Automatic Number-Plate Recognition

R A. Lotufo A. D. Morgan A. S. Johnson

Faculty of Engineering
University of Bristol
Bristol BS8 1TR

1 Introduction
“WEc engineers and the police often wish to be able to uniquely identify road vehicles at different
points in a road network for a number of nssons[l],such as:
automatic toll-collection, road pricing schemes, and operation of ear parks for revenue col-
lection;
fleet control for monitoring and optimisiig the movements of particular types of vehicles
(e.g. heavy goods vehicles);
speed- and weight-limit enforcement purpoees;
crime detection/prevention; and
tr&c data coUection (e.g. vehicle journey times, origin/destination and entry/&t flow
patterns, etc.) for trafEc operation and planning.

Until quite recently, vehicle identification has largely been a labour intensive manual activity,
aided where possible by electromechanical actