Final

Acknowledgement
First and foremost, I offer my sincerest gratitude, to the Almighty, for bestowing me with an opportunity to complete my work.
I am thankful to our Principal, Prof.Shine P. James, for his valuable support

and advice. I would also like to thank our HOD, Mr.Yasim Khan M. Asst. Prof.
ECE Department, for sharing his invaluably constructive criticism and friendly
advice during the project work.
I am highly indebted to our MTech coordinator Ms.Flower Abraham, Asst.

Prof. ECE Department for giving necessary guidance and arranging all facilities
to complete my work.
I would like to express my gratitude to my project guide Ms.Ajishya Liza

Idiculla, Asst. Prof. ECE Department whose support and supervision helped
me to coordinate my project.
I am indebted to all teaching and non-teaching staff of the Department of Electronics Engineering for their co-operation and support. Last but not the least I wish
to express my sincere thanks to all my friends for their goodwill and constructive
ideas.
Abstract
Demand for intelligent systems for vehicle control is increasing in the recent years, to ensure safeness and convenience. The
rapid increase of on road vehicles and also the rise in human needs
have paved way for the same. Numerous intelligent systems were
developed so for the identification of vehicles. This work is dependent on image processing techniques for vehicle identification. It
includes two portions, a descriptor and a classifier. Log Gabor filter descriptors are used for the extraction of features and effective
edge detection of the vehicles in the captured frame. The objects are
classified as vehicle or non-vehicle with an SVM classifier. Number
plate extraction is appended to vehicle identification, to improve the
relevance of the work. A pre-trained neural network is employed to
extract the characters in number plates of identified vehicles.
Contents
INTRODUCTION
1.1
Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2
Organization of the project . . . . . . . . . . . . . . . . . . . . .
LITERATURE SURVEY
2.1
Vehicle detection . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.1
Background subtraction method . . . . . . . . . . . . . .
2.1.2
Using object features . . . . . . . . . . . . . . . . . . . .
2.1.3
Object based approach . . . . . . . . . . . . . . . . . . .
2.1.4
Color transform model . . . . . . . . . . . . . . . . . . .
2.1.5
Detection using shadow . . . . . . . . . . . . . . . . . .
License plate extraction . . . . . . . . . . . . . . . . . . . . . . .
2.2
PROBLEM FORMULATION
13
3.1
14
Block diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . .
iii
3.2
3.1.1
Image acquisition . . . . . . . . . . . . . . . . . . . . . .
15
3.1.2
Image enhancement and gray scale conversion . . . . . .
16
3.1.3
Log gabor filters . . . . . . . . . . . . . . . . . . . . . .
17
3.1.4
Support Vector Machine classifiers . . . . . . . . . . . . .
18
ALGORITHM . . . . . . . . . . . . . . . . . . . . . . . . . . .
20
3.2.1
Vehicle detection . . . . . . . . . . . . . . . . . . . . . .
20
3.2.2
License plate extraction . . . . . . . . . . . . . . . . . .
21
3.2.3
Verification . . . . . . . . . . . . . . . . . . . . . . . . .
22
3.2.4
Number extraction . . . . . . . . . . . . . . . . . . . . .
22
WORK DONE AND RESULTS
23
4.1
Vehicle identification . . . . . . . . . . . . . . . . . . . . . . . .
23
4.2
Number plate extraction . . . . . . . . . . . . . . . . . . . . . .
28
CONCLUSION
30
FUTURE SCOPE
31
List of Figures
3.1
Block Diagram of Proposed Method . . . . . . . . . . . . . . . .
15
3.2
Sample input images . . . . . . . . . . . . . . . . . . . . . . . .
16
4.1
Sample input image
. . . . . . . . . . . . . . . . . . . . . . . .
23
4.2
Gray scale enhanced image . . . . . . . . . . . . . . . . . . . . .
24
4.3
Log gabor filter output . . . . . . . . . . . . . . . . . . . . . . .
24
4.4
Selecting camera pose . . . . . . . . . . . . . . . . . . . . . . . .
25
4.5
Input images are selected . . . . . . . . . . . . . . . . . . . . . .
25
4.6
Result for single vehicle image . . . . . . . . . . . . . . . . . . .
27
4.7
Result for single non-vehicle image . . . . . . . . . . . . . . . .
27
4.8
Neural network training completed . . . . . . . . . . . . . . . . .
28
4.9
Detected number KL05J8798 . . . . . . . . . . . . . . . . . . . .
28
4.10 Detected number SAP729 . . . . . . . . . . . . . . . . . . . . . .
29
4.11 Detected number 5RLS . . . . . . . . . . . . . . . . . . . . . . .
29
4.12 No number detected . . . . . . . . . . . . . . . . . . . . . . . . .
29
List of Tables
4.1
Calculated feature sets for vehicle samples . . . . . . . . . . . . .
26
4.2
Calculated feature sets for non-vehicle samples . . . . . . . . . .
26
vi
List of Abbreviations
ADAS
Advanced Driver Assistance Systems
ANPR
Automatic license plate recognition
BAM
bidirectional associative memories
CCTV
Closed Circuit Television
HOG
Histogram Oriented Graph
ITS
intelligent transportation system
LPR
License Plate Recognition
PCA
Pricipal Component Analysis
ROI
Region of Interest
SVM
Support Vector Machine
SVM
Support Vector Machine
vii
College of Engineering Poonjar
Image based vehicle verification
Chapter 1
INTRODUCTION
Vehicle identification systems find high importance in the intelligent traffic
systems. The increase in the number of vehicles and the increase in the human
population and needs justify the same. Thus, as days pass, the complexity of
controlling and identifying vehicles is increasing and much difficult to solve. Intelligent transportation systems find wide application at such scenario since they
improve mobility, productivity and foremost the safety.
Vision-based object detection from a moving platform becomes particularly

challenging in the field of ADAS . It is an overall system that provides a driver
with essential information, automate difficult or repetitive tasks, and lead to an
overall increase in car safety and comfort. These safety features are designed to
avoid collisions and accidents.In this context, onboard vision-based vehicle verification strategies become critical, facing challenges derived from the variability
of vehicles appearance, illumination, and vehicle speed.
With the aim of reducing injury and accident severity, precrash sensing is becoming an area of active research among automotive manufacturers, suppliers and
universities. A detailed study of the same can be found in [27]. These efforts have
produced several prototypes and solutions, based on rather different approaches.
Department of Electronics & Communication Engineering
Visual traffic surveillance has attracted significant interest in computer vision, because of its tremendous application prospect. A traffic surveillance system needs
to detect vehicles and classify them if possible.
Techniques for vehicle identification implemented with image processing have

high demand today. The grounds are the flexibility, reduced cost and the ease
of implementation. The complexity increases since the vehicles are of different
shapes, colors size etc. A key to this issue covets with a system employing a descriptor and a classifier. The habitual methods employ generation of hypothesis
and further verification [2]. In the former stage, the vehicle portions are pointed
from the captured frame. The vehicle portions are found using several features
like color[3], edges[4], motion[5] etc. At the later stage, the identified feature sets
verify the result. Nowadays, learning based methods find wide application since
the processor speeds has increased. Such methods usually classify the objects to
two classes, where a bunch of samples are trained for desired features that enable
classification. These methods involve the need of descriptors and usual descriptors include the HOG[7][8], PCA[6] etc.
Gabor filters forms such a descriptor which typically uses the statistics like
mean and variance for classification of the detected objects [10]-[13]. Their relevance is reflected in the work of Daugman[9], which describers the similarity of
the Gabor functions to that of visual system in mammals and hence in human.
This has aroused its application including image compression [19], segmentation
[20], [21], scanning and retrieval of image data [16], [15], object tracking [24],
texture classification [18], [22], and [23] and extracting features for further classification [14], [15], and [25] Other than these merits, several drawbacks exist
associated with bandwidth and information redundancy. These are used with var-
ious direction and scales for performing feature extraction. Because of limited
bandwidth, several filters are required for wide spectrum coverage. Also, a major
Gabor response is low frequency.
This work employs Log Gabor filter descriptor introduced by Field [26], to
overcome the disadvantages mentioned above. These have 0 DC components and
hence the risk of redundant information in the lower frequency range is avoided.
Symmetry is an important feature of this descriptor in log axis in place of linear
frequency axis. The equation below, gives its frequency response.
LGm,n = (f, ) =
1.1
exp
2
log ff
2(log)2

exp
( n )
2
2

(1.1)
Objective
New information and control technologies that make vehicles smarter are now
arriving on the market either as optional equipment or as specialty after-market
components. These technologies are being developed and marketed to increase
driver safety, performance, and convenience. It is under this circumstance that the
ADAS systems find more application and the interest on vehicle detection based
on image analysis find more importance due to its low cost, flexibility and processing capabilities.
The main objective of this work is to detect the vehicle using Log Gabor filter.
Vehicle detection based on image analysis mainly depends on classifying the object in the captured image as either vehicle or non-vehicle.
To detect any vehicle, edge detection and feature extraction are the important
two stages. Feature extraction can be done with a help of a Gabor filter bank at
different scales and orientations. In this work a bank of 8 Gabor filters with 2
different wavelengths and 4 different orientations is used. The support vector machine classifiers are used for feature extraction. The SVM classifiers are used to
classify the needed set of information from the training set of data.
Later, optical character recognition technique has been employed to extract

the number plates of the identified vehicle. This is implemented with the help of
a pre-trained nueral network.
1.2
Organization of the project

This thesis report is organized as follows. Chapter 2 gives a summary of
literature survey. The following chapter describes the problem formulation and
algorithm. The discussion of the result is shown in chapter 4 and concluding the
report by chapter 5. The future scope of the project is discussed in chapter 6.
Chapter 2
LITERATURE SURVEY
There are various types of approaches in traffic surveillance; one of the most
common approaches is vision based. This approach more often used to analysis
vehicle from aerial images or videos. The vision of vehicle will differ according to
the camera points, illumination conditions and backdrop situations. This chapter
provides the comparative study on technical analysis of existing methods which
is extended to overcome the disadvantages of the techniques mentioned earlier.
2.1
2.1.1
Vehicle detection
Background subtraction method
Background extraction is an efficient preprocessing technique for traffic observation and surveillance. It reduces the amount of insignificant information in the
image sequences and thus accelerates the processing. Background images tend to
be motionless over a long period of time, and moving object images only contain
foreground objects. Change detection [12], [13] is the simplest method for segmenting moving objects. An adaptive background update method [14] has been
proposed to obtain background images over a given time period. He et al. [15] apDepartment of Electronics & Communication Engineering
plied the Gaussian distribution to model each point of a background image, using
the mean value of each pixel during a given time period to assemble the image.
However, all these methods had drawbacks with the efficiency of the overall result
obtained.
2.1.2
Using object features
Vehicle images observed from rear or frontal view are in general symmetrical in
horizontal and vertical directions. This observation was used as a cue for vehicle
detection in the early 90s [16]. An important issue that arises when computing
symmetry from intensity, however, is the presence of homogeneous areas. In these
areas, symmetry estimation is sensitive to noise. In [17], information about edges
was included in the symmetry estimation to filter out homogeneous areas. When
searching for local symmetry, two issues must be considered carefully. First, we
need a rough indication of where a vehicle is probably present. Second, even
when using both intensity and edge maps, symmetry as a cue is still prone to false
detections, such as symmetrical background objects, or partly occluded vehicles.
Although few existing systems use color information to its full extent for HG,
it is a very useful cue for obstacle detection, lane/road following, etc. Several
prototype systems investigated the use of color information as a cue to follow
lanes/roads, or segment vehicles from background [18]. Similar methods could be
used for HG, because non-road regions within a road area are potentially vehicles
or obstacles. The lack of deploying color information in HG is largely due to
the difficulties of color-based object detection or recognition methods in outdoor
settings. The color of an object depends on illumination, reflectance properties of
the object, viewing geometry, and sensor parameters. Consequently, the apparent
color of an object can be quite different during different times of the day, under
different weather conditions, and under different poses.
Using shadow information as a sign pattern for vehicle detection was initially
discussed in [19]. By investigating image intensity, it was found that the area underneath a vehicle is distinctly darker than any other areas on an asphalt paved
road. A first attempt to deploy this observation can be found in [20], though there
was no systematic way to choose appropriate threshold values. The intensity of
the shadow depends on the illumination of the image, which in turn depends on
weather conditions. Therefore the thresholds are not, by no means, fixed. In [21],
a normal distribution was assumed for the intensity of the free driving space. The
mean and variance of the distribution were estimated using Maximum Likelihood
(ML). It should be noted that the assumption about the distribution of road pixels might not always hold when true. For example, rainy weather conditions or
bad illumination conditions will make the color of road pixels dark, causing this
method to fail.
2.1.3
Object based approach
As mentioned in [2], Ashley C. Holt, Edmund Y.W. Seto et al which presents object oriented approach for detecting and classifying vehicles from high-resolution
aerial videos. This approach has three steps for detecting the vehicles.
Segmentation
Training
Validation
For first step it takes some training images from total number of aerial view
images depending upon time or number of frames. Then it proceed with optimization of segmentation parameters for segmentation and is repeated for training
samples which involve multi resolution segmentation and spectral difference segmentation then it transfer to validation part of object accuracy assessment with the
training samples. In this model the circumstance of applicability to high spatial

resolution tenuously sensed data, and to address the need for a quantitative, usersupervised method for choosing best segmentation parameters. It developed an
objective metric which is the number of training object matched with maximize
area matched and which is minimizes below and over segmentation for preferred
images in objective primitives. By using this method the authors identify good
result for static and gray scale images.
2.1.4
Color transform model
Luo-weitsai, Jun-weihsieh et al[5] proposed a new method for vehicle detection using color transform model and edge map. This method introduces different
phases for detection. In the first phase a color transform model recognize vehicle
pixels from the image backgrounds. This technique reduced number of random
samples unlike previous methods it takes only less number of images for training. All color participation pixels are projected on color space. In the second
phase a classification is carried out using Bayesian classifier which detects pixels
corresponding to the vehicle or not and it generates different vehicle premise for
each detected vehicles. The final phase is validation of hypothesis by coefficient
of wavelet transform, edges and corners of images. This method is efficient and
takes less number of training samples and results in accurate vehicle detection on
static images.
2.1.5
Detection using shadow
4. T. Gandhi et al. [4] discusses the issues and approaches involved in developing a mobile vehicle-mounted system to detect, classify, and log the surrounding
vehicles in a database for efficient query-based retrieval. This system consists of
three components,
1. Vehicle sensing, detection and tracking

2. Feature extraction and classification
3. Database storage and retrieval
Relevant research in each of these components is described in order to guide the
development of such a system. Such a system has applications including traffic analysis from mobile probes, analyzing driver behavior based on surrounding
vehicles, as well as surveillance from mobile platform.
2.2
License plate extraction
One of the most important topics of ITS is the LPR and Number plate extraction is hotspot research area in the field of image processing. Automatic license
plate recognition system is to extract vehicle license plate from a digital image.
ANPR system has been intensively studied in many countries. Due to the different types of number plates being used, the requirements of an automatic license
plate recognition system are different for each country. LPR systems have many
potential applications in intelligent traffic systems, such as the payment of parking fee, highway toll fee, traffic data collection, traffic monitoring systems, traffic
law enforcement, and security control of restricted areas etc. This type of application puts high demands on the reliability of the System. The purpose of this
research is to develop an application which recognizes license plates from vehicles. Generally, LPR was developed to identify vehicles by the contents of their
license plates. The system takes images of automobiles as input and processes
them. Once a license plate is detected, its digits are recognized and displayed on
the User Interface. This work focuses on the design of a single algorithm used
for extracting the license plate from a single image, isolating the characters of the
plate and identifying the individual characters.

ANPR recognizes a vehicles plate number from an image by digital camera.
It is fulfilled by the combination of a lot of techniques such as image acquisition
i.e. capturing the image of real image of plate localizing the license plate character
segmentation i.e. locating and identifying individual character on the plate, optical
character recognition. The recognition problem is generally sub-divided into four
parts are Image acquisition i.e. capturing the image of the license plate, Preprocessing the image i.e. localizing the license plate, Character segmentation
i.e. locating and identifying the individual symbol image on the plate, Optical
character recognition. For example, in India the norm is printing the license plate
number in black color on white background for private vehicles and on a yellow
background for commercial vehicles. Number plate is a pattern with very high
variations of contrast. If the number plates are very similar to the background it is
difficult to identify the location. The morphological operation is reused to extract
the contrast feature withinin the plate [29].
Major works in license plate recognition includes image segmentation.
Here a region of interest is obtained and the further extraction is done based on
this. Some of the previous works done in this area is mentioned below.
Lotufo, Morgan and Johnson proposed automatic number-plate recognition using optical character recognition techniques. Johnson and Bird proposed
knowledge-guided boundary following and template matching for automatic vehicle identification. Fahmy proposed BAM neural network for number plate reading. Its appropriate for small numbers of patterns. Nijhuis, Ter Brugge, Helmholf
J.P.W. Choi and Kim proposed the method based on vertical edge using Hough
transform (HT) for extracting the license plate. E.R. Lee, P.K. Kim and H.J. Kim
used neural network for color extraction and a template matching to recognize
characters. Yoshimura and Etoh used Gabor jets projection to form a feature vec-
10
tor for recognizing low resolution gray-scale character. Hontani et.al. proposed
a method for extracting characters without prior knowledge of their position and
size in the image[31].
Chittode J S et al. developed algorithm which is applied on the car park systems to monitor and manage parking services. Algorithm is developed on the basis
of morphological operations and used for number plate recognition. Optical character is used for the recognition of characters in number plate. Peng H et al. presented an algorithm named Document Image Recognition. DIR is most effective
approach which is used to find most similar template for input image in a database.
The algorithm is developed on the basis of global matching of CBP Chunyu C et
al. presented a technique for recognition of number plate from vehicle image.
This technique is implemented using MATLAB and characters are recognized using edge detection segmentation and pre processing of image. Lekhana G.C et al.
developed an efficient real time on-line Number plate recognition system. NPR
algorithm works in different steps firstly image acquisition, using fusion of spectral analysis characters are segmented and characters are recognized. Paunwala
C.N et al. proposed a methodology which finds ROI using morphological processing and directional segmentation. The ROI is the area which includes the
number plate from which alphanumeric characters are recognized. This method
is tested on different databases which contain images. Singh M et al. developed
an efficient approach works on opening and closing of morphological operations.
Firstly localization of plate in image has been done then skew correction is done
for segmentation process of alphanumeric characters. Recognization is done using the template matching. Kranti S et al. presented a methodology for number
plate extraction named Feature based number plate localization . This methodology mainly deals on two methods edge detection and window filtering method.
Both methods are used in this methodology and give efficient results. Ganapa-
11
thy V et al. developed a methodology for Malaysian vehicles. This methodology

is mainly based on Hough transform and morphological analysis and results extraction of number plate with 95 percentage accuracy. Othman K et al. used an
approach which is texture based approach and worked on edge information for
localization and recognition. Multi layer perceptron and neural network are used
for segmentation of alphanumeric characters of license plate[32].
12
Chapter 3
PROBLEM FORMULATION
The whole work consists of three parts - testing, training and number plate
extraction. In the testing portion, the candidate object is tracked by extracting
the log gabor features. Log-Gabor filters are designed as Gaussian functions on
the log axis, which is in fact the standard method for representing the spatial frequency response of visual neurons. Their symmetry on the log axis results in a
more effective representation of the uneven frequency content of the images: redundancy in lower frequencies is reduced, and the response of the filter in the
linear frequency axis displays a tail in the higher frequencies that adapts the frequency fall-off of natural images. Furthermore, log-Gabor filters do not have a DC
component, which allows an increase in the bandwidth, and hence fewer filters are
required to cover the same spectrum.
The results obtained at the testing part is validated along with identifying the
objects in the input images to be vehicle or not, with the help of a trained system. In particular, this is usually addressed as a two-class supervised classification problem in which a set of samples are trained in search of specific feature
descriptors of the vehicle and the nonvehicle classes. The system training is done
with more images and the feature values are extracted and stored. Later, the num-
13
ber plates are read from the images, identified to have a vehicle object. Thus, the
overall system developed, will identify the objects in each input image to be a
vehicle or not and also reads the number plates of the identified objects.
The advantages are:

The response of each filter in the bank is typically represented by statistics
such as the mean and the standard deviation, both for characterization of
vehicles or other objects.
These statistics can be extracted over the whole image or in smaller sub windows. For instance, in the candidate image is divided into nine overlapping
sub-windows.
Other representations of texture based on Gabor filters are analyzed. In
this paper, an alternative representation of images for vehicle classification
using log-Gabor filters instead of Gabor filters is proposed and evaluated.
The overall work is seperated into several steps like, Image acquisition, Image
enhancement, Gray scale conversion, Log gabor filters, Vehicle verification and
Extracting number plate.
3.1
Block diagram
The block diagram for the proposed method is as shown in Figure 3.1. A detailed
explanation of each modules is mentioned in this section. However, the overall
work focuses in identifying the vehicle object in the image given. Thereafter,
the number plates are read and displayed from the images identified to have a
vehicle object in it. The number is read through an iterative process and the usual
segmentation process is avoided in this work.
14
Figure 3.1: Block Diagram of Proposed Method
3.1.1
Image acquisition
The method starts with the image collection using a camera on the vehicle dashboard, or on vehicle overhead. The captured images are stored in either .jpeg or
.png form. The images were affected with noise as a result of changing illumination, weather factors, vehicle speed etc. Hence, an image enhancement was done
prior to the major work. Here I have used the vehicle images taken from the open
GTI database. The images include both vehicle and non vehicle images, captured
from cameras fixed at front, left, right and farther positions. Fig 3.2 shows sample
15
images used in this work.
Figure 3.2: The image samples used for training the descriptor. These images
are taken from the open GTI database. Four sets of images (a), (b), (c) and (d)
are taken where in the camera position is changed. (a): Far end position, (b):
Front end position, (c): Right end position and (d): Left end position
3.1.2
Image enhancement and gray scale conversion
All the captured images will certainly consist of non-vehicular objects like
human, road, buildings etc, which needs to be filtered to detect the vehicle. To
perform the same, we enhance the image. The input images are resized so that
issues of artifacts are reduced. To perform the same, we enhance the image with
the surrounding pixels. The resultant image is converted to gray scale. Later, we
take the DFT, perform filtering operation, perform IDFT.
16
3.1.3
Log gabor filters
Log-Gabor filters are designed as Gaussian functions on the log axis, which is in
fact the standard method for representing the spatial frequency response of visual
neurons. Their symmetry on the log axis results in a more effective representation
of the uneven frequency content of the images: redundancy in lower frequencies
is reduced, and the response of the filter in the linear frequency axis displays a
tail in the higher frequencies that adapts the frequency falloff of natural images.
Furthermore, log-Gabor filters do not have a DC component, which allows an increase in the bandwidth, and hence fewer filters are required to cover the same
spectrum. The main contribution of this paper is the design of a descriptor based
on log-Gabor functions for vehicle verification instead of state-of-the-art descriptors based on Gabor functions. These functions have not been previously used for
vehicle representation purposes.
Feature extraction
Several strategies can be taken to define the feature vector from the result of
Gabor filtering. Some of them are discussed in [24]. and include the use of raw
Gabor responses, thresholded Gabor features, Gabor energy features and grating
cell operator features. These are compared [24] in a set of generic texture images based on the Fisher criterion. The former two deliver poor results, whereas
good performance is achieved by Gabor energy features and grating cell operator
features. However, the latter are not suited for vehicle classification, as they require that a system of at least three bars is present in their receptive fields [24].
Gabor energy features (which combine the response of symmetric and antisymmetric Gabor filter) are thus selected. These features result in very large vectors
(as large as the size of the image) and thus entail heavy training and classification.
17
Therefore, statistical moments are usually preferred [17]-[]19],[30]. and will also
be adopted in this paper. In particular, three moments are analyzed.
The mean , the standard deviation and the skewness of the data distribution.
m,n =
s
m,n =
m,n
1 XX
| Jm,n (x, y) |
R.C x y
1 XX
(| Jm,n (x, y) | m,n )2
R.C x y

3
1 X X (| Jm,n (x, y) | m,n )
=
R.C x y
m,n
(3.1)
(3.2)
(3.3)
The feature vector is thus composed of the mean and variance of the images resulting of applying the Gabor or log- Gabor filter bank to the input image:
Log gabor based Classifier

The maximum frequency, and the scaling between center frequencies, scaling
factor a, are also adjusted so that the same range of frequencies is covered. Remarkably, each filter of the log-Gabor bank can encompass a larger bandwidth
than the traditional Gabor function, therefore theoretically a smaller impact in the
performance is to be observed when decreasing the number of scales. the best
performance is selected for each value of N.The results in the tables confirm the
expected average increase in performance obtained with log-Gabor filters with
respect to Gabor filters.
3.1.4
Support Vector Machine classifiers
SVM is a state-of-the-art classication method introduced in 1992 by Boser,

Guyon, and Vapnik . The SVM classier is widely used in bioinformatics (and
other disciplines) due to its high accuracy, ability to deal with high-dimensional
18
data such as gene expression, and exibility in modeling diverse sources of data.
SVMs belong to the general category of kernel methods . A kernel method is an
algorithm that depends on the data only through dot-products. When this is the
case, the dot product can be replaced by a kernel function which computes a dot
product in some possibly high dimensional feature space. This has two advantages: First, the ability to generate non-linear decision boundaries using methods
designed for linear classfiers. Second, the use of kernel functions allows the user
to apply a classier to data that have no obvious fixed-dimensional vector space
representation. The prime example of such data in bioinformatics are sequence,
either DNA or protein, and protein structure.
SVM are primarily two-class classifiers that have been shown to be an attractive and more systematic approach to learning linear or non-linear decision
boundaries [29],[30]. Given a set of points, which belong to either one of the two
classes, SVM finds the hyperplane leaving the largest possible fraction of points
of the same class on the same side, while maximizing the distance of either class
from the hyperplane. This is equivalent to performing structural risk minimization
to achieve good generalization [29] [30]. Given l examples from two classes
(x1 , y1 ) (x2 , y2 ) ... (xl , yl ) , xi RN , yi {1, +1}
(3.4)
finding the optimal hyper-plane implies solving a constrained optimization problem using quadratic programming. The optimization criterion is the width of the
margin between the classes. An SVM model is a representation of the examples
as points in space, mapped so that the examples of the separate categories are divided by a clear gap that is as wide as possible. New examples are then mapped
into that same space and predicted to belong to a category based on which side of
the gap they fall on.
19
3.2
ALGORITHM
3.2.1
Vehicle detection
Testing descriptor
Step 1: The test image is selected and stored in the variable im
Step 2: Image in im is converted to gray scale and stored in b.
Step 3: These images are displayed prior to the calculation of Log gabor features.
Step 4: Obtain the feature values, using the steps mentioned below.
1. Convert image to double.
2. Take fourier transform of the image.
3. Calculate filter angle.
4. Initialize filter wavelength.
5. Calculate absolute angular distance.
6. Calculate angular filter components.
Step 5: Set the values at the center filter.
Step 6: Multiply angular spread to correct the quadrants.
Training for feature set
Step 1: Collect images of four different camera poses and organize them to four
folders, for both positive and negative samples.
Step 2: Input the camera pose required N through interactive window.
1. If N=1, only front pose is required.
20
2. If N=2, front and left pose required

3. If N=3, front, left and right pose required.
4. If N=4, front, left, right and far pose required.
Step 3:
Obtain the database location and the number of images required. Read
upto the number of images entered, from the folder given. Store each images to a
pre-allocated structure.
Step 4: GOTO Step 4 of the testing section and repeat it for the count of images
entered. Repeat each step, upto step 8 in the testing part for all images. Store the
result to feature.mat file.
Verification
The system is input with several images from the database and the objects are
identified to be vehicle or not, at this stage.
Step 1: Input the camera pose required N, through interactive window.
Step 2: Perform Step 2 to 4 in the training stage.
Step 3:
Compare the feature set values obtained, with the values stored in fea-
ture.mat file. Output the result as vehicle if value 1 is returned and as nonvehicle,
if the function returns value 0.
Step 4:
Calculate ratio of detected vehicles to the total input images given.
Output the result as the overall accuracy of the system
3.2.2
License plate extraction
Neural network training

Step 1:
The images in the database folder are read one by one and saved as a
matrix form. The training is performed with nntrainig tool in matlab.

21
Step 2: A total of 22 characters, including alphabets and numbers are taken for
system training. We have considered 50 samples for each of the characters taken.
Step 3:
These are read and stored in a matrix form, and we form a bigger
matrix by repeating a single 50x22 matrix, 50 times. This helps in extracting each
features of every single character sample, one after other.
Step 4: Repeat the steps. Epoch values are given to decide the training iteration
limit.
3.2.3
Verification
Step 1: Read the video sample

Step 2: Convert it into frames
Step 3: Convert each frame image to grey.
Step 4: Statistical moments are calculated and saved to data.
Step 5: Classification into vehicle and non vehicle species is done using svm.
Step 6:
If the detected frame has a vehicle object, then number extraction is
done.
3.2.4
Number extraction
Step 1: The region of interest is selected and the dimensions are stored.
Step 2: Cropped and re sized to exact dimensions.
Step 3: We convert this image to rectangle matrix.
Step 4:
Further, scanning is performed in the matrix, to extract the digit and
character patterns, using the pretrained patterns.

Step 5: The patterns are recognized and displayed.
Step 6: The whole process is repeated for each frame captured.
22
Chapter 4
WORK DONE AND RESULTS
4.1
Vehicle identification
The performance of the proposed method is validated with a test image. The results of the initial testing stage are given below. Figure 4.1 shows the original rgb
image taken for testing purpose. The test image is in jpeg format, with dimension
400x226 and size of 14kb.
Figure 4.1: Sample input image
23
The input image is converted to gray scale. Gray scale images only contain one
image plane, containing the gray scale intensity values. This reduces the computational complexity. Also, the reduction in the image size can improve the speed
of operation. Figure 4.2 shows the gray scale converted image.
Figure 4.2: Gray scale enhanced image

The log gabor filter is applied to this image, to detect the edges which appear in
the frequency domain as high frequencies. These detected edges can be used as
the input to a feature extraction module. Figure 4.3 shows the edges detected with
the Log Gabor filter.
Figure 4.3: Log gabor filter output

24
An interactive interface has been designed to improve the practicality of the system. The following figure shows the user interactive menu list for selecting the
camera position. With this option, the value od N is decided. The variable N is
assigned four values as follows.
1. If N=1, front pose only.
2. If N=2, front and left
3. If N=3, front, left and right
4. If N=4, front, left, right and far
Figure 4.4: Selecting camera pose

Further the log gabor energy features are calculated for both vehicle and nonvehicle samples. The following figure shows the interface for selecting the positive and negative samples. Each portions are included with four camera positions.
Figure 4.5: Input images are selected

25
The feature sets are calculated for the images selected. The following tables shows
the feature values calculated for vehicle samples with camera in front position.
Table 4.1: Calculated feature sets for vehicle samples
Object sample Camera Pose
Vehicle
Front Pose
Vehicle
Front Pose
Vehicle
Front Pose
Vehicle
Front Pose
Vehicle
Front Pose
Vehicle
Front Pose
Vehicle
Front Pose
Vehicle
Front Pose
Vehicle
Front Pose
Vehicle
Front Pose
Mean
3.1662
3.3965
3.3567
3.3544
2.96
3.0146
3.0875
3.1722
2.9702
3.095
Sigma Skewness
1.181
1.3537
1.3572
1.3859
1.9342
1.7281
1.4802
1.8172
1.2146
1.5467
1.821
2.4998
1.7598
2.2418
1.9832
2.1422
1.8826
2.0821
2.0393
1.9249
Similarly, the following tables shows the feature values calculated for nonvehicle samples with camera in front position.
Object sample Camera Pose

Non-Vehicle
Front Pose
Non-Vehicle
Front Pose
Non-Vehicle
Front Pose
Non-Vehicle
Front Pose
Non-Vehicle
Front Pose
Non-Vehicle
Front Pose
Non-Vehicle
Front Pose
Non-Vehicle
Front Pose
Non-Vehicle
Front Pose
Non-Vehicle
Front Pose
Mean
3.6052
2.9086
2.5876
3.0948
3.8125
3.7975
4.0487
4.3227
3.2481
4.5744
Sigma Skewness
0.3045
0.0269
0.7459
1.3743
0.7443
0.8422
0.5401
-0.485
1.3459
2.4179
0.4246
7.1077
0.9756
2.8729
1.3092
2.5698
0.6314
0.2046
1.8135
1.2711
Table 4.2: Calculated feature sets for non-vehicle samples

After calculating the feature set, we check for vehicle objects in the images. Here,
the images are taken from the open GTI database, to spot the object to be vehicle
26
or not. The following figure shows the result obtained for a single vehicle image,
along with the feature set values.
Figure 4.6: Result for single vehicle image

The following figure shows the result obtained for a single non-vehicle image,
along with the feature set values.
Figure 4.7: Result for single non-vehicle image
27
4.2
Number plate extraction
The neural network training is performed with nntraining tool in the Matlab2012.
The following figure shows the window for completed training.
Figure 4.8: Neural network training completed

The figure below shows the results of the number read with the iterative scanning
of input images.
Figure 4.9: Detected number KL05J8798
28
Figure 4.10: Detected number SAP729
Figure 4.11: Detected number 5RLS
Figure 4.12: No number detected
29
Chapter 5
CONCLUSION
In this work the potential of Log Gabor filters as descriptors for supervised
vehicle classification has been utilized. A Log Gabor filter based descriptor has
been employed to extract features of the input image. The tabulated feature sets
for the vehicle and non-vehicle samples are used to identify the object in the given
input set. This classification is done using a support vector machine classifier,
such that the complexity is reduced. With this system, the vehicle numbers were
identified, using the character recognition and neural network. The initial part
is used to judge the identified object to be a vehicle or not. And, vertical and
horizontal scanning is performed over the enhanced image of vehicles, to find the
number. This method is comparitively simple as it does not use the usual methods
of segmentation and spotting ROI.
30
Chapter 6
FUTURE SCOPE
The system is currently developed with a limited set of alphanumeric characters
for training the neural network. This can be enhanced with more font varieties to
enable the identification of vehicle plates of different countries. The system can
also be applied in localized vehicle plate identification systems, with a definite
set of alphanumeric character set. The system can be extended into real time systems such that vehicle numbers captured with a CCTV can be read. This system
finds wide application in automatic parking systems, toll collection units, traffic
surveillance systems etc.
31
REFERENCES
[1] J.Arrospide and L.Salgado, Log Gabor filter for image based vehicle verification, IEEE, Transcations on Image Processing,Vol22,no.6,June2013.
[2] Z. Sun, G. Bebis, and R. Miller, On-road vehicle detection: A review IEEE
Trans. Pattern Anal. Mach. Intel.., vol. 28, no. 5, pp. 694711, May 2006.
[3] N. Dalal and B. Triggs, Histograms of oriented gradients for human detection in Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit., vol.
1. Jun. 2005, pp. 886893.
[4] T. Gandhi and M. M. Trivedi, Video based surround vehicle detection, classification and logging from moving platforms: Issues and approaches in
Proc. IEEE Intell. Veh. Symp., Jun. 2007,pp.. 10671071.
[5] N. Dalal and B. Triggs, Histograms of oriented gradients for human detection in Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit. , vol.
1. Jun. 2005, pp. 886893.
[6] T. Barbu, Gabor filter based face recognition technique,Proceedings of the
Roman Academy,series a,Vol. 11,no. 3/2010,pp. 227-283.
[7] J. Zhou,D. Gao,D. Zhang, Moving vehicle detection for automatic traffic
monitoring, IEEE Trans. Vehicular Technology,Vol. 56,n0. 1,Jan. 2007.
[8] M,M,Rashid,A. Musa,M. A. Rahman,N. Farahana,A. Farhana, Automatic

parking management system and parking fee collection based on number
plate recognition,International Journal of Machine Learning and Computing,Vol. 2,no. 2,April,2012.
[9] S. Asthana,N. Sharma,R. Singh, Vehicle number plate recognition using
multiple layer back propagation neural networks,IJCTEE,Vol. 1,Issue1.
,July2011.
[10] Meenakshi,R. B. Dubey,
Vehicle license plate recognition sys-
tem,International Journal of Advanced Computer Research,Vol. 2,no.

4,Issue. 7,2012.
[11] M. M. M. Fahmy, 1994, Automatic Number-plate Recognition: Neural Network Approach, Proceedings of VNIS94 Vehicle Navigation and Information System Conference,31 Aug, 1994.
[12] J. B. Kim, C. W. Lee, K. M. Lee, T. S. Yun, and H. J. Kim, Wavelet-based
vehicle tracking for automatic traffic surveillance, in Proceedings of IEEE
International Conference on Electrical and Electronic Technology, Vol. 1,
2001, pp. 313-316.
[13] G. L. Foresti, V. Murino, and C. Regazzoni, Vehicle recognition and tracking from road image sequences IEEE Transactions on Vehicular Technology, Vol. 48, 1999, pp. 301-318.
[14] Y. K. Jung, K. W. Lee, and Y. S. Ho, Content-based event retrieval using semantic scene interpretation for automated traffic surveillance IEEE
Transactions on Intelligent Transportation Systems, Vol. 2, 2001, pp. 151163.
[15] Z. He, J. Liu, and P. Li, New method of background update for videobased vehicle detection, in Proceedings of IEEE Conference on Intelligent
Transportation Systems, 2004, pp. 580-584.
[16] A. Kuehnle, Symmetry-based recognition for vehicle rears Pattern Recognition Letters, vol. 12, pp. 249258, 1991.
[17] M. Bertozzi, A. Broggi, A. Fascioli, and S. Nichele, Stereo vision-based
vehicle detection, IEEE Intelligent vehicle symposium, pp. 3944, 2000.
[18] D. Guo, T. Fraichard, M. Xie, and C. Laugier, Color modeling by spherical influence field in sensing driving environment, IEEE Intelligent Vehicle
Symposium, pp. 249254, 2000.
[19] H.Mori and N. Charkai, Shadow and rhythm as sign patterns of obstacle
detection,International Symposium on industrial exlectronics, pp. 271 277,
1993.
[20] E. Dickmanns, et. al. The seeing passenger car vamors-p, International
Symposium on Intelligent Vehicles94, pp. 2426, 1994.
[21] C. Tzomakas and W. Seelen, Vehicle detection in traffic scenes using
shadows,Tech. Rep. 98-06, Institut fur neuroinformatik, Ruht-universitat,
Bochum, Germany, 1998.
[22] 2S. E. Grigorescu, N. Petkov, and P. Kruizinga, Comparison of texture features based on Gabor filters, IEEE Trans. Image Process., vol. 11, no. 10,
pp. 11601167, Oct. 2002.
[23] D. Koller, N. Heinze and H. Nagel, Algorithm characterization of vehicle
trajectories from image sequences by motion verbs, IEEE Conf. on Computer Vision and Pattern Recognition, pp. 9095, 1991.
[24] M. Betke, E. Haritaglu and L. Davis, Multiple vehicle detection and tracking in hard real time, IEEE Intelligent Vehicles Symposium, pp. 351356,
1996.
[25] Z. Sun, G. Bebis, and R. Miller, Improving the performance of on-road
vehicle detection by combining gabor and wavelet features, The IEEE Fifth
International Conference on Intelligent Transportation Systems, September,
2002, Singapore
[26] H. Cheng, N. Zheng, and C. Sun, Boosted Gabor features applied to vehicle
detection, in Proc. 18th Int. Conf. Pattern Recognit., 2006, pp. 662666.
[27] P.Anishiya, Prof S.M, Mary Joan, Number Plate Recognition for Indian
Cars Using Morphological Dilation and Erosion with the Aid of OCRs,
IACSIT, Vol4, 2011, pp 115-119.
[28] A.S. Johnson, B.M. Bird, 1990, Number-plateMatching for Automatic Vehicle Identification, IEEE Colloquium on Electronic Image and Image Processing in Security and Forensic, Aprl, 1990.
[29] M.M.M. Fahmy, 1994, Automatic Number-plate Recognition : Neural Network Approach, Proceedings of VNIS94 Vehicle Navigation and Information
System Conference, 31 Aug-2 Sept, 1994.
[30] H.J. Choi, 1987, A Study on the Extraction and Recognition of a Car Number Plate by Image Processing, Journal of the Korea Institute of Telematics
and Electronics, Vo1.24, pp. 309-3 15,1987.
[31] Serkan Ozbay, and Ergun Ercelebi, Automatic Vehicle Identification by
Plate Recognition World Academy of Science, Engineering and Technology
International Journal of Electrical, Computer, Electronics and Communication Engineering Vol:1, No:9, 2007.
[32] Manisha Rathore and Saroj Kumari Tracking number plate using MatLab International Journal in Foundations of Computer Science Technology (IJFCST), Vol.4, No.3, May 2014.

Final

Uploaded by

Document Information

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

Final

Uploaded by

Copyright:

Available Formats

Acknowledgement

I am thankful to our Principal, Prof.Shine P. James, for his valuable support

I am highly indebted to our MTech coordinator Ms.Flower Abraham, Asst.

I would like to express my gratitude to my project guide Ms.Ajishya Liza

Organization of the project . . . . . . . . . . . . . . . . . . . . .

Background subtraction method . . . . . . . . . . . . . .

Using object features . . . . . . . . . . . . . . . . . . . .

Object based approach . . . . . . . . . . . . . . . . . . .

Color transform model . . . . . . . . . . . . . . . . . . .

Detection using shadow . . . . . . . . . . . . . . . . . .

License plate extraction . . . . . . . . . . . . . . . . . . . . . . .

Image enhancement and gray scale conversion . . . . . .

Log gabor filters . . . . . . . . . . . . . . . . . . . . . .

Support Vector Machine classifiers . . . . . . . . . . . . .

License plate extraction . . . . . . . . . . . . . . . . . .

WORK DONE AND RESULTS

Number plate extraction . . . . . . . . . . . . . . . . . . . . . .

Block Diagram of Proposed Method . . . . . . . . . . . . . . . .

Sample input images . . . . . . . . . . . . . . . . . . . . . . . .

Sample input image

Gray scale enhanced image . . . . . . . . . . . . . . . . . . . . .

Log gabor filter output . . . . . . . . . . . . . . . . . . . . . . .

Selecting camera pose . . . . . . . . . . . . . . . . . . . . . . . .

Input images are selected . . . . . . . . . . . . . . . . . . . . . .

Result for single vehicle image . . . . . . . . . . . . . . . . . . .

Result for single non-vehicle image . . . . . . . . . . . . . . . .

Neural network training completed . . . . . . . . . . . . . . . . .

Detected number KL05J8798 . . . . . . . . . . . . . . . . . . . .

4.10 Detected number SAP729 . . . . . . . . . . . . . . . . . . . . . .

4.11 Detected number 5RLS . . . . . . . . . . . . . . . . . . . . . . .

4.12 No number detected . . . . . . . . . . . . . . . . . . . . . . . . .

Calculated feature sets for vehicle samples . . . . . . . . . . . . .

Calculated feature sets for non-vehicle samples . . . . . . . . . .

Advanced Driver Assistance Systems

Automatic license plate recognition

bidirectional associative memories

Closed Circuit Television

Histogram Oriented Graph

intelligent transportation system

License Plate Recognition

Pricipal Component Analysis

Support Vector Machine

Support Vector Machine

College of Engineering Poonjar

Image based vehicle verification

Vision-based object detection from a moving platform becomes particularly

College of Engineering Poonjar

Image based vehicle verification

Techniques for vehicle identification implemented with image processing have

Department of Electronics & Communication Engineering

Image based vehicle verification

College of Engineering Poonjar

Department of Electronics & Communication Engineering

College of Engineering Poonjar

Image based vehicle verification

Later, optical character recognition technique has been employed to extract

Organization of the project

Department of Electronics & Communication Engineering

College of Engineering Poonjar

Image based vehicle verification

College of Engineering Poonjar

Image based vehicle verification

Using object features

College of Engineering Poonjar