used to authenticate the x-MIP.
Also, we assume MN is in collocated mode.
Analysis of an FA mode is not presented here,
because it has almost the same results. The
subscripts mn;mg; vpn; xha; and iha refer to MN,
Mobile Gateway, IPsec-based VPN gateway, xHA, and i-HA, respectively. Also, the following
(13)
(14)
S ifh mRh iLhh iI fh .
where m is the number of MNs connected to the

mobile network. In addition, the HOTP
authentication method [5] used in the proposed
architecture is an event synchronization method
that performs client authentication through a
single handshake. In contrast, the HTTP digest
authentication method adopted by many other
security protocols is based on the challengeresponse paradigm and requires two handshakes
between the client of the SIP server and the
Diameter server. This is because the SIP Registrar
469
uses UAR/UAA and MAR/MAA commands for
the Diameter server for user authentication and
authorization when the client of the SIP server
transmits SIP REGISTER and INVITE messages.
Therefore, the cost incurred in the use of HTTP
digest authentication can be calculated as follows:
reduce the signaling cost for handoff while

supporting the VPN.
Table 5. Table Type Styles.
Parameter
Mx
M i o
Rh 2anvg 2areg 4 Anvg ,reg 2 M ,
M i i
R f 2anvg 4a pro 2areg 4 Areg , pro 4 Apro , reg
4 Breg , alg Breg ,cn 2M ,
Test
Tter
dx
ex
L ff 3bnvg 5bpro 3breg balg 5Breg , pro 5B pro ,reg
Wx , y
2 Breg , alg 2 M ,
Z x, y
L fh 3bnvg 4breg balg 5Bnvg ,reg 2 Breg ,alg Breg ,cn
2U 2 M ,
Lhf 3bnvg 5bpro 5breg 2balg 5 Breg , pro 5 B pro ,reg
2 M ,
4000
I ff bmn bnvg 2 Bmn ,nvg ,
3500
I fh bmn bnvg breg bcn 2 Bmn ,nvg 2 Breg ,cn .
3000
4.2 Numerical Results

This section provides the numerical results for the
analysis presented in Section 4. The analysis was
validated by extensive simulations using ns-2 [35].
As discussed in Section 4, the signaling cost
function consists of the transmission cost and the
processing cost. We assume that the transmission
cost is proportional to the distance between the
source and destination nodes, and the processing
cost includes the processing and verifying SIP
messages [36], [37], [38], [39]. Also, the
transmission cost of a wireless link is higher than
that of a wire line. To illustrate performance,
reasonable values were chosen for the parameters,
as listed in Table 5. Additionally, to compare our
designs signaling cost with that of IETF MVPN,
we assumed the x-HA was optimally collocated
with the VPN gateway and AAAF, and the i-HA
was collocated with the SIP Proxy 2/Registrar.
Also, the AAAH in IETF MVPN as collocated
with the Diameter server in the SeSIP. One of the
major objectives in the proposed SeSIP is to
Signaling Cost(per unit Time)
I hf bmn bnvg breg bcn 2 Bmn ,nvg 2 Breg ,cn ,
Definition
The x-MIP registration cost.
The i-MIP registration cost when MN is located
outside the intranet.
The i-MIP registration cost when MN is located inside
the intranet.
The establishment cost of IPsec tunnel.
The termination cost of IPsec tunnel.
The processing cost for MIP registration at Node x.
The processing cost for IPsec message at Node x.
The transmission cost of MIP registration between
Node x and Node y.
The transmission cost of IPsec message between Node
x and Node y.
The total cost for x-HA to process and transmit
HAR/HAA and AMR/AMA messages to AAAF and
AAAH.
IETF-MVPN
With SIP-NVG(use HTTP)
Without SIP-NVG(use HTTP)
2500
2000
1500
1000
500
0
10
Average Residence Time
Figure 11. Comparison of various signalling costs versus

residence time(1)
Fig. 11 presents a comparison among the

signaling costs with IETF MVPN, with and
without SIP-NVG using HTTP. Also Fig. 12
presents a comparison among the signaling costs
with SIP-NVG using HTTP, and using HOTP. As
defined above, m is the number of MNs attached
to the mobile network. In addition, we assume
N 7 , m 5 , c 10 , and 5. In the SIPbased protocol, SIP re-INVITEs and SIPs
registration must be performed during each
handoff. Therefore, the signaling cost for a SIPbased protocol might be higher than for MIP.
However, Fig. 11 shows that a method with SIPNVG has lower signaling costs for handoff than in
IETF MVPN. This is because IETF MVPN
470
requires time to establish the three tunnels.
Compared to the mobile network without SIPNVG, the method with SIP-NVG reduces handoff
signaling cost significantly, since SIP-NVG
performs registration in the SIP Registrar on
behalf of the entire mobile network when it moves
to a new subnet, whereas, without SIP-NVG, all
MNs must update their locations individually.
4000
With SIP-NVG and use HTTP
SeSIP(With SIP-NVG and HOTP)
3500
450
3000
400
2500
350
Signaling Cost(per unit time)
Signaling Cost(per unit Time)
with each technique reaches the maximum number

allowed in the mobile network. Comparing c 10
and c 15 in the SIP-based techniques, more
sessions exist in the mobile network when c 10 .
Hence, greater signaling cost occurs for a reINVITE. Because moving from home network to
foreign network causes higher signaling cost than
other types of handoffs, revisiting the home
network frequently results in higher signaling cost.
2000
1500
1000
500
0
C=10 Without SIP-NVG (use HOTP)

C=10 With SIP-NVG (use HTTP)
C=10 With SIP-NVG (use HOTP)
300
250
200
150
100
10
Average Residence Time
And, Fig. 12 shows that SeSIP(with SIP-NVG

using HOTP) has lower signaling costs for
handoff than in SIP-NVG using HTTP. Further,
our results show that the HOTP authentication
method reduces signaling cost, compared to the
HTTP digest authentication method, by reducing
the required handshakes from two to one. Handoff
signaling cost decreases when the average network
duration time increases, i.e., when the mobile
network has relatively low mobility. Fig. 13, 14
demonstrates the average signaling cost for
handoff versus , the number of sessions in the
mobile network. The parameters are set as in Fig.
11, 12 except that 0.1 . Similar to Fig. 11, 12,
the proposed SeSIP has less signaling cost for
handoff than without SIP-NVG using HOTP or
with SIP-NVG using the HTTP digest method. We
also that when increases, the average cost for
SIP-based solutions increases too. The reason is
that with more ongoing sessions, more reINVITEs are needed to maintain session
continuity. Besides, when is larger than 20, the
costs of all techniques presented in Fig. 13, 14
remain almost constant. This is because, when
approaches 20, the number of ongoing sessions
10
Average Residence Time(1/gamma)
Figure 13. Comparison of signalling cost with and without

SIP-NVG and using HTTP or HOTP Method (C=10)
500
C=15 Without SIP-NVG (use HOTP)

C=15 With SIP-NVG (use HTTP)
C=15 With SIP-NVG (use HOTP)
450
Signaling Cost(per unit time)
Figure 12. Comparison of various signalling costs versus

residence time(2)
400
350
300
250
200
150
100
1
10
Average Residence Time(1/gamma)
Figure 14. Comparison of signalling cost with and without

SIP-NVG and using HTTP or HOTP Method (C=15)
5 CONCLUSIONS
Although the IETF standard has proposed a
mobile VPN architecture, it is designed for the
movement of a signal node only. In addition, IETF
MVPN has large overhead for transmitting realtime packets, because it requires one IPsec tunnel
and two MIP tunnels. On the other hand, there has
been no efficient way to support mobile VPN in
NEMO, even though NEMO supports network
mobility. This paper presents a novel method for
supporting MVPN in NEMO that ensures that the
471
session is maintained continuously when the
whole network moves, and it proposes using the
HOTP-based authentication method to shorten the
processing time of the signaling that continuously
occurs to maintain the session. In addition,
security is enhanced in our design through the
integration of NEMO and VPN.
We analyzed the design and performance of our
proposed design, and results indicate that the
proposed SeSIP based on SIP is well suited to
real-time service. Although SIP-based mobility
management can easily support routing
optimization, there may be an upswing in the
handoff signaling costs, because many signaling
messages are transmitted to maintain the session
in progress with SIP in NEMO. In the proposed
SeSIP, a URI list is used to signify the SIP proxy
server instead of transmitting signaling messages
individually to each node. Therefore, the signaling
cost is reduced. User authentication using the
existing HTTP digest authentication method
requires many handshakes, increasing the
signaling cost. In contrast, the proposed SeSIP
using HOTP-based authentication considerably
reduces the number of handshakes needed with the
authentication server, thus reducing the signaling
cost. The SIP proxy server and the Diameter
server are responsible for authentication and
authorization. Also, the ALG receives a command
from the SIP proxy server to process the security
information for the data transmission, depending
on MIDCOM architecture. ALG is responsible for
converting and relaying the protected and
unprotected data. Thus, unauthorized data cannot
pass the ALG in the Internet. This paper examined
a method for efficient management of group
mobility and cost savings for real-time services
through the integration of mobile VPN and
NEMO. NEMO, currently in the early stage of
research, is expected to be further realized through
the convergence of various technologies, policies,
and methods, such as the path optimization
method for efficient services, multi-homing
technology, and methods for services in the
inclusive mobility network.
These need to be researched in the period
ahead. For commercial service, research should
also be conducted to develop technology to enable
fast detection of movement.
6 ACKNOWLEDGEMENT
This research was supported by Basic Science
Research Program through the National Research
Foundation of Korea (NRF) funded by the
Ministry of Education (NRF-2010-0024695).
This article is a revised and expanded version of a
paper entitled Secure SIP-based Mobility
Management Scheme for Cost-Optimized NEMO
Environments presented at The International
Conference on Digital Information, Networking,
and Wireless Communications (DINWC2014)
held on June 24-66, 2014 at Ostrava, Czech
Republic.
7 REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
V. Schena and G. Losquadro, FIFTH Project Solutions

Demonstrating
New
Satellite
Broadband
Communication System for High Speed Train, Proc.
IEEE Vehicular Technology Conf., pp. 2831-2835, May
2004.
WirelessCabin Project, http://www.wirelesscabin.com,
2011.
V. Devarapalli, R. Wakikawa, A. Petrescu, and P.
Thubert, Network Mobility (NEMO) Basic Support
Protocol, IETF RFC 3963, Jan. 2005.
S. Kent and R. Atkinson, Security Architecture for the
Internet Protocol, IETF RFC 2401, Nov. 1998.
D. MRaihi, M. Bellare, F. Hoornaert, D. Naccache, and
O. Ranen,HOTP: An HMAC-Based One-Time
Password Algorithm,RFC 4226, December 2005.
S. Vaarala and E. Klovning, Mobile IPv4 Traversal
Across IPsec-Based VPN Gateways, IETF RFC 5265,
June 2008. 7.
D. Harkins and D. Carrel, The Internet Key Exchange
(IKE),IETF RFC 2409, Nov. 1998.
J.-C. Chen, Y.-W. Liu, and L.-W. Lin, Mobile Virtual
Private Networks with Dynamic MIP Home Agent
Assignment, Wireless Comm. and Mobile Computing,
vol. 6, no. 5, pp. 601-616, Aug. 2006.
J.-C. Chen, J.-C. Liang, S.-T. Wang, S.-Y. Pan, Y.-S.
Chen, and Y.-Y. Chen, Fast Handoff in Mobile Virtual
Private Networks, Proc. IEEE Intl Symp. World of
Wireless Mobile and Multimedia Networks (WoWMoM
06), pp. 548-552, June 2006.
S.-C. Huang, Z.-H. Liu, and J.-C. Chen, SIP-Based
Mobile VPN for Real-Time Applications, Proc. IEEE
Wireless Comm. And Networking Conf. (WCNC 05),
pp. 2318-2323, Mar. 2005.
Z.-H. Liu, J.-C. Chen, and T.-C. Chen, Design and
Analysis of SIP-Based Mobile VPN for Real-Time
Applications, IEEE Trans. Wireless Comm., vol. 8, no.
11, pp. 5650-5661, Nov. 2009.
A. Dutta, F. Vakil, J.-C. Chen, M. Tauil, S. Baba, N.
Nakajima, and H. Schulzrinne, Application Layer
Mobility Management Scheme for Wireless Internet,
Proc. IEEE Intl Conf. Third Generation Wireless and
beyond (3G Wireless), pp. 379-385, May 2001.
472
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
D. Vali, S. Paskalis, A. Kaloxylos, and L. Merakos, An

Efficient Micro-Mobility Solution for SIP Networks,
Proc. IEEE GLOBECOM, pp. 3088-3092, Dec. 2003.
M. Baugher, D. McGrew, M. Naslund, E. Carrara, and
K. Norrman, The Secure Real-Time Transport Protocol
(SRTP),IETF RFC 3711, Mar. 2004.
J. Arkko, E. Carrara, F. Lindholm, M. Naslund, and K.
Norrman, MIKEY: Multimedia Internet KEYing, IETF
RFC 3830, Aug. 2004.
P. Calhoun, J. Loughney, E. Guttman, G. Zorn, and J.
Arkko,Diameter Base Protocol, IETF RFC 3588, Sept.
2003.
M. Garcia-Martin, M. Belinchon, M. Pallares-Lopez, C.
Canales, and K. Tammi, Diameter Session Initiation
Protocol (SIP) Application, IETF RFC 4740, Nov.
2006.
P. Srisuresh, J. Kuthan, J. Rosenberg, A. Molitor, and
A. Rayhan, Middlebox Communication Architecture
and Framework, IETF RFC 3303, Aug. 2002.
M. Handley and V. Jacobson, SDP: Session Description
Protocol, IETF RFC 2327, Apr. 1998.
J.-C. Chen and T. Zhang, IP-Based Next-Generation
Wireless Networks. John Wiley and Sons, Jan. 2004.
H. Schulzrinne, S. Casner, R. Frederick, and V.
Jacobson, RTP: A Transport Protocol for Real-Time
Applications, IETF RFC 3550, July 2003.
E. Rescorla, SSL and TLS: Designing and Building
Secure Systems. Addison Wesley, 2001.
D. Geneiatakis, G. Kambourakis, C. Lambrinoudakis, T.
Dagiuklas, and S. Gritzalis, A Framework for Protecting
a SIP-Based Infrastructure against Malformed Message
Attacks, Computer Networks, vol. 51, no. 10, pp. 25802593, July 2007.
D. Geneiatakis and C. Lambrinoudakis, An Ontology
Description for SIP Security Flaws, Computer Comm.,
vol. 30, no. 6, pp. 1367-1374, Mar. 2007.
D. Geneiatakis, G. Kambourakis, and T. Dagiuklas, A
Framework for Detecting Malformed Messages in SIP
Networks, Proc. 14th IEEE Workshop Local and
Metropolitan Area Networks, 5 pp. -5, Sept. 2005.
J. Xie and I.F. Akyildiz, A Novel Distributed Dynamic
Location Management Scheme for Minimizing
Signaling Costs in Mobile IP, IEEE Trans. Mobile
Computing, vol. 1, no. 3, pp. 163-175, July-Sep. 2002.
W. Ma and Y. Fang, Dynamic Hierarchical Mobility
Management Strategy for Mobile IP Networks, IEEE J.
Selected Areas Comm., vol. 22, no. 4, pp. 664-676, May
2004.
R. Rummler, Y.W. Chung, and A.H. Aghvami,
Modeling and Analysis of an Efficient Multicast
Mechanism for UMTS, IEEE Trans. Vehicular
Technology, vol. 54, no. 1, pp. 350-365, Jan. 2005.
S. Fu, M. Atiquzzaman, L. Ma, and Y.-J. Lee, Signaling
Cost and Performance of SIGMA: A Seamless
Handover Scheme for Data Networks, Wireless
Communications and Mobile Computing, vol. 5, no. 7,
pp. 825-845, Nov. 2005.
Y.-B. Lin, Reducing Location Update Cost, IEEE/ACM
Trans. Networks, vol. 5, no. 1, pp. 25-33, Feb. 1997.
D. Gross and C.M. Harris, Fundmentals of Queueing
Theory. John Wiley and Sons, 1998.
M.M. Zonoozi and P. Dassanayake, User Mobility
Modeling and Characterization of Mobility Patterns,
IEEE J. Selected Areas Comm., vol. 15, no. 7, pp. 12391252, Sept. 1997.
Y. Fang and I. Chlamtac, Teletraffic Analysis and
Mobility Modeling of PCS Networks, IEEE Trans.
Comm., vol. 47, no. 7, pp. 1062-1072, July 1999.
[34]
[35]
[36]
[37]
[38]
[39]
[40]
P. Calhoun, T. Johansson, C. Perkins, T. Hiller, and P.

McCann, Diameter Mobile IPv4 Application, RFC
4004, Aug. 2005.
The
Network
Simulatorns-2,
http://www.isi.edu/nsnam/ns, 2011.
J. Xie and I.F. Akyildiz, A Novel Distributed Dynamic
Location Management Scheme for Minimizing
Signaling Costs in Mobile IP, IEEE Trans. Mobile
Computing, vol. 1, no. 3, pp. 163-175, July-Sep. 2002.
W. Ma and Y. Fang, Dynamic Hierarchical Mobility
Management Strategy for Mobile IP Networks, IEEE J.
Selected Areas Comm., vol. 22, no. 4, pp. 664-676, May
2004.
R. Rummler, Y.W. Chung, and A.H. Aghvami,
Modeling and Analysis of an Efficient Multicast
Mechanism for UMTS, IEEE Trans. Vehicular
Technology, vol. 54, no. 1, pp. 350-365, Jan. 2005.
S. Fu, M. Atiquzzaman, L. Ma, and Y.-J. Lee, Signaling
Cost and Performance of SIGMA: A Seamless
Handover Scheme for Data Networks, Wireless
Communications and Mobile Computing, vol. 5, no. 7,
pp. 825-845, Nov. 2005.
T.C. Chen, J.C. Chen, Z.H. Liu, Secure Network
Mobility (SeNEMO) for Real-Time Applications,
IEEE Trans. Mobile Computing, vol. 10, no. 8, Aug
2011.
473
Fall Detection System Design by Smart Phone

Yung-Gi Wu and Sheng-Lun Tsai
Department of Computer Science & Information Engineering
Chang Jung Christian University
Tainan city, Taiwan
wyg@mail.cjcu.edu.tw
ABSTRACT
Fall detection is one of the major issues in health care
filed.
Falls can cause serious injury both in
physiology and psychology, especially to the old
people. A reliable fall detector can provide rapid
emergency medical care for the fallen down people.
Thus, a reliable and effectively fall detection system is
necessary. In this paper, we propose a system which
utilizing mobile phones as a detector to detect the
falling. When fall accident occurs, the system has
three response procedures for help.
The first
procedure is transmitting the emergency message to
the related people for help. The second procedure
shows the users status and location on the map of
webpage, according to users GPS location and status.
The third procedure makes the alarm sound; its
purpose is to let the person who nearby the user can be
noticed that the user needs help. First, using a
waist-mounted mobile phone to capture accelerometer
of the human body and adopt the DCT (Discrete
Cosine Transform) to analyze the value of
accelerometer to distinguish the activities of daily
living (ADL) and falls. ADL consist of walking,
standing and sitting.
We utilized a tri-axial
accelerometer in mobile phone to capture the signal
and transmit it to the server by way of Internet. We
adapt two judgments achieved in Server, first judgment
is based on an adaptive threshold for detecting the
energy by DCT; the setting of adaptive threshold
include height, weight and gender. The second
judgment is according to the tilt of smart phone.
Experimental results show that this method can detect
the falls effectively; in addition, it is more portable
than other devices as well.
KEYWORDS
Fall detection; health care; mobile phone; DCT;
accelerometer;
1 INTRODUCTION
bbandydd@hotmail.com
The average age of the population of the world is
increasing because the increment in the average life
span. The number of persons with older ages are
increasing faster the higher the age is considered.
According to Taiwan population statistics, the
population rate of aged 65 or over is expected to rise
from 10.4% in 2008 to approximately 37.5% in 2056.
Among older persons, approximate 25~35% of them
experienced fall-related injury more than one time per
year.
The fall in the elderly persons is one of the major
public health problems, due to it causes a serious injury
and obstacle to their independent living. If the
emergency medical treatments were not in time, these
injuries may result in disability, paralysis and even
death. Once the falling is occurred, specific persons
such as family, colleagues, and friends should be
notified to assist the people as soon as possible. It is
believed that the earlier the fall is notified; the lower is
the rate of morbidity-mortality. Thus, reliable, fast
and effective fall detection notified system is getting
more important. A typical fall detection system has
two major parts: the detection part and notify part.
The detection part is able to detect the fall efficiently
and accurately. Thus it needs a notify part to notify
the related persons. In this paper, we use message,
GPS and alarm sound to achieve the notification. A
reliable fall detection system cant just detect the fall;
it must have the ability to help the user from a danger
situation. The necessary functions in fall detection
system should include computing ability (compute the
energy by DCT for detecting the fall), media player
(make an alarm sound so that the person who nearby
the user can be notice that user needs help), Wi-Fi or
3G (transmit the accelerometer value to server) and
communication (automatic send a text message to
related person); these functions contain the detection
part and notify part. Due to mobile phone naturally
combines the two parts; therefore, we utilize mobile
phone as the major part in our fall detection system.
1.2 Motivation and Objectives
1.1 Background
474
In the recent years, the mobile phone is portable

and convenient. They become the most popular
mobile devices. The functionality of mobile phone
development has been enhanced more than before.
Since the mobile phones have already become an
essential in modern life, this paper proposes a
waist-mounted mobile phone to capture the
accelerometer of the human bodys movement. The
advantages of using mobile phone are portable,
convenient and popularization. Accelerometer built
in the smart phone is low-cost, flexible, and it is an
accurate method for the analysis of posture and
movement, which can be applied in fall detection.
Accelerometer-based systems have been shown that
can accurate measure both dynamic and static activities
in both long and short term situations.
Our fall detection system has two major
components: the server (computer) and the client
(mobile phone). The server is used to calculate the
value of accelerometers to judge the falling or not, the
judgment of falling is based on an adaptive threshold;
include height, weight and gender. Once fall is
judged, it will notify the related persons by automatic
sending a text message for help and the mobile phone
will make an alarm sound so that the person who
nearby the user can be notified that user needs help.
We dont use the mobile phone to judge, because the
computer hardware specifications are more advanced
than the mobile phone and so is the computing ability.
The client captures the values of accelerometer to
server via the wireless channel for analyze. In
addition, the client transforms the GPS location of
mobile phone to server per 30 seconds. It will update
the current location on the map of webpage
simultaneously. The webpage presents the users
status and location on the map; therefore, related
persons can always obtain the users situation on the
web.
In this paper, we will describe a system based on 3
tri-axial accelerometers in a mobile phone to detect the
activities of ADL and falls. First, we will convert the
value of 3 tri-axial accelerometers to Sum Vector
Magnitude (SVM) and transform the SVM to server.
Secondly, adopt a well-known transform method as
Discrete Cosine Transform (DCT) to analysis SVM to
distinguish ADL and falls in server. In this paper, we
apply an adaptive threshold and the tilt of smart phone
to judge the fall. Experimental results show that this
method can detect the falls effectively and more
portable than other devices.
Mobility tele-monitoring as this system is a
growing area, which enables the subjective monitoring
of the health status of elderly people living
independently in their own homes. It provides the

clinician with continuous quantitative data that can
indicate an improvement or deterioration in a patients
condition. Tele-monitoring also reduces the cost of
providing care to elderly subjects by moving care from
the traditional hospital/nursing home setting into the
home, thus making more efficient use of healthcare
resources.
The rest of the paper is divided into four sections.
Sec.2 describes an overview of existing fall detection
systems. Sec. 3 proposes our fall detection method.
The experimental result is presented in Sec. 4. The
system interface is represented in Sec.5. Sec. 6 is the
conclusion and extensions of this paper.
2 RELATED WORKS
There are numbers of different approaches for the
fall detection have revealed in recent years. Recent
fall detection systems can be classified to four classes,
which are user activated alarms, visual detection,
wearable sensor and mobile phone. We give these
four classes as follows, 2.1 describes the user activated
alarms, 2.2 describes the visual detection, 2.3 describes
the wearable sensor and 2.4 describes the mobile
phone.
2.1 User Activated Alarms

The first class is user activated alarms. In [1]
and [2], the system requires the person to push a button
on a wearable device, which will simultaneously notify
the hospital. But it relies on the elderly persons
ability to push the alarm button for help. Sometimes,
the fall will lead to loss of consciousness or severe
injuries resulting in inability to push the alarm button.
The advantage of this method is the judgment relies on
him/her without complicate computing.
The
disadvantage of this method is if the person loss of
consciousness then the system is useless.
2.2 Visual detection

The second class is visual detection of a fall
[3][4][5][6].
3 sets up the camera for testing
two-month in an apartment. The camera mounted in
the corner of the room for monitoring the whole room
as possible. 4 makes an alarm when a fall is detected.
They use a simple algorithm for computing motion
vectors and respond fall accidents. They have ability
to distinguish fall accidents and normal daily activities,
such as sitting down, walking and squatting down.
Their system has robust to monitor a persons spatial
position in a room with a pet. 5 provides an
automatic fall detection system, focus on monitoring
the elderly person in a nursing home or in their home.
475
They use a 3D camera to extract images and combine
with a context model. Body orientation can be
obtained from extracting the body posture. The
context model is using for finding visual fall detection,
depending on the location and time of when the fall
events happened. 6 proposes a visual-based fall
detection system, which uses an omni-directional
camera for a better detection perspective. They
propose a simple algorithm of threshold and decision
tree based on the body angle and length variation.
In this class, they use image processing technique
for capturing the images of body and then detecting
visual falls events. In this way, it has limitations on
detecting the fall events, the body is a moving object
but the equipment is fixed. Moreover, due to the
privacy issues, its not suit to monitor persons for a
long time.
2.3 Wearable sensors

The third class is using the wearable sensors.
The most common method for fall detection is using
tri-axial accelerometers. 7 describes a system can
continuously monitoring patients with ease of use,
comfortable. By using accelerometer for measuring
posture, energy expenditure and movement of patient
which are important parameters of the system. 8 uses
a waist-mounted tri-axial accelerometer. Detecting
the change of body angle from standing to lying down.
When a large negative accelerometer value occurred, it
indicates a fall event.
[9] places a tri-axial
accelerometer behind the wearers ear and utilized 3
thresholds for detecting the fall : if all the xy-plane
much larger 2g; if all spatial components before the
impact larger than 0.7m/s; if all spatial components
higher larger 6g. [10] places two sensors on the body,
one on the thigh, and the other on the chest. When
the first hit occurs, it will check if the second hit is
detected within two seconds. If it happens, the
system will calculate the change of hitting angles to
judge the fall. [11] proposes a system that can
monitor the steps, falls and daily activities of elderly
person. Using the elastic-band for mounting the
sensors on the upper body and legs. It is important
for the sensor cannot move freely and the elastic-band
must be fastened. The advantage of this method is
accurately detect the fall. But the disadvantage of this
method is the sensor has not the communication
functions for help.
2.4 Mobile phone

Some authors proposed methods by using mobile
phone which has accelerometer to detect ADL and falls.
[12] embeds a tri-axial accelerometer in a mobile

phone, uses the wireless channel to connect to Internet,
and using 1-Class SVM (Support Vector Machine)
algorithm for the pre-processing, KFD (Kernel Fisher
Discriminant) and K-NN (Nearest Neighbor) algorithm
for the precise classification. [13] proposes a fall
detection system by mobile phones, experimental
equipment has a magnetic sensor in the mobile phone
and a magnetic accessory. The strength between
magnetic sensor and magnetic accessory indicate the
relative position. They put the mobile phone in the
right pocket of the pants and put the magnetic
accessory above the knee of left leg. 14 proposes
an energy efficient method for fall detection; in order
to improve the battery lifetime and computing ability
of mobile phone The power-saving method is using
the manually setting variety activities with different
sampling rate for this purpose.
They utilize
Time-Domain features and Frequency-Domain features
for classified activities.
Time-Domain features
include total magnitude and tilt; Frequency-Domain
features include wavelet coefficients.
3 SYSTEM DESCRIPTIONS AND

IMPLEMENTATION
3.1 Working Procedure
Fig. 1
Working procedure of our system
Our proposed methods workflow is illustrated in

Fig. 1.
The goal was to design a small and
lightweight system that can be worn comfortable.
First, user must key in gender, height and weight. Then,
the accelerometer of mobile phone will capture
acceleration values and transmission both acceleration
values and user data to server. Server captures the
last 50 samples of accelerations in tri-axial and convert
them into SVM values. After obtaining the SVM
476
values, the system calculates the energy by DCT
to determine whether the energy exceed than threshold.
If the energy exceeds the threshold and the phone is tilt
then client will make an alarm sound. If the alarm
sound alert exceeds 15 seconds, mobile phone will
automatic send an emergency message for help and
shows the location and situation of user on webpage.
3.2 Fall Detection Algorithm

Originally, the accelerometer is wearing on the arm
or wrist, just like a watch, but it has shown that the
frequent and severe movements of the arm in ADL
make it difficult to use the acceleration to determine
whether the fall is occurred or not. Therefore, we use
a belt worn around the waist in our experiments. The
most common method for fall detection is using a
tri-axial accelerometers. According to Newtons law
we know that
F=m
(1)
where m is the mass and is the acceleration of the
object.
Fall will produce an up to down force; therefore,
it is reasonable to use acceleration to analyze the fall.
The tri-axial accelerometer can capture the
accelerations of three axes in human body. 15
proposed a sum vector magnitude (SVM) to capture
the acceleration of human body to detect the falls.
= 2 + 2 + 2
(2)
effectively algorithm.
However; refer to Fig. 2, it is
obvious that it is difficult to distinguish ADL and fall
by a threshold only. The ADL are defined as walking,
standing and sitting.
The fatal concern in fall
detection is to make it distinguishable from various
ADL.
In this paper, we utilized a well-known transform
method Discrete Cosine Transform (DCT) in digital
signal processing to analyze the values of SVM.
Transform coding is a kernel component of
contemporary image/video processing applications.
Transform coding relies on the premise that signals
exhibits a certain level of correlation with their
neighboring signals, so that we utilized DCT transform
the spatial domain into frequency domain into find the
features of SVM.
In this paper, we use the
one-dimensional DCT 16 to transform SVM to get its
feature for classification furthermore.
The most
common DCT definition of a 1-D sequence of length N
is given as (3) :
1
1
(2 + 1)
() = () () [
]
2
2
=0
() =
20
30
0
AC
20
SVM
15
if u=0,
Walking
Sitting
15
10
25
2
,
2
c(u)=1,
o.w.
for u=0,1,2,,N-1, N=50 in our experiment.
F(0) denotes the DC and the others are refer to as ACs.
where , , are the acceleration of x-axis,

y-axis and z-axis, respectively.
Walking
Standing
Sitting
(3)
Standing
Fall
-5
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49
-10
-15
-20
Data
10
Fig. 3
0
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49
Data
Fig. 2
50 SVM values in four cases.
Fig. 2 contains four curves which show the 50
SVM values in the four cases of walking, standing,
sitting and fall, respectively. X-axis denotes 50 data
in the last 5 seconds and Y-axis denotes the SVM
values. In 15, it claimed that if SVM > 1.8g, then the
situation must be fall, and then proposed a simple and
DCT (DC is excluded) values of

the four cases
Fig. 3 illustrates the four curves of DCT
coefficients of the four cases, respectively. X-axis
denotes the 49 data (DC excluded) and Y-axis denotes
the AC values.
477
mobile phone mounted on the user and starts a timer.
Once the user does not manually turn off the alarm
within 15 seconds, the system will send an emergency
message for help automatically. In addition, the
mobile phone will make an alarm sound so that the
persons who nearby the user can be noticed that user
needs help; therefore, they can quickly help the user
who is injured for medical care. At the same time,
the map of webpage will present that the situation of
user is dangerous and the location of accident
occurred.
Fig. 4
DC values of the four cases
Refer to Fig. 4, the DC is hard to distinguish so that
we do not take it into consideration for classification.
Due to the DCT owns excellent property of the energy
centralization; most applications of DCT are data
compression. In this paper, we use DCT to get the
energy of SVM to be the feature for classification.
We utilized an array that consists of the last 50 samples
of the SVM. When a new sample is added to the
array, all other samples will be shifted one place and
the oldest sample is eliminated. Then we use (4) to
calculate 49 AC values to get the energy of SVM.
2
= 49
(4)
=1 ()
Notice that, the DC coefficients are removed in the
four cases. DC represents the average of the samples.
Refer to Table 1, in which we list the energies of the
four cases by choosing different amount ACs. The
energy is used to distinguish the fall and ADL. The
selection of threshold will be addressed in the next
section.
Table 1.
Energy by using different amount of ACs

Energy
25 ACs
20 ACs
15 ACs
10 ACs
9 ACs
Walking
306.59
261.01
85.35
26.56
25.50
Standing
0.57
0.53
0.31
0.26
0.26
Sitting
171.93
154.96
149.06
124.09
103.11
Fall
1070.45
997.36
838.56
280.79
204.53
Energy
8 ACs
7 ACs
6 ACs
5 ACs
4 ACs
Walking
25.43
25.42
25.42
23.13
16.59
Standing
0.23
0.23
0.18
0.17
0.15
Sitting
91.25
70.04
50.25
38.46
20.82
Fall
122.83
117.63
116.80
111.09
77.10
The system records the data in the recent 5 seconds

to detect if a fall occurs. If a fall is detected, the
server will transmit a signal that triggers an alarm in
4.1 The training phase
In this section, we will depict how the system
designs and works. In our experiment, the mobile
phone is mounted on the waist. We implement the
system in Java, with Eclipse and Android 2.1 SDK.
We collect data in the four cases including walking,
standing, sitting and fall to get the thresholds.
Because the fall may cause serious injury, we didnt
test falls with real elderly people.
Detailed data are given in Table 2.
Table 2. Volunteer data
#Volunteer
Gender
Age
Height (m)
Weight (kg)
#Volunteer
Gender
Age
Height(m)
Weight(kg)
#Volunteer
Gender
Age
Height(m)
Weight(kg)
a
male
23
1.70
70
h
male
23
1.70
70
o
female
21
1.60
60
b
male
24
1.75
55
i
male
23
1.70
100
p
female
37
1.60
45
c
male
23
1.72
65
j
male
22
1.67
75
q
female
36
1.60
50
d
male
25
1.76
60
k
male
21
1.83
97
r
female
36
1.55
49
e
male
23
1.62
52
l
female
24
1.57
50
f
male
23
1.76
106
m
female
19
1.50
42
g
male
21
1.78
55
n
female
20
1.60
50
We have four situations; each situation was

carrying out three times, so we gather about 216
samples. We assumed that the values of fall are
higher than others, so we eliminate those values of
ADL that are higher than the values of fall. Although
we gather about 216 samples, but some samples are
quite different to the other samples for the same one
person. In that case, we eliminate that the different
data. Fig. 5 shows three curves of three falls for one
same person. In that figure, X-axis denotes the 49
data (DC excluded) and Y-axis denotes the AC values
after DCT. We can observe that the curve of first fall
is different to others. For the accuracy of the
experiment, we eliminate those apparent different
samples. Finally, we gather 200 samples for training
to get the threshold.
478
20
First fall
Second fall
15
10
5
AC
0
-5
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49
-10
-15
Data
Fig. 5
The curve of value of DCT
4.2 Fixed Threshold

Fig. 8
Selecting the first 15 AC values
Fig. 9
Fig. 10
In (4) we calculate 49 AC values to get the energy

of SVM. Nevertheless, DCT has low-frequency,
middle-frequency and high-frequency and it owns
excellent property of energy centralization. Therefore,
calculate 49 AC values are improper. The most
valuable data is in the low-frequency band, those data
in middle-frequency and high-frequency bands are not
worse referencing. We try to find out the coefficients
which can be used for determining the fall.
Fig. 6
Fig. 7
479
Fig. 11
Fig. 12
Fig. 13
Fig. 14
Fig. 15
Refer to Fig. 6 to Fig. 15, they show that selecting
the first 6 AC values to get the energy of SVM as given
in (5) will achieve a better accuracy and rapid
computing time.
= 6=1 ()2
(5)
Detailed data of the E values in four cases for each
volunteers are given in Table 3. Each person tests three
times, sequence is 1, 2 and 3, such as a1, a2 and a3. 0
shows the curve of fixed threshold in dash line,
Table 3. Fixed threshold and energy of each
volunteer
480
height and weight to improve the performance. The
adaptive threshold is divided into two types, one is for
male as given in (6) and another is for female in (7).
It is well-known that the strength of woman is lower
than man, so the adaptive threshold for female is lower
than male; the adaptive thresholds calculated by (6)
and (7) and the adaptive threshold for every volunteer
person are given in Table 4. We can see that each
threshold of person is different, due to the mass of
body on everyone is different.
=
1200
Walking
Standing
Sitting
1000
800
Energy
600
400
200
Table 4.
#Volunteer
threshold
#Volunteer
threshold
#Volunteer
threshold
#Volunteer
threshold
#Volunteer
threshold
9.8
2
9.8
=
(6)
0.75
(7)
Adaptive threshold of each volunteer

a
118.68
e
97.088
i
169.55
m
68.6
q
71.77
b
88
f
167.67
j
131.77
n
71.77
r
74.95
c
107.65
g
85.058
k
141.92
o
86.13
d
94.912
h
118.68
l
74.54
p
64.59
0
a b c d e f g Volunteers
h i j k lm n o pq r
4.4 Phone tilt
Fig. 16 The curve of fixed threshold

It is well-known that the strength of woman is lower
than man, so the fixed threshold for female is lower
than male. The threshold of man is 86 and woman is
74. We can find the fall energy of q3 is lower than
fixed threshold, so the fall is not detected. The fall
energy of q3 is lower than others and it close to the
walking energy of q1; therefore it causes the fall
energy of q3 becomes mis-judgment. Due to the
different person has different body mass, if the number
of samples changed more, we are afraid that more fall
events will be escaped. So we proposed an adaptive
threshold for detecting the ADL and fall, in the next
sub-section.
In this paper, we keep receiving the E
of the user and judge the falls by the adaptive
thresholds. The reason to adopt adaptive threshold is
that different person has different body mass, so it is
not suitable for a fixed threshold.
Fig. 17 Accelerometer values in normal

status
0 shows the forces of gravity detected by the
phone when the program is running, the gravity of
X-axis is -9.8m/2 , the gravity of Y-axis is 0.0 m/2 ,
the gravity of Z-axis is 0.0 m/2 . The reason of
using phone tilt is that some daily activities, like
walking or sitting may cause large force as well. It
will trigger alarm sometimes, in order to solve this
problem so that we apply the phone tilt for detection.
Fall usually cause the body to sharply tilted, we
utilized this feature to determine the fall.
4.3 Adaptive Threshold

Different to [17], which adopt two fixed thresholds
for male and female, respectively. In this paper, we
propose an adaptive threshold based on users gender,
481
Fig. 18 Wearing on the waist
0 shows that the user wears the phone on the
waist. When the phone is horizontal to the ground,
only the gravity of X-axis forces the gravity of ground.
Once the fall is occurred, the phone tilt status will be
changed.
35
Walking
30
25
20
SVM
15
10
5
0
h i j k lm n o pq r
Fig. 21 The curves of maximum value of

SVM
Fig. 19 The fall occurred
0 shows the situation of fall occurred. When the
fall happened, the user will be in lying status. In this
situation, the forces the gravity of ground will impose
to the Y-axis. We assume that when the energy is
larger than adaptive threshold and force the gravity of
ground has changed to other axis instead of X-axis, the
system will judge the situation of user to be fall.
1200
Walking
Standing
Sitting
1000
800
Energy
600
400
200
0
h i j k lm n o pq r
Fig. 20 Four kinds of tilt judgment

As 0 shows, gy represents the frontward and
backward, gz represents the leftward and rightward.
We utilize this four kinds of tilt judgment to obtain the
phone tilt. As mentioned before, fall will cause the
body to sharply tilt.
4.5 Experimental results

We compare our method with the maximum value
of SVM 15.
Fig. 22 The curves of our proposed method

0 contains four curves which show the maximum
values of SVM in the four cases of walking, standing,
sitting and fall, respectively. X-axis denotes the 18
volunteers and Y-axis denotes the SVM values. The
curve of walking, sitting and fall are confusing, it is
hard to distinguish ADL and fall. Only the curve of
standing is different with others.
0 illustrates the four curves of energy by DCT
coefficients of the four cases, respectively and one
curve of threshold. X-axis denotes the 18 volunteers
and Y-axis denotes the energy of DCT values. We
can obvious see that the fall has separated to ADL by
an adaptive threshold. Refer to the above two figures,
we can clearly see that the curves in the first figure is
confusing. It is difficult to distinguish with ADL and
fall. We collected 200 samples for detecting the fall.
The user must key in his or her height, weight and
gender to the system. After user starts the system, the
monitoring keeps running in Android, collecting and
recording the values from sensor. If a fall is detected,
the alarm component works to make an alarm sound to
notify the persons nearby and sending an emergency
message to the related person for help. At the same
time, the map on the webpage shows the location and
the situation of the user is dangerous, as shown in 0.
482
Users must key in their gender, height and weight to
get the adaptive threshold.
Table 7. Tester test result
Number
of tests
#Volunteer
False detection
T1
T2
T3
T4
Standing
10
Walking
10
Sitting
10
Fall
10
Fig. 23 The map on the webpage

Number
of tests
4.6 The testing phase

#Volunteer
In order to test system stability, we test ADL and

fall by myself. We test each activity 50 times.
Detailed data are given in Table 5. There are some
false detection during tests. This test doesnt adapt
the phone tilt method, when the energy of ADL is large,
the system will false detect the ADL as the fall event.
But the angle of ADL isnt as large as the fall, so we
take the phone tilt into account to solve this problem.
Table 5. Test results of ADL and fall
Numbe
r
False
Reason
Accuracy
No
100%
large strength
96%
50
large strength
98%
50
No
100%
of tests
detecti
on
Standing
50
Walking
50
Sitting
Fall
Table 6.
Tester data
False detection
T5
T6
T7
T8
Standing
10
Walking
10
Sitting
10
Fall
10
To evaluate the proposed method, we invite 8

persons to test our system. As given in Table 7, we
test ADL and falls. In order to meet practical
situation, we ask the testers to mount the phone and
wear it in a period of time; in additions, they will act
the falls randomly during the time. Mobile phone is
mounting on the users waist. Test result shows that
alert is not occurred during the test of ADL. When
the testers act the falls randomly, the system detect the
fall and make an alarm sound. All of fall events can
be detected and transmit the help message including
location to the related person soon accurately.
5.SYSTEM INTERFACE
5.1 The interface of server
#Volunteer
T1
T2
T3
T4
Gender
female
female
male
male
Age
23
23
30
51
Height (m)
1.7
1.54
1.65
1.72
Weight (kg)
68
47
65
62
Threshold
86.471
72.831
116.988
102.691
#Volunteer
T5
T6
T7
T8
Gender
male
female
male
male
Age
53
22
37
37
Height (m)
1.62
1.5
1.74
1.78
Weight (kg)
60
37
76
95
Threshold
112.026
60.433
123.002
146.92
The system is carried out to other 8 persons who

are not included in the training step to yield the
adaptive threshold. Detailed data are given in Table 6.
This test adapts both adaptive threshold and phone tilt.
Fig. 24 Server obtains the values from the client

0 shows the message obtained from the client.
483
The nine values from top to the bottom are 1) the
gender of the user, 2) the height of the user, 3) the
weight of the user, 4) the value of SVM, 5) the
accelerometer value of X-axis, 6) the accelerometer
value of Y-axis, 7) the accelerometer value of Z-axis, 8)
the latitude, 9) the longitude. 1), 2) and 3) are used
for adaptive threshold. Calculating 4) to obtain the
energy. 5), 6) and 7) are used for judgment the phone
tilt. 8) and 9) show the users location.
5.2 The interface of client
has gathered 50 data, it will start detecting.
Fig. 27 The fall occurred

0 shows the graph when the fall occurred.
When the system detects the energy exceeds the
threshold and phone tilt, it will alert and start to
countdown 15 seconds.
Fig. 28 Sending emergency message

Fig. 25 Users personal information
0 shows the screen for the user to key in personal
information. User has to key in his/her gender, height,
weight for calculate adaptive threshold, IP for
connecting the server and telephone number for
sending emergency message. After keying in that
personal information, press the start button to start
detecting.
Fig. 26 After pressing the start button

The screen is shown in 0. After pressing the
start button, there are three values appear. 118.685 is
the adaptive threshold which is calculated by gender,
height and weight. The current SVM value is 9.226.
The bottom of value is remaining data, after the system
0 shows if the alarm exceeds 15 seconds, the

system will automatic send an emergency message to
the telephone number that key in the system by user for
further assistance.
6 CONCLUSION
Tri-axial accelerometers are used for fall detection

in many academic articles, but it is a pity that rarely
fall detection uses the mobile phones as platform for
fall detection. Mobile phone is highly portable, and it
consists with accelerometer, gyroscope, wireless
channel and communication component.
All
necessary components are already integrated therein.
In this paper, we propose a method that combines
the digital signal processing technology and fall
detection together; in addition, we utilize mobile
phones as a sensor. Moreover, we propose an
adaptive threshold and phone tilt for fall detection.
This is based on body mass of the user, so it is suitable
for everyone. In addition, we also utilize phone tilt
method for the more precise detection in practical
situation. Experimental results show that our method
achieves good detection performance.
We are planning to extend our proposed method to
other situations, such as traffic accident report system,
fall detection on bicycle. Make this system to be a
484
practical necessity for elderly person in daily life. We
also plan to build a system based on Google Map,
which can collect and show the locations of falls are
occurred. It can provide information that whether the
place is dangers or not, allows relevant agencies to
adopt suitable actions. By monitoring multiple smart
phones, medical center can know all of the patients at
the same time.
7 REFERENCES
1.
2.
11. Vermeiren D., Weyn M., and Ron G.D. Detecting

human motion: introducing step, fall and ADL
algorithms, Electronic Healthcare lecture notes of the
Institute for Computer Sciences, Social Informatics
and Telecommunications, vol. 27, pp. 62-69. (2010)
12. Zhang, T., Wang J., Liu P. and Hou J. Fall detection
by embedding an accelerometer in cellphone and using
KFD algorithm, International Journal of Computer
Science and Network Security, vol. 6, pp. 277-284.
(2006)
Hori T., Nishida Y. and Aizawa H. Sensor Network

for Supporting Elderly Care Home, Proceedings of
IEEE Sensors, vol. 2, pp. 575-578, Vienna Austria.
(2004)
13. Jiangpeng D., Xiaole B., Zhimin Y., Zhaohui S., Dong
Allin S.J., Bharucha A., Zimmerman J. Wilson D.,

Roberson M.J., Stevens S., Wactlar H. and Atkeson
C.G. Toward the Automatic Assessment of
Behavioral Disturbances of Dementia, Proceedings of
2nd Intl Workshop Ubiquitous Computing for
Pervasive Healthcare Applications, pp. 25-29, Seattle
USA. (2003)
14. Yunji L., Xingshe Z., Zhiwen Y., Bin G. and Yue Y.
3.
Sixsmith A., Johnson N. A smart sensor to detect the

falls of elderly, IEEE Pervasive Computing, vol. 3,
pp. 42-47. (2004)
4.
Fu Z., Culurciello E., Lichtsteiner P., Delbruck T.

Fall detection using an address-event temporal
contrast vision sensor, Proceedings of IEEE
international Symposium on Circuits and Systems, pp.
424-427, Seattle, Washington, USA. (2008)
5.
3-axis accelerometer system, Wearable and

autonomous systems, LNEE 75, pp. 259-275. (2010)
Jansen B., Deklerck R. Context aware inactivity

recognition for visual fall detection, Proceeding of
pervasive health conference and workshops, pp. 1-4,
Innsbruck, Austria. (2006)
6.
Miaou S., Sung P., Huang C. A customized human

fall detection system using omni-camera images and
personal information, Proceedings of 1st distributed
diagnosis and home healthcare conference, pp. 39-42 ,
Arlington, Virginia. (2006)
7.
Mathie M., Basilakis J., and Celler B.G. A system for

monitoring posture and physical activity using
accelerometers, 23rd Annual IEEE Engineering in
Medicine and Biology Society, vol. 4, pp. 3654-3657,
Istanbul Turkey. (2001)
8.
Salleh R., MacKenzie D., Mathie M., and Celler B.G.

Low power tri-axial ambulatory falls monitor,
Proceedings of 10th International Conference on
Biomedical Engineering, pp. 223-235, Singapore.
(2000)
9.
Lindemann U. Evaluation of a fall detector based on

accelerometers: a pilot study, Med Biol Eng Comput,
vol. 43, pp. 548-551. (2005)
X. Mobile phone-based pervasive fall detection,

Personal Ubiquitous Comput, vol. 14, pp. 633643.
(2010)
Energy Efficient Recognition Based on Low
Resolution Accelerometer in Smart Phones,
Advances in Grid and Pervasive Computing - 7th
International Conference, pp. 122-136, China. (2012)
15. Mathie M.J., Lovell N.H., Coster A.C.F., and Celler

B.G. Determining Activity Using a Tri-axial
Accelerometer, Engineering in Medicine and
Biology, 24th Annual Conference and the Annual Fall
Meeting of the Biomedical Engineering Society
EMBS/BMES Conference, vol. 3, pp. 2481-2482,
Houston, Texas, USA. (2002)
16. Khayam S.A. The Discrete Cosine Transform (DCT):

Theory and Application,
http://www.dcd.zju.edu.cn/~jun/Courses/Multimedia2
011.../complementary/DCT_Theory%20and%20Appli
cation.pdf, pp.4-6. (2003)
17. Wu Y. G. and Tsai S. L. , The Analysis of G-Sensor

by Discrete Cosine Transform for Fall Event on Smart
Phone , Proceeding of The International Conference
on Digital Information, Networking, and Wireless
Communications (DINWC2014), p.37-44,Ostrava,
Czech, (2014)
10. Mostarac P., Hegedus H., Jurcevic M., Malaric R. and

Lay-Ekuakille A. Fall detection of patients using
485
Persian/Arabic Handwritten Digit Recognition Using Local Binary Pattern

Mohsen Biglari*, Faezeh Mirzaei, Jalil Ghavidel Neycharan
biglari@shahroodut.ac.ir, Department of Computer Engineering, Shahrood University of Technology, Shahrood, Iran
fmirzaei@grad.kashanu.ac.ir, Department of Computer Engineering, University of Kashan, Kashan, Iran
jalil.ghavidel@gmail.com, Department of Computer Engineering, Shahrood University of Technology, Shahrood, Iran
ABSTRACT
Automated handwritten character recognition
seems to be necessary due to the increasing number
of Persian/Arabic handwritten documents. A new
approach for Persian/Arabic handwritten digit
recognition has been proposed in this paper. This
approach employs Local Binary Pattern (LBP) operator
as the base feature extraction method. Although this
operator has shown great performance in research areas
such as context and object recognition, but it has not
been used in Persian/Arabic handwritten digit
recognition problem. First step in the proposed approach
involves smoothing, converting black and white input
image to grayscale intensity image and resizing it to a
fixed size. In the next step, input image is divided into
several blocks. LBP operator is applied to each block to
extract features. Finally, these features are used to train
a multi-layer perceptron neural network with circular
approach. Empirical results shows that the proposed
approach has a very good generalization accuracy
(99.72%) on Hoda dataset with 60000 train and 20000
test samples. This accuracy is the best among the stateof-the-art methods.
KEYWORDS
Persian/Arabic Handwritten Digit Recognition, Local
Binary pattern, Multi-Layer Perceptron, Binary to
Grayscale Conversion, Hoda DataSet
1 INTRODUCTION
Automated handwritten character recognition
seems to be necessary due to the increasing number
of Persian/Arabic handwritten documents.
Handwritten character recognition research for
Latin script family started over 50 years ago, but
its rather a new emerging research area for
Persian/Arabic language with great achievement in
the past few years. Persian and Arabic digit
handwritings are exactly the same. So, all the

descriptions in this text works for both, however for
the sake of shortness, only Persian phrase will be
used from now on.
The characteristics of Persian digit handwriting are
different from the Latin. Some digits in Persian can
be written in different forms. For example, five can
be written like an empty circle or like 5.
Distinguishing the first form from zero is very
difficult in some cases. There are more other
challenges like this which make Persian
handwritten digit recognition harder than the Latin.
Good performance of gradient based features in
Latin digit recognition [1], has encouraged
researchers to consider them in Persian digit
recognition. Therefore, although many approaches
have been proposed in this area, but our focus was
only on gradient based one. For example, authors in
[2] achieved 99.15% recognition rate by extracting
features from gradient image on Hoda dataset [4].
Authors in [3] have used 16 directions instead of 8
directions which improved recognition rate by
0.32% (99.15% recognition rate) on the same
dataset. But it is slower than the base method. A
new classifier combining method has been
proposed by [5]. This model is composed of four
RBF networks as experts and another one as input
gate. The gate network learns how to divide input
space among experts. This method has reached
95.3% recognition rate and outperforms previous
network combination methods. Modified ChainCode Direction Frequency in surrounding pixels of
images has been employed by [6]. This paper
achieved 99.02% accuracy using multilayer SVM
neural network on Hoda dataset.
This paper presents a new feature extraction
approach for Persian handwritten digit recognition
problem based on Local Binary Pattern (LBP)
operator. LBP has been successfully applied in
486
context, object and face recognition problems. High
success rate of LBP is a good inspiration to use it in
Persian handwritten digit recognition. This
operator is the main feature extraction method in
the proposed approach. Also Multi-Layer
Perceptron (MLP) has been used as the base
learner. Experimental results show that our
proposed approach is more accurate and faster than
the previous ones.
The rest of this paper is organized as follows.
Section two gives a brief review of LBP operator.
The proposed approach is explained in section
three. Section four includes the experimental
results and the final section concludes the paper.
sampling points on a circle with radius R. Figure 2

shows different samples of this operator and
formula 1 explains this operator. , (, ) is
the interpreted decimal value of pixel (x, y).
shows the intensity value of the central pixel and
plays the same role for the neighboring pixels. S(x)
is calculated as follows:
LBPP, R ( x, y) i 0 s( g i g c )2 i
(1)
1 x 0
s ( x)
0 x 0
(2)
P 1
2 LOCAL BINARY PATTERN

Ojala et.al. [7] proposed Local Binary Pattern
(LBP) for texture description. LBP is one of the
most popular and successful binary descriptors in
context, object and face recognition [8-12]. This
operator is also robust to monotonic gray-scale
changes.
The basic LBP operator works with a 33 window.
Center of this window is placed on each pixel of the
image and is compared with other pixels in the
window. The results are put along and make a
binary string which is interpreted as a decimal
number. Histogram of these decimal numbers is the
output features. Figure 1 illustrates this process
with more details.
Figure 2: Extended LBP with different P and R
Uniform patterns are another extension of the basic

LBP operator [10]. A binary pattern is called
uniform if it has at most two changes from one to
zero and vice versa. This operator is presented with
LBPPu,2R . Although the number of uniform patterns to
the total number of patterns is 58/256 (23%) but it
is shown empirically in [10] that LBP8u,12 and LBP8,1
have almost the same performance. Histogram of
uniform LBP operator has only one bin for nonuniform patterns which leads to total number of 59
bins.
3 THE PROPOSED APPROACH
Figure 1: The basic LBP operator
The basic LBP operator has been extended to use

circular neighborhoods. This adjustment helps it to
catch more structural features of entities like faces,
objects, and characters [10]. This extended
versions sampling points are on a circle with
arbitrary radius. , refers to an operator with P
Our proposed approach is made of two main parts

which both of them are very critical. The first part
is image normalization that converts binary image
to grayscale and normalize its size. The second part
is feature extraction. LBP operator is used in a way
that both global and local features are extracted.
Finally these features are fed to a multi-layer
perceptron network. Figure 3 shows the explained
process diagram.
The proposed approach has several advantages.
The first and the most important one is its high
accuracy. Experimental results show that our
method is the most accurate among current
487
methods. The second advantage which make our
method easy to use in real application, is its
simplicity. This algorithm can be implemented very
fast and also has a high speed. Runtime report is
presented in the Experimental results section.
Figure 4: The used filter for smoothing and converting

binary image to grayscale image
Figure 5: Applying the filter to the left binary image and

converting it to a grayscale image
3.1.2 SIZE NORMALIZATION

LBP operator dictates that input images should have the
same size. Height to width ratio of Persian alphabet is
larger than one. Therefore, empirically, size of input
images has been set to 28 20 . Figure 6 represents this
process in more details.
Figure 3: The proposed preprocessing and feature extraction
approach diagram
3.1 PREPROCESSING
3.1.1 SMOOTHING AND CONVERTING TO
GRAYSCALE
LBP method can only be used for grayscale images,

therefore the input binary image should be
converted to a grayscale image. Figure 4 shows a
filter that can do this conversion. This filter can
smooth the input image and convert it to a grayscale
image. Figure 5 presents an example image of digit
eight (8). After applying this filter to the input
image, all the edges will be blurred and other
uniform areas remains the same. So LBP operator
will focus only on the edges and the non-edge parts
will have no place in the extracted histogram. Any
digit has its specific number and type of edges.
Therefore by extracting edges information, we can
discriminate different type of digits.
This method has several advantages over the

conventional methods. Persian digits like five and
zero have many structural similarities and one of
their major differences is their size. The proposed
approach preserves their size differences.
Moreover, this method preserves height to width
ratio that is a critical issue for some digits like one
and zero that can fill the whole image after resizing
and produce hard cases to classify.
Figure 6: The proposed algorithms flowchart for image size

normalization
488
3.2 FEATURE EXTRACTION
LBP abilities in Persian handwritten digit
recognition have not been evaluated yet. Uniform
LBP has been employed to obtain more compact
and efficient feature vector. This operator is applied
in two forms (figure 7). These two obtained
histograms are put along to get the final feature
vector which is used in the training process.
1. The uniform LBP operator is applied to whole
image to get a histogram with 59 bins.
2. After dividing the image into 4 equal blocks,
Uniform LBP is applied to each block and the
resulting histograms are concatenated.
Figure 8: Difference between single-region LBP and multiregions LBP
Multi-region LBP has been used because, singleregion LBPs output histogram is not discriminant
enough. For example consider figure 8 which
shows two Persian digits, seven (7) and eight (8)
which have reverse form of each other. Almost all
the edges in one image exist in the other one. If we
apply LBP to the whole image, it will produce
similar histograms. To overcome this problem, the
input image is divided into some blocks and LBP is
applied to them, therefore the results will no longer
be similar (figure 8).
Hoda dataset [13] has been used to evaluate and

compare the proposed approach to state-of-the-art
methods. The proposed approach obtained the first
place in the Persian handwritten digit competition,
which held during The First Conference on Pattern
Recognition and Image Analysis (PRIA 2013)
[13]. Results of this competition have been reported
in this section too. Hoda and Pars [13] datasets have
been used to weigh the registered systems in this
competition.
4.1 HODA DATASET
Hoda is the largest available dataset of handwritten
Persian digits. It is composed of 60000 train and
20000 test samples. For each digit, varying from
zero to nine, there is 6000 train and 2000 test
samples. This dataset can be obtained from [4].
Figure 9 shows some sample images from this
dataset.
Figure 9: Sample images from Hoda dataset

Figure 7: The proposed approach for feature extraction using
LBP operator
489
4.2 PARS DATASET
Table 1: Comparison of the proposed method and the best

previous methods
This dataset gathered by previously mentioned

competition [13] and has not been officially
released yet. It has 81863 samples which are
divided into 59451 train and 24412 test samples.
Figure 10 shows some sample images of this
dataset.
Figure 10: Sample images from Pars dataset
4.3 TRAINING PROCESS

A multi-layer perceptron, consisting of one hidden layer
with 30 neurons is used in the training process. 295
inputs (feature vector length) and 10 outputs (0 to 9) are
other specification of this network.
Hoda dataset has 60000 training samples. Because of
this large number of training samples, two different
approaches have been considered for training process:
1. Training at once: In this approach, the whole dataset
is fed to network in training process.
2. Circular training: In this approach, training set is
divided into some subsets. All subsets are used for
training in successive turns. This process is repeated
several times, which prevents the network from
forgetting the first subsets.
The proposed method accuracy has been compared

with several other methods. The results have been
summarized in table 1. The results indicate several
points. Circular training is almost as good as
training at once and running times of these two
approaches are very close. In some cases, training
with the whole dataset is not feasible. For example,
sometimes training dataset is relatively too large to
fit in the RAM. In these cases Circular training can
be a good alternative.
Method
Ref. [5]
Ref. [14]
Ref. [15]
Ref. [6]
Ref. [15] (5-fold)
Proposed Method
(Circular training)
Proposed Method
(At once training)
Precision (%)
95.30
97.99
98.71
99.02
99.37
99.59
99.72
Although the proposed method has the best accuracy

among the other methods, it is much simpler than the
others and has much less features which makes it very
fast. Table 2 and table 3 report the results of the
competition [13]. These tables compare the methods
proposed in [13] on Hoda and Pars datasets. Name of
the universities indicates only their representatives. As
the results show, the proposed method outperforms all
other methods on both datasets. Although these systems
are only trained on Hoda dataset, the proposed method
still has a better result than the others on Pars dataset.
Experimental results of [13] shows that our method

can recognize a digit in 180 milliseconds while it
takes 400 milliseconds for the second place team.
Table 2: Comparison of the proposed method on Hoda
dataset in competition [13]
Rank
4
3
2
1
Team
Shahid Beheshti University
Amirkabir University
Birjand & Hormozgan
University
Proposed Method
Precision (%)
94.63
95.94
97.39
99.59
Table 3: Comparison of the proposed method on Pars dataset

in competition [13]
Rank
4
3
2
1
Team
Shahid Beheshti University
Amirkabir University
Birjand & Hormozgan
University
Proposed Method
Precision (%)
96.07
97.81
98.94
99.53
5 CONCLUSION AND FUTURE WORK

In this paper, a new feature extraction method for
Persian handwritten digit recognition has been
490
proposed. Experimental results show that LBP
operator can extract very discriminating features
for handwritten digit recognition. Extraordinary
recognition rate (99.72%) of the proposed approach
makes it one of the best methods that has been
proposed in this research area.
Advantages of this method can be categorized as
follows:
High discriminative power of the extracted
features
Easy implementation
Very low feature extraction time
Very low computational overhead
On the other hand, one of the drawbacks of this
system can be the number of its features. Feature
vector length can be reduced to 50% of the current
length without noticeable performance fall. This
helps the system to operate about two times faster.
Methods like PCA, LDA and evolutionary
algorithms can be considered as good feature
reduction and selection methods. Also LBP
operator can be modified to be more suitable for
Persian handwritten digit recognition. We think it
will be a good idea to extend LBP and combine it
with a feature selection method.
[6]
[7]
[8]
[9]
[10]
[11]
[12]
7 REFERENCES
[13]
[1]
[2]
[3]
[4]
[5]
C.-L. Liu, K. Nakashima, H. Sako, and H.

Fujisawa, "Handwritten digit recognition:
benchmarking of state-of-the-art techniques,"
Pattern Recognition, vol. 36, pp. 2271-2285, 2003.
M. Nahvi, K. Kiani, R. Ebrahimpour, Gradient
feature extraction improvement based on discrete
cosine transform for handwritten Persian digits
recognition, 18th Iranian Conference on Electric
Engineering (ICEE2010), pp. 3066-3071, 2010.
H. Khosravi, A. A. Kabir, Introduction of two fast
and efficient features for handwritten Persian digits
recognition, 4th Iranian conference on Machine
Vision and Image processing, pp. 2023-2029,
2007.
H. Khosravi, E. Kabir, "Introducing a very large
dataset of handwritten Farsi digits and a study on
their varieties", Pattern Recognition Letter, vol. 28,
pp. 1133-1141, 2007.
R. Ebrahimpour, A. Esmkhani, and S. Faridi,
"Farsi handwritten digit recognition based on
[14]
[15]
[16]
[17]
mixture of RBF experts," IEICE Electronics

Express, vol. 7, pp. 1014-1019, 2010.
A. Alaei, P. Nagabhushan, and U. Pal, "Fine
classification of unconstrained handwritten
persian/arabic numerals by removing confusion
amongst similar classes," in Document Analysis
and Recognition, 2009. ICDAR'09. 10th
International Conference on, 2009, pp. 601-605.
T. Ojala, M. Pietikinen, and D. Harwood, "A
comparative study of texture measures with
classification based on featured distributions,"
Pattern Recognition, vol. 29, pp. 51-59, 1996.
H. K. Ekenel, M. Fischer, E. Tekeli, R.
Stiefelhagen, and A. Ercil, "Local binary pattern
domain local appearance face recognition," 2008,
pp. 1-4.
D. Maturana, D. Mery, and . Soto, "Face
recognition with local binary patterns, spatial
pyramid histograms and naive Bayes nearest
neighbor classification," 2009, pp. 125-132.
T. Ojala, M. Pietikainen, and T. Maenpaa,
"Multiresolution gray-scale and rotation invariant
texture classification with local binary patterns,"
Pattern Analysis and Machine Intelligence, IEEE
Transactions on, vol. 24, pp. 971-987, 2002.
L. Snchez Lpez, "Local Binary Patterns applied
to Face Detection and Recognition," 2011.
H. M. Vzquez, E. G. Reyes, and Y. C. Molleda,
"A new image division for LBP method to improve
face recognition under varying lighting
conditions," 2008, pp. 1-4.
Birjand University, "Competitions Report," The
First Iranian Conference on Pattern Recognition
and Image Analysis (PRIA 2013), 2013.
R. Ebrahimpour and S. Hamedi, "Hand written
digit recognition by multiple classifier fusion
based on decision templates approach," in The
International Conference on Computer, Electrical,
and Systems Science, and Engineering, 2009, pp.
245-250.
A. Alaei, U. Pal, and P. Nagabhushan, "Using
modified contour features and SVM based
classifier for the recognition of Persian/Arabic
handwritten numerals," in Advances in Pattern
Recognition,
2009.
ICAPR'09.
Seventh
International Conference on, 2009, pp. 391-394.
G. E. Hinton, M. Revow, and P. Dayan,
"Recognizing handwritten digits using mixtures of
linear models," Advances in neural information
processing systems, pp. 1015-1022, 1995.
T.-F. Gao and C.-L. Liu, "High accuracy
handwritten Chinese character recognition using
491
LDA-based compound distances," Pattern
Recognition, vol. 41, pp. 3442-3451, 2008.
[18] L. S. Oliveira, R. Sabourin, F. Bortolozzi, and C.
Y. Suen, "Feature selection using multi-objective
genetic algorithms for handwritten digit
recognition," in Pattern Recognition, 2002.
Proceedings. 16th International Conference on,
2002, pp. 568-571.
492
Navigation Error Reduction in Swarm of UAVs

Shibarchi Majumder1, Rahul Shankar2, Mani Shankar Prasad3
1,3
Amity Institute of Space Science & Technology, INDIA
2
Amity Institute of Nuclear Science & Technology, INDIA
1
shibarchi.majumder@student.amity.edu, 2rahul.shankar@student.amity.edu, 3msprasad@amity.edu
1. ABSTRACT
For large area surveillance, a number of UAVs are
preferred. In such situation, navigation becomes a
prime factor, which reduces the chance of collision.
UAVs primarily depend on GPS, which has an order of
inaccuracy. Conventional correction methods like
differential GPS needs extra hardware and cannot
improve stochastic errors like multipath error and
receiver error. In this paper we demonstrate a novel
technique to reduce the GPS error, in a swarm of
UAVs.
2. KEYWORDS
GPS Error, Navigation Error, GPS Error Reduction,
UAV Swarm, Position Correction, Error offset, Intra
Swarm Communication;
3. INTRODUCTION
A group of cooperative UAVs, also called a
swarm of UAVs sometimes very effective for
certain tasks like mapping, target searching etc.
Portability,
lesser
take-off
and
landing
requirements, easy handling along with low
manufacturing and maintenance costs make it
more desirable than expensive and large UAVs.
However, the low cost and small size principle
puts a limit to the onboard sensors, resulting in a
significant inaccuracy in navigation with others.
IMU (Inertial Measurement Unit) is a popular
solution for such problem, but precise IMUs are
expensive and heavy. On the other hand, GPS,
which is a less expensive solution, comes with a
lower accuracy. Commercial GPS provides an
accuracy of the order 15 to 100 meters.
#
Correspondence: shibarchi@hotmail.com/ msprasad@amity.edu
A widely used method to reduce GPS error is to

put reference stations to correct the GPS data
(DGPS), which limits the operational area as well
as increases the cost.
Unlike inertial navigation system, error
distribution in GPS remains almost linear over
time. However, it could be difficult to estimate
some error parameters accurately due to its
modeling constraints. Errors like SA (Selective
Availability), Ionosphere delay, propagation delay
remains same over a large area, while receiver
error and multipath error varies with different
hardware and location respectively. Receiver error
could be caused by receiver thermal noise or
receiver clock error and its hard to model receiver
error due to its stochastic nature. In most cases, the
clock is a quartz-crystal oscillator whose accuracy
is of the order of 1-10 ppm. Multipath Propagation
Error (MPE) is caused when objects in the vicinity
of the receiver antenna reflect GPS signals,
resulting in one or more secondary propagation
paths. These secondary-path signals superimpose
on the desired direct-path signal and distort its
amplitude and phase. Errors in multipath cannot be
reduced by DGPS, since they depend on local
reflection geometry near the receiver antenna. In a
receiver without multipath protection, C/A code
ranging errors of 10m or more can be experienced.
Multipath errors not only cause large code ranging
errors but also severely degrade the ambiguity
resolution process required for carrier phase
ranging such as that used in precision surveying
applications.
This paper demonstrates a new technique to
reduce error in GPS for networked swarm of
UAVs, where the accuracy in position
493
measurement is improved using the error in
relative distance between every member UAV in
the swarm.
sender UAV calculates the distance of other

UAVs.
4. INTRA-SWARM COMMUNICATION
To reduce the error in position measurement by
correlating the GPS data with other members in
the swarm, it is necessary to determine the relative
distance between the members as well as to get the
GPS provided data for every UAV.
Figure 2: A timeline showing the various delays in packet
transmission
4.1. Total Propagation Time

( )
Figure 1: An example of inter UAV communication. In a

system of three UAV's, each UAV has two different
communication links with other UAVs.
Where,
is response pulse receiving time
is request pulse transmitting time and
is predefined delay time.
Therefore, distance between two UAVs is,
(
)
( )
where
4.2. Navigation Error Deduction
Each UAV of the swarm is equipped with a

transponder
which
is
capable
of
transmitting/receiving the GPS data with other
UAVs. For easy identification, a simple data
format of unique ID and time delay ( ) of
transmission is used.
To calculate the distance between each member in
the swarm, each UAV is featured with a DME
type transponder with a unique bandwidth for
identification. As the operational area is too small
for the propagation time of the EM pulse across
the distances to be significant, a predefined delay
time is assigned to every transponder. In a
sequential manner, each UAV sends a request
pulse and after receiving the pulse other UAVs
wait for the pre-determined delay assigned to them
and send back a response pulse. From the known
bandwidth and delay assigned to each UAV, the
The DME measurement gives the real time

relative distance measurement of the UAVs in the
vicinity. Although GPS data received through
trans-receiver can also give a calculated relative
distance, there may be in errors or inaccuracies in
them (inherent due to C/A code errors).
Figure 3: Diagram showing measurement of error between

measured and calculated distances
4.3. Calculated Distance Equation

( )
494
where
and
are distance measured and
distance calculated respectively. Errors between
all UAVs are calculated by the above equation.
For a network of n number of UAVs,
number of error equations can be formed and error
for each UAV can be calculated.
5. CORRECTION
and receiver clock error, the error in distances is

Gaussian in nature. Given these parameters, a
Gaussian model probability density function is
associated with each UAV location data whose
mean value is located at the GPS indicated center.
Thus the probability of finding the UAV
at a
point whose distance from the indicated center is
would be given by
(
The correction process is divided into three stages,

namely the primary, secondary and the tertiary
correction stages. However, not all UAVs in the
swarm will follow through with all three stages
since this would be inefficient in terms of
processing load as well as accuracy of results as
erratic values may creep into the system due to
large number of corrections made.
( )
The value of is set for each UAV such that the

chances of occurrence of the true location of the
UAV outside the circle of radius tends to zero
i.e.
5.1. Correction Criteria
5.3. Correction Stage I
The question of which swarm member should

follow which correction stages is decided using
the sum-distance rule.
Corrections based on relative distances are most
effective for the UAV which is surrounded on all
directions by other UAVs and has the minimum
summed up distance from others.
Stage I correction involves estimating the most

probable location of the UAV within the margin of
GPS error by isolating the region within which the
UAV must be in order to agree with the distance
measured by every other member.
The location of each UAV
can be corrected
using the measured distance between
and
( ) . For UAV
another UAV
, the
( )
th
where
is the sum of the distances of the j
UAV from all the UAVs and
is its distance
th
from the i UAV. From the above calculation the
middle UAV (whose value of
is minimum) is
determined. Once this is done, the criterion is
applied as follows:
( )
( )
5.2. Preliminary Calculation
Before the correction process, certain preliminary
calculations must be done in order to facilitate
them. As we are considering only multipath error
( )
( )
( )
distribution function of another UAV ( ) is

reconstructed at a point represented by the position
vector from the GPS center of
given by
( )
(7)
Where is the unit vector in the direction of
and ( ) is the error from the calculated distance
equation between and .
Since ( ) is always less than GPS error radius
( ) for any pair of members
and , there will
exist a region of overlap of the two probability
distribution functions, which collectively give rise
to a local sum distribution function as shown in
Figure 4.
495
5.4. Correction Stage II
Stage II correction is performed on the middle
UAV ( ) after all the members have corrected
centers ( ) after execution of stage I.
New error values are calculated between the
calculated distances between members corrected
centers( ), and their measured distances
( )
( )
( )
( )
( )
Let there be a point represented by the position
vector with respect to ( )
( )
( )
Figure 4: The figure shows the reconstruction of the
probability distribution function of the location of
near
and the region of overlap
Similarly, the probability distribution functions of

the locations of all UAVs
in the
swarm are reconstructed around the GPS center of
and are summed up. Since we know that the
true location of
must lie within the region of
overlap, this sum function is normalized within
this region.
Let
be the set of all points within the overlap
region. A new probability density function is
assigned to the location. This is given by
( )
{
(
(
( )
)
( ))
( )
Where is the position vector of a point from the

center of
and () is the probability density
function at that point.
The corrected GPS centre ( ) of the UAV
is
given by the centre of gravity of the distribution
( )
(
(
( ( )
))
( ))
( )
Therefore, the corrected middle members location

will now be given by:

( )
After the stage II correction, the error in location
has been reduced to its minimum as reducible by
our algorithm. Hence, this location is considered
to be the true location.
5.5. Correction Stage III
Now that
has a correct location pinpointed
( )
( )
Where
( )
Where
is the direction angle of
as provided
by the antenna, and and are the unit vectors
along the positive and axes.
If is the position vector of the location of
as calculated from the current centre, then the
position vector of its accurate location will be
given by:

()
( )
Therefore, the GPS center of each non-middle
member
of the swarm is shifted by () to
complete the error reduction process for the whole
swarm.
496
6. TEST RESULTS
In order to test the algorithms accuracy, five GPS
receivers were placed at different locations (whose
location coordinates are known) across the Amity
University campus and the distances between
these locations were calculated. The data was
provided to the algorithm for error reduction. The
results obtained are as shown in figures 5, 6 and 7.
Figure 7: The figure shows the results of GPS error

reduction. Here, receiver 4 (refer figure 5) was taken as the
middle member and the black triangles represent the
corrected locations provided by the algorithm.
Figure 5: The figure shows the various locations across the

Amity University campus where GPS receivers were placed.
The coordinates shown were of locations manually
pinpointed on google maps.
The results shown are of a single run of the

algorithm. The error at each location is similar
since stage three corrections is simply a reverse
engineering of the location of the members of the
swarm based on measured distance and directional
data. The algorithm was run 20 more times and
each time, absolute errors were calculated. The
results are graphically represented in figure 8.
Figure 8: Absolute errors obtained for corrected positions of

middle members (represented by grey squares) and average
absolute errors in indicated positions (represented by black
rhombi)
Figure 6: In this figure, the red rhombi represent the GPS
indicated positions of the receivers placed at locations
represented by the yellow stars. In other words, the yellow
stars are the true positions and the red rhombi are the
indicated positions.
497
Location
Indicated
position
True position
Corrected
position
77
77
77
77
77
77
77
77
77
7
7
77
Abs.
error
3.
77
77
77
7
7
77
77
Table 1: The table shows the true coordinates (obtained

from google maps), indicated coordinates (from the
receiver), and the corrected coordinates of the positions of
the receivers in serial order, along with the absolute error on
the corrected position.
4.
5.
7. CONCLUSION
6.
In this paper a way to reduce the GPS position

error has been explained. The above mentioned
way is not only able to reduce the receiver error
and multipath error but also able to operate in nonGPS-connectivity.
During the scenario when the swarm must explore
a region where there is no clear sky visibility, it
was seen in our trials that if even a single UAV in
the swarm had a low intrinsic GPS error, the
accuracy of the estimated locations of all other
members of the swarm are greatly heightened, thus
proving very useful for isolated explorations.
In the near future we shall attempt to perform
corrections on GPS velocity for a networked
swarm of UAVs.
7.
8.
9.
10.
11.
12.
8. ACKNOWLEDGEMENT
This research was done during post graduation
phase, as a part of UAV Systems in Institute of
Space Science and Technology, Amity University
Uttar-Pradesh, INDIA.
13.
receiver clock errors: impact on positioning

accuracy. Measurement Science And Technology,
Uk, 2009.
K D Park, J L Davis, Per O J Jarlemark, P Elosegui,
J E Normandeau, B E Corey, A E Niell, C E
Meertens, and V A Andreatta. Multipath
characteristics of GPS signals as determined from
the Antenna and Multipath Calibration System
B R K Rao, A D Sharma, Y R Kumar. Technique
to reduce multipath GPS signals Current Science,
Vol. 90, No. 2, 25 January 2006
M S Grewal, L R Weill, A P Andrews Global
Positioning Systems, Inertial Navigation, and
Integration 2004
G Xu GPS: Theory, Algorithms and
Applications 2007
B H Wellenhof, Lichtenegger, E Wasle GNSS
Global Navigation Satellite Systems 2007
Kitagawa, Monte Carlo lter and smoother for
non-Gaussian nonlinear state space models,
Journal of Computational and Graphical Statistics
5(1), pp. 125, 1996.
D Fox, W Burgardy, F Dellaert, S Thruny. Monte
Carlo Localization: Efcient Position Estimation
for Mobile Robots. American Association for
Articial Intelligence, 1999
R C Arkin, T Balch. Cooperative multiagent
robotic systems. Editors: D Kortenkamp, R P
Bonasso, and R Murphy, Articial Intelligence and
Mobile Robots. MIT/AAAI Press, Cambridge, MA,
1998.
T R. Balch. The impact of diversity on
performance in multi-robot foraging. In Agents
99, Seattle, WA, 1999.
T S Rappaport, Wireless Communications:
Principles and Practice, Pearson Education India,
2009.
P Z. Peebles, Radar Principles, John Wiley &
Sons, 01-Jan-2007
9. REFERENCES
1.
2.
B R Townsend, P C Fenton. A Practical Approach

to the Reduction of Pseudorange Multipath Errors
in a Ll GPS Receiver. Townsend & Pat Fenton,
1994
T K Yeh, C Hwang, G Xu, C S Wang and C C Lee.
Determination of global positioning system (GPS)
498

IJDIWC - Volume 4, Issue 4

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

IJDIWC - Volume 4, Issue 4

Uploaded by

Copyright:

Available Formats

A MULTI TREE BASED APPROACH FOR PERFORMANCE ANALYSIS IN

ANALYTICAL APPROACH OF COST-EFFECTIVE AND SECURE SIPChulhee Cho, Jae-Young Choi,

FALL DETECTION SYSTEM DESIGN BY SMART PHONE

Yung-Gi Wu, Sheng-Lun Tsai

PERSIAN/ARABIC HANDWRITTEN DIGIT RECOGNITION USING Mohsen Biglari, Faezeh Mirzaei,

Shibarchi Majumder, Rahul

NAVIGATION ERROR REDUCTION IN SWARM OF UAVS

WSN based AFPS provides instant bullet

Target Boxes with

Hi-Gain 14 dBi Outdoor

capability. Target boxes were Powder Coated with

1.1 Target Box Assembly

1.2 Commander Console

The target units connect to the commander and

1.4 Target Box Assembly

Figure 2: Automatic Target Box

Bullet-Impact Sensor: The most critical

Figure 3: Target Box Assembly

Target box with mil specs circular connectors is

Figure 4: Circular Connectors for AFPS Target Box

1.5 Bullet-Impact Sensor

the target sheet where the PVS is located was

Figure 5: Bullet-Impact Sensor Mechanical Placement

Target enclosure has a base-box and a target sheet.

were used for testing. Amplified PVS signals were

Figure 6: Amplified Response of Bullet-Impact Sensor

The finalized system was installed at 10 different

Figure 7: Control Board Circuit Design

The main parts of Control board circuit are:

Bullet Impact Counter: Analog comparator

Motor Driver: IRF540 Power MOSFETs were

1.7 WiFi Communication Module

Figure 8: Moxa WiFi Communication Module

DESIGN OF COMMANDER CONSOLE

Figure 9: Commander Console Labview GUI

Figure 10: AFPS Commander Console

down on each bullet hit but popup again

Figure 11: Bullet-Impact Over-Count & Corrected Signal

A testing methodology was established in which

of AFPS was assembled at the firing range with a

To further improve the efficiency of the AFPS,

Figure 14: AFPS - Third Field Testing Results

Figure 12: AFPS - First Field Testing Results

Bullet-Impact sensing circuit was improved

Figure 15: AFPS - Fourth Field Testing Results

Figure 13: AFPS - Second Field Testing Results

The four phases of AFPS field testings and

Figure 16: AFPS Average Software Count Accuracy

SCALCOM-EMBEDDEDCOM'09., pp.566-570, 25-27

Information Transparency in Admission Selection of State Islamic Higher Education

Right to information, Transparency, Continuous

3. Marking process (Flagging) any of unfair or

Figure 1. SPMB-PTAIN Admission Process

2. DATA PROCESSING FUNCTION IN

To run the workflow described in figure 2, the

scoring and passing team for scoring and rank

Figure 2. SPMB-PTAIN data processing

The processes that starts from LJU delivery

4. DATA MATCHING CONTROL IN DATA

From several issues that occur as shown in Table

c. Attendance list of examinees

PROCESSING SYSTEM (SMUPDA)