You are on page 1of 9

License Plate Recognition:

A Brief Tutorial

© wikimedia commons

Christos-Nikolaos E. Anagnostopoulos

1. Introduction A computer vision and character recognition algorithm

W
ith the rapid development of public trans- for license plate recognition can be used as a core for
portation system, automatic identification intelligent infrastructure like electronic payment sys-
of vehicles become more and more practi- tems (toll payment, parking fee payment), freeway and
cal in many applications during the past two arterial management systems for traffic surveillance.
decades. Nowadays, Intelligent Transportation Systems Moreover, as increased security awareness has made
(ITS) are having a wide impact in people’s life as their the need for vehicle based authentication technolo-
scope is to improve transportation safety and mobil- gies extremely significant the proposed system may be
ity and to enhance productivity through the use of employed as access control system for monitoring of
advanced technologies. ITS are divided into intelligent unauthorized vehicles entering private areas.
infrastructure systems and intelligent vehicle systems. All vehicles, world-wide, should have a license
number, which is their principal identifier despite the
Digital Object Identifier 10.1109/MITS.2013.2292652 fact that it can be deliberately altered in fraud situa-
Date of publication: 21 January 2014 tions or replaced (e.g., with a stolen plate). Therefore,

IEEE Intelligent transportation systems magazine • 59 • Spring 2014

1939-1390/14/$31.00©2014IEEE
ITS rely heavily on robust License Plate Recognition (LPR) ground. Alternatively, a triggering signal occurs when
systems. As a simple definition, a LPR system is an inte- illumination or color change above a predefined thresh-
grated hardware and software module that identifies the old appears in a selected part of the image (virtual loop).
content of the plate (i.e. Vehicle Identification Number— ■■ F ree flow: The pipeline does not receive signal from
VIN) and generates the appropriate sequence of ASCII software or hardware device. In this case the cap-
characters. The focus of this article is to briefly describe turing device (usually video-cameras or lately smart
such core modules as well as to discuss other related topics cameras) takes images continuously and the imaging
and future trends. software searches and tracks license plates according
to specific rules.
2. Installation/Hardware
Setting up an LPR system is a crucial procedure that plays 2.2 Auxiliary Units
the most important role in the successful operation of the In addition, in order to overcome the problem of varying
system. The proper planning should take into consideration illumination, Infra-Red (IR) auxiliary units are used. The
several installation issues, besides the core modules for LPR. idea of using them emerged from the nature of the license
Important factors that should be addressed relates to appro- plate surface, which usually are made from retro-reflective
priate timing of image acquisition (i.e. triggering devices materials, which cause them to shine when a bright light is
and camera calibration) and to varying illumination condi- directed towards them. This attribute makes license plates
tions encountered during a twenty-four hour period. The perfect candidates for cameras that are sensitive to infra-
former is addressed by selecting suitable triggering devices red illumination. IR units are employed in scientific papers
and calibrating carefully the capturing device (either cam- and in the majority of the commercial systems, since they
eras or video-cameras), while the latter is tackled by using contribute significantly to the plate localization problem
auxiliary illumination units that do not interfere with the and they increase the sharpness and clearness of the plate
human visual system or distract the driver’s attention. image to be forwarded to character recognition modules.
Such cameras, employing a narrow band-pass filter and
2.1 Triggering Devices coupled with infrared illuminators make ideal license plate
Triggering devices deals with image acquisition at the capture devices. Illumination in the area of 950 nm is ideal
proper time and belong in one of the follow three categories: for capturing retro-reflective license plates while avoiding
■■ Sensor/loop trigger: The plate recognition pipeline distraction to drivers, since it is unobtrusive, invisible and
includes a hardware sensor (magnetic loop detector, eye-safe. The illumination unit is an array of Light Emit-
laser scanner, infrared sensor), which is usually located ting Diodes (LEDs), which were incorporated into a secure
at the roadside of the experimental site (access point, housing in a manner that allows them to operate in any out-
lanes). Whenever a vehicle has been detected by the door condition. The camera incorporates a filter eliminat-
sensor, the image acquisition process initiates (either ing any light that falls outside the range 920 nm–980 nm.
one or multiple images are stored).
■■ Software trigger: In this case, there is no physical sen- 2.3 Camera Calibration
sor in the pipeline. The image acquisition process begins Problems can arise from aspects related to various settings
using sophisticated change/movement detection algo- of the camera and specifically its calibration and position-
rithms. Usually, change detection methods compare each ing. Important parameters that affect the overall perfor-
image frame against a predetermined reference back- mance are the angle of vision in horizontal (pan) and ver-
tical direction (tilt) (see Figure 1). Large values in these
parameters deteriorated the performance of the system,
since they alter significantly the appearance of plate and
the characters to be identified by OCR. In such a case, the
imaging system probably would need an image restoration
technique, prior to LPR.
Another issue that should be addressed carefully is the
shutter speed of the camera (if a still camera is used). Low
V shutter speed is a particular problem if the targets are
H moving at an excessive speed, causing a blurry appearance
of the vehicle and the license plate. To avoid blurring it is
ideal to have the shutter speed of the selected camera set
less than 0.001 seconds for free flow applications. In slow-
moving traffic or in access/control/billing points where
Fig 1 Pan (H) and tilt (V). the vehicles stop when the shutter speed may be lower.

IEEE Intelligent transportation systems magazine • 60 • Spring 2014


4
1 1
2 8
2
3 10
4 3 {
5
6 9 4 7
3
7 5 D = 1,8”

8 6

Fig 2 Artificially created images for measuring discrimination ability. Fig 3 CCD representation of the example.

Moreover, angle of vision (pan and tilt) also influences CCD width = 1.8 ) cos { (inches)

the value of the shutter speed. Lower values of pan and tilt CCD width = 36, 6 mm (2)
decrease the need of high shutter speed as the difference
CCD height = 1.8 ) sin { (inches)
between frames in this case is smoother. Usually the shut- 
ter speed is selected by a trial an error procedure accord- CCD width = 27, 5 mm. (3)
ing the application needs.
Another question that should be solved is the calcu- In order to calculate the object distance the equation (4)
lation of the effective distance in relation to the camera is used:
specifications. The effective distance is a function that
f
has several parameters related to the camera, the dis- = D ,(4)
l L
crimination ability of the LPR algorithm and the physical
dimensions of the smallest plate to be identified. There- where, f, l, D and L are depicted in Figure 4.
fore, suppose that: Another important measure is the physical dimensions
■■ A camera used in a LPR testbed has 3.2 million effective of one pixel in the CCD. This can be found from the effec-
pixels (e.g. 2048 # 1536 pixel images), CCD at 1.8 inch tive pixels and the actual dimensions of the CCD. Thus, the
with a typical width/height ratio equal to 4/3, and lens square pixel dimensions are given by the equations:
3x adjusted to f = 24 mm (equivalent to 114 mm camera
format). CCD_width
pixel_width =
■■ D iscrimination ability of the algorithm is 20 # 68 pixels horizontal effective pixels

and (note that 20 pixels is typically the lowest height of
= 36.6 mm = 17.9nm (5)
identifiable characters in most OCR engines) 2048
■■ T he smallest plate to be identified n is a rectangular CCD_height
pixel_height =
object of 33 cm width and 10 cm height. vertical effective pixels

We have to note that, the discrimination ability indi- = 27.5 mm = 17.9nm. (6)
cates the smallest plate that can be identified by the system 1536
correctly by the LPR system (including the OCR module as
well). Discrimination ability is usually measured by trial The dimension of the license plate in the CCD (Figure 5)
and error or by artificially created images that contain is calculated as:
many copies of a cropped vehicle image with its plate in
different resolutions and size (see Figure 2). license plate in the CCD : width = 68 pixels x 17.9 nm

The effective distance for successful license plate recog-  = 1.21 mm (7)
nition is calculated as follows. Suppose that a typical CCD
license plate in the CCD : height = 20 pixels x 17.9 nm
has a width/height ratio equal to 4/3 as shown in Figure 3:
Therefore, the angle { is:  = 0.36 mm. (8)

3
{ = arctan ( ) & { = 36, 8 degrees .(1)
4 Hence, the object distance is:

Since the CCD is at 1.8 inches its dimension in cm is given L# f 0.33 (m) # 24 (mm)
D= = . 6.5 meters .(9)
by (2) and (3): l 1, 21 (mm)

IEEE Intelligent transportation systems magazine • 61 • Spring 2014


Lens

Real Image CCD


Object
Object
Dimension L
Dimension l
1,21 mm
0,36 mm
License Plate

Object Distance Focal Length


D f

Fig 4 Representation of parametes f, l, D and L. Fig 5 Dimensions of a plate in the CCD.

The above distance D, is actually the maximum dis- to motorway exit). To this end, practical applications of VrI
tance between the target (vehicle) and the camera using are numerous focusing on analysis of travel behavior for
the specific camera and lens. Of course the above-calcu- transport planning purposes. Specifically, the derivation
lated distance could be increased if more expensive lenses of section travel times (time taken by a vehicle to go from
and special equipment are employed. A useful link for one point to another) is useful to transportation engineers
selecting cameras and lenses can be found at http://www. for traffic operations, planning and control. Accurate travel
cctv-information.co.uk/i/ An_Introduction_to_ANPR. times and densities can be instrumental in travel reliabil-
ity, feedback control, vehicle routing, traffic assignment and
3. Applications traveller information systems. If vehicles are tracked along
consecutive points, then partial origin/destination demands
3.1 Security Applications can even be measured instead of estimated [1],[2]. In addi-
In the past decades, the issue of security has become more tion, the problem of analog toll billing could be addressed
significant and the need for effective security systems has with a robust VrI system, as shown in Figure 6.
intensified. Many areas were marked as restricted such as Moreover, Vehicle Manufacturer/Model Recognition
borders, government buildings, embassies, power plants (VMMR) is a sophisticated computer vision application
and military camps, since illegal access can have serious based on LPR that appeared lately in the literature [3],[4].
consequences for homeland security. LPR check points For VMMR, the spatial position of a license plate in a cap-
provide control management, granting access only to vehi- tured image or video frame plays an important role in
cles authorized to enter limited zone areas. the segmentation of a distinctive reference area from the
vehicle (see Figure 7). From this area, which is usually
3.2 Ticket Issuing, Billing
Ticket issuing in parking lots and tolls are actually among
the first and most common applications of LPR. They are
also the easiest LPR implementation, since vehicles either Vehicle B
Analog Billing:
stop before the check point (parking lots), or pass with low Reidentified
Vehicle A
speed from a specified lane (toll). Applications include the Vehicle B
detection of various traffic violations (e.g. speeding, bus
lane occupancy) and the identification of stolen or unin-
sured vehicles. For the latter task, an updated record of
such vehicles by the authorities is crucial.
Vehicle A Vehicle A
Reidentified
3.3 Other Applications
License plate recognition, as a means of vehicle identifica-
tion, may be further exploited in various ways, such as traffic
management, analog billing in highways and Vehicle Manu-
facturer/Model Recognition (VMMR). Traffic management Vehicle B
and analog billing is based on Vehicle re-Identification
(VrI). VrI is the process of matching vehicles from one point
on the roadway to another (e.g., from motorway entrance Fig 6 Analog toll billing through license plate recognition.

IEEE Intelligent transportation systems magazine • 62 • Spring 2014


proportional to the license plate, important local features As far as extraction of the plate region is concerned,
could be retrieved that as the logo area or other vehicle techniques based upon combinations of edge statistics, tex-
“biometrics” (i.e. headlight shape, dimensions, distance ture and mathematical morphology are the most popular.
from symmetry axis etc.). Of course, robust classification They are based on the property that license plates can be
schemas and appropriate training are critical issues to be
addressed. Beside vehicle classification, vehicle matching
and/or verification are also possible, assuming that a tem- Table 1. Categorization of methods for license plate detection
plate image of a vehicle has been archived into a database in images.
in advance.
Approach Description of Major Processing
4. Image Processing/Analysis in LPR Binary image processing
Image analysis is the extraction of meaningful and useful
information from digital images by means of Digital Image Mathematical morphology Tophat operations
Processing (DIP) techniques. DIP is a form of signal pro- Edge features Vertical edges
cessing for which the input is a digital image or a video Vertical Sobel operators
sequence. The output of DIP may be either an image or a Binary algorithms Connected component analysis
set of features useful for solving application-based prob- Shape analysis, geometric attributes/spatial
lems. In a License Plate Recognition (LPR) pipeline, the moments
input is a color or grayscale image and the output is a string Gray level processing
of characters. LPR typically includes three steps, license
Global image analysis Filtering and enhancement
plate detection, character segmentation and optical char-
acter recognition. In all three steps, there is plenty space to Partial image analysis Horizontal image scanning (N row distance)
implement image analysis and artificial intelligence sche- Statistical measurements Statistical block processing
mas. This section provides a brief reference of relevant lit- Statistical Haar-like features selected by
AdaBoost training schemes
erature, according to their major methodology in the three
typical steps of license plate recognition. Details on perfor- Region segmentation Dynamic/adaptive thresholding
mance, execution time, image resolution and other relative (adaptive thresholding) Mean shift segmentation
issues in the literature can be found in [5]. Hierarchical representations Quadtree decomposition and vector
quantization
4.1 License Plate Detection P robabilistic object tracking Condensation algorithm and differential
(Image Processing, Artificial Intelligence) in video evolution
License plate detection aims at the spatial identification of Transforms Gabor filters in block-based processing
the license plate within the input image. To this end, a fea- Hough transform, windowed-HT
ture extraction process initiates to reduce the amounts of Radon transform
Wavelet transform
data in the image, which is usually a coarse-to-fine strategy.
Generalized symmetry transform
Color processing

Color model transformation Color segmentation in RGB/HSI model


Fuzzy sets Color and spatial positioning fuzzy sets in
Mask Wmask = n : Wplate RGB/HSI model
Fuzzy c-means
Hmask = k : Hplate
Texture features Texture-based measurements
[Example: n = 4,
License Plate k = 2] Histogram manipulation Histogram intersection
Classifiers
W mask Statistical classifiers Nearest neighbor search
Mahalanobis distance
Support vector machines
H mask
Computational intelligence Genetic algorithms, genetic programming,
discrete time cellular neural network, pulse
License Plate H plate coupled neural network, time delay neural
network, multilayered feedforward neural
W plate
network, convolutional neural network,
fuzzy neural network
Fig 7 Segmentation of a distinctive frontal view of a vehicle.

IEEE Intelligent transportation systems magazine • 63 • Spring 2014


viewed as irregularities in the texture (or brightness) of
the image and therefore abrupt changes in the local char- Table 2. Categorization of shape analysis for binary objects.
acteristics of the image, manifest probably its presence.
Shape Analysis
Since this method does not depend on the edge of license-
plate boundary, it can be applied to an image with unclear 1. Topological 3. Moments 4. Shape orientation
license-plate boundary and can be implemented simply and attributes descriptors
fast. As far as extraction of the plate region is concerned, 1.1 Euler number 3.1 Surface 4.1 Bounding box:
a fast categorization of methods that were reported in the 1.2 Convex hull 3.2 Center of gravity     • height, width,
literature is shown in Table 1, along with the description of area, ratio
the main processing method. 1.3 Perimeter 3.3 Inertia 4.2 Image-oriented
An extremely crucial step in License Plate detection bounding box:
(as well as in character segmentation in the following 1.4 Area 3.4 Oriantation     • height, width,
stage) is Connected Component Analysis (CCA). CCA is area, ratio
a vital technique in binary image processing that scans 2. Geometrical 3.5 Major ellipse axis 4.3 Radius
an image and labels its pixels into components based on attributes measurements:
pixel connectivity, (either 4-connected or usually 8-con- 2.1 Compactness 3.6 Minor ellipse axis     • min, max,
nected). Once all groups of pixels have been determined, angle, ratio
each pixel is labelled with a value according to the com-
ponent it was assigned to. Extracting and labelling of
various disjoint and connected components in an image is and simplest one. Projections are used for character seg-
central to many automated image analysis applications as mentation in many text recognition systems. Obtaining a
many helpful measurements and features in binary objects binary image, the idea is to add up image columns or rows
may be extracted. Such features include area, orientation and obtain a vector (or projection), whose minimum values
and aspect ratio, just to name a few of them that are very allow us to segment characters (see Figure 9). It should be
frequently integrated to image processing algorithms for emphasized that many character segmentation modules
license plate detection (see Figure 8). Then, using simple incorporate more than one methods (e.g. adaptive thresh-
filtering techniques, binary objects with measurements olding, CCA and object filtering followed by the projection
that exceed the desired limits can be eliminated in the fol- method etc.). Usually, there is a preprocessing phase that
lowing algorithmic steps. Full shape attributes for binary removes the noise, enhances the image, and moreover, it
image processing are listed in Table 2. detects, centers and normalizes the license plate image.
The output of the preprocessing module is still an image
4.2 Character Segmentation (Image Processing) representing the corrected/enhanced license plate.
The license plate candidates determined in the plate loca- Character segmentation is needed to perform charac-
tion stage are now examined in the character’s segmenta- ter recognition, which fully relies on isolated characters.
tion phase. A large diversity of techniques to segment each Incorrectly segmented characters are not likely to be suc-
character after plate localization in the image has been cessfully recognized. In fact, most of the recognition errors
developed as shown in Table 3, where they are grouped in the LPR systems are not due to missing recognition
according their methodology. Reviewing the literature, power but to segmentation errors.
it was evident that the method that exploits vertical and
horizontal projections of the pixels is the most common 4.3 Optical Character Recognition
For the recognition of segmented characters, numerous
algorithms that appear in optical character-recognition
applications, utilizing statistical classifiers, computational
Object 1 intelligence architectures and common pattern matching
Measurement Value techniques are employed as shown in Table 4. As already
Area 268 Pixels discussed at the section related to the calculation of the
Aspect Ratio 3,5 effective distance, character height should be at least 20
pixels if character recognition is required. Note however,
Orientation -7 Degrees
Object 1 that sometimes even this resolution is not enough as dirt,
Euler Number -8
physical damage, and unpredictable shadows degrade the
Compactness 0.54
recognition performance.
.... .... Practically, OCR should successfully handle any ambi-
guity that may arise due incorrectly segmented characters
Fig 8 Shape analysis in binary object 1 (license plate). from the previous character segmentation step. Very good

IEEE Intelligent transportation systems magazine • 64 • Spring 2014


Table 3. Categorization of methods for character segmentation.
Detected LP
Approach Description of Major Processing

Binary image processing


Corrected LP
Projections Vertical and horizontal projections of the
pixels in the binary image
Vertical Projections
Binary algorithms Connected component analysis
Shape analysis and geometric attributes/
spatial moments
Measurements: height, width, area, Segmentation
orientation, centre of gravity

Mathematical morphology Thinning, thickening, pruning techniques


Fig 9 Shape analysis in binary object (license plate).
Contours Bug following
Shape driven active contour model
Table 4. Categorization of methods for plate character
Gray level processing recognition.

Histogram manipulation Intensity-gradient scattergrams Approach Description of Major Processing


Valley search in the intensity distribution
Cumulative distribution functions, Classifiers
background/foreground entropy
Statistical classifiers Hidden Markov models
R egion segmentation (local/ Block processing, threshold in Support vector machines
adaptive thresholding) each block Likelihood models in tree hierarchy
Transforms Hough transform Computational intelligence Multi layered feedforward neural networks
Classifiers Adaptive ressonance theory neural networks
Self organized neural networks
Statistical classifiers Markov random fields Probabilistic neural networks
Hidden Markov chains Learning vector quantization neural network

Pattern matching
results have been reported using neural networks and sta-
Template matching Normalized cross correlation
tistical classifiers. Given also that OCR engines are quite Hausdorff distance
mature and that they are continuously improved over time, Root mean squared error (RMSE) for all the
the developers focus their attention to OCR improvement template shifts
in sets of ambiguous characters (1/I, 0/O, 0/D, 2/Z, 8/B, Image partition: zoning, projections, contour
distance, segment count
and 5/S) rather than redesigning or retraining character
recognition modules.

A = (P # I ) %, (10)
5. Performance Calculation
Due to the wide range of applications of LPR systems there
is no actually a common consensus on how developers or where P and I are plate detection and character recogni-
end-users can assess safely the overall performance. Dif- tion percentage.
ferent approaches have been reported according to the spe- On the other hand, the second method assesses overall
cific nature of each test-bed and the special conditions that performance in a more pessimistic (strict) formula. Total
are needed. This means that the performance is strongly character recognition rate is now related exponentially to
affected by the characteristics of the physical installation, the number of characters expected n. Therefore, equation
the operation prerequisites (in-door or out-door), the vehicle (10) is converted to equation (11) below.
flow etc. Basically, three methods are usually implemented.
The first method to measure the overall success rate A = (P # I n ) %.(11)
is by calculating separately the percentage of: (i) license
plates correctly detected and (ii) characters correctly rec- The difference between the above methods is obvious.
ognized. Both percentages should be verified by an expert Suppose that we have a sufficient test set that includes 3000
supervising a sufficient test sample. This total accuracy A images of single vehicle, and each plate contains 7 charac-
is figured as shown in Equation (10). ters. If the expert verifies 96.0% correct LP detection (2880

IEEE Intelligent transportation systems magazine • 65 • Spring 2014


over 3000) and 98.0% overall character segmentation in a the tag to continuously power the tag and its RF communi-
follow-up analysis (19757 over 20160), then we have 94.1% cation circuitry, whereas Passive RFID relies on RF energy
success using equation (1). However, assuming an inde- transferred from the reader to the tag to power the tag.
pendent character recognition rate of 98.0% for each one There are several issues to be taken into consideration.
of the 7 characters in the plate (as equation 2 indicates) the For instance, passive tagged vehicles can only be read
overall performance falls to 83.3%. individually from a reader no more than 12 meters (40
The third (and most accurate) method dictates that feet) mounted over the lane. In addition, another limita-
we have to consider only the percentage of the correct tion occurs from the line-of-sight requirement and the fact
interpretation of entire plate content. Thus, equation 11 that the reader may process one vehicle each time, making
becomes as follows: such an approach inadequate in heavy traffic situations.
On the other hand, active tags in license plates (i) work
A = (P # C ) %.(12) in longer range (up to 100 meters) without the necessity
for line-of-sight, (ii) transfer bigger data packages and (iii)
In equation 12, C is the percentage of cases with suc- operate in any speed and direction, as they process data
cessful recognition of all characters in the plate. Consider- faster. However, they are battery-operated and the instal-
ing the example of the previous paragraph, if the 20160- lation cost is multiple.
19757 = 403 erroneously classified characters correspond
to the first character of different plates, then equal num- 7. Discussion
ber of license plates is not identified correctly. As a result,
the number of plates with correct interpretation of entire 7.1 Operation Challenges
content is 2880-403 = 2477, and therefore C equals to The challenges associated with LPR can be attributed to
(2447/2880)%-85.0%. Then, the overall performance fol- the following factors:
lowing equation (3) falls to 81.6%. ■■ Environmental/illumination conditions: The major
In summary, we could say that the most important indi- problems revolve around the varying light levels
cation for the evaluation of a license plate recognition sys- encountered during a twenty-four hour period and the
tem is the definition of the conditions under which the per- effect those lighting changes have on the image cap-
formance is estimated. In addition, the proper indication tured for license plate location and processing. In any
of the failure cases and the various limitations during the outdoor environment, illumination not only changes
process are essential for determining the system accuracy. slowly as daytime progresses, but may change rapidly
The article written by Liam Keilthy [6] may provide a good due to changing weather conditions or passing objects
starting point for the development of a generic approach for (i.e., clouds).
operators to ANPR system performance assessment. ■■ D ifferent dimensions/type: This is an important restric-
tion when a LPR system is expected to handle a vari-
6. Radio Frequency Identification (RFID) ety of plate formats according to the standardization
RFID is used to improve production logistics, automate regulations in each country. Color, size, character fonts,
access control and parking, secure border roadways, alphanumeric sequence, official stamps, province and
automate toll collection, manage traffic flow. RFID tag is country codes are some of the varying parameters that
a memory chip, which could be implanted in the license may confuse a LPR system.
plate during the vehicle registration process by the appro- ■■ Occlusion: Even in a perfectly fine tuned LPR system
priate transportation authority, becoming a standard iden- with correct settings, plates may be partially occluded
tifier for all vehicles. The data stored on this RFID tag can by other objects. For instance, in an image with a group
be retrieved anytime via fixed and mobile reading devices of vehicles, some vehicles may be partially occluded
by radio communication at any point needed (sometimes by others.
even in heavy traffic). ■■ Plate condition: The appearance of plates is directly
In conjunction with camera-based systems RFID could affected by unpredictable reasons such as dirt, shadows
drastically improve LPR performance in cases where the from bumpers, sun reflections, improper placement and
latter fails. Instead of having image/video processing and generally unacceptable physical appearance (e.g. dis-
OCR, RFID tags could broadcast vehicle identity to nearby colored or distorted characters etc).
receivers under all out-door conditions.
Active RFID and Passive RFID are fundamentally differ- 7.2 Super-Resolution
ent technologies that are often evaluated together. While Enhancing license plate text in real traffic videos is a
both use radio frequency energy to communicate between a challenging problem for LPR which is not sufficiently
tag and a reader, the method of powering the tags is different. addressed in the literature and still has plenty of room for
Active RFID uses an internal power source (battery) within research. In special occasions, (e.g. terrorism acts), human

IEEE Intelligent transportation systems magazine • 66 • Spring 2014


operators need the support of unconventional image/video load Matlab Code and sample images. Sample images may
processing techniques to decipher the content of a license be also downloaded from the Medialab License Plate Recog-
plate. This can be achieved with super-resolution (SR). nition database at http://www.medialab.ntua.gr/research/
Super-resolution is a method that works effectively when LPRdatabase. html. Medialab License Plate Recognition
several low resolution images contain slightly different database contains a large image and video dataset that
views of the same object. Thus, global information about has been collected and grouped according several criteria
the object is gathered. An SR image of the license plate is (type and color of plates, illumination conditions, various
obtained by fusing the information derived from multiple, angles of vision and indoor or outdoor images with or with-
subpixel shifted, and noisy low-resolution plate views. The out infra-red auxiliary units. We anticipate that research-
best case is when an object moves smoothly in the video, ers engaged in LPR or in related projects will report their
such as most cases in LPR. Motion detection and tracking results on this publicly available set or alternatively will
(i.e. motion compensation for finding corresponding areas contribute to the enrichment of this test database.
in subsequent frame) is used to create an SR license plate
view. Excellent works on this topic can be found in [7],[8]. About the Author
Christos-Nikolaos E. Anagnostopou-
7.3 LPR Goes Mobile los was born in Athens, Greece in 1975.
From 2008 developers moved away from static installa- He received his Mechanical Engineer-
tions, providing also mobile LPR (m-LPR). Usually m-LPR ing Diploma from the National Tech-
indicates that the platform is designed to operate in situa- nical University of Athens (NTUA)
tions where mobility is required or checkpoints occasion- in 1998, and the Ph.D. degree from
ally need to be moved. the Electrical and Computer Engi-
Mobile LPR, which includes also LPR units installed inside neering Dpt., NTUA in 2002. From 2008, he serves the
(on-board) vehicles, is now possible since new technology University of the Aegean as Assistant Professor in the
offers us cheaper and smaller processing units, cameras and Cultural Technology and Communication Department.
smart programmable cameras. All this equipment may be He is a member of the Greek chamber of Engineers and
portable or even be installed in vehicles, allowing real time member of IEEE. His research interests include image
LPR. Various mobile architectures are proposed, in which processing, computer vision, neural networks and artifi-
algorithms are able to compensate for certain variables that cial intelligence. He has published more than 120 papers
can affect the LPR ability to produce an accurate result. It in journals and conferences, in the above subjects as well
is evident that mobility affects the overall performance and as other related fields in informatics. He also serves as
there are noteworthy challenges related with m-LPR. The associate editor for the IEEE Intelligent Transportation
major challenge is the need to operate in various surround- Systems Magazine.
ings. So, all problems affecting LPR systems apply to m-LPR
with the extra issue of the unpredictable environment. If we 8. References
[1] C. C. Sun, G. S. Arr, R. P. Ramachandran, and S. G. Ritchie, “Vehicle
also consider the case of on-board LPR, where camera views
Reidentification using multidetector fusion,” IEEE Trans. Intell.
change as the vehicle moves, the LPR algorithms should ful- Transport. Syst., vol. 5, no. 3, pp. 155–164, Sept. 2004.
fill stricter criteria and parameterizations. In these cases, [2] C. N. Anagnostopoulos, T. Alexandropoulos, V. Loumos, and E. Kaya-
fas, “Intelligent traffic management through MPEG-7 vehicle flow
action cameras should be employed that are equipped with surveillance,” in Proc. IEEE John Vincent Atanasoff Int. Symp. Modern
digital image stabilization and lens distortion correction Computing, 2006, pp. 202–207.
[3] A. Psyllos, C. N. Anagnostopoulos, and E. Kayafas, “Vehicle model
algorithms ensuring that footage recorded will look stable. recognition from frontal view image measurements,” Comput. Stand.
The digital image stabilization technology corrects camera Interfaces, vol. 33, no. 2, pp. 142–151, 2011.
[4] H. Yang, L. Zhai, L. Li, Z. Liu, Y. Luo, Y. Wang, H. Lai, and M. Guan, “An
shakiness for smoother recording, while lens distortion cor-
efficient vehicle model recognition method,” J. Softw., vol. 8, no. 8, pp.
rection offers the option to automatically remove the fish-eye 1952–1959, 2013.
look that is common with action cameras. [5] C. N. Anagnostopoulos, I. Anagnostopoulos, I. Psoroulas, V. Loumos,
and E. Kayafas, “License plate recognition from still images and video
sequences: A survey,” IEEE Trans. Intell. Transport. Syst., vol. 9, no. 3,
7.4 Conclusions pp. 377–391, Sept. 2008.
[6] L. Keilthy, “ANPR System performance,” in Proc. Parking Trend Int.,
In this paper we present a brief tutorial on the well-studied, June 2008.
but not fully solved, problem of LPR. The main goal is to [7] Y. Tian, K. H. Yap, and Y. He, “Vehicle license plate super-resolution
using soft learning prior,” Multimedia Tools Applicat., vol. 60, no. 3,
provide the young researcher with sufficient information
pp. 519–535, 2012.
about this topic and possible solutions giving emphasis on [8] K. V. Suresh, G. M. Kumar, and A. N. Rajagopalan, “Superresolution
image processing techniques. To this end, the reader is of license plates in real traffic videos,” IEEE Trans. Intell. Transport.
Syst., vol. 8, no. 2, pp. 321–331, June 2007.
encouraged to visit www.lpr-tutorial.info in order to down-


IEEE Intelligent transportation systems magazine • 67 • Spring 2014

You might also like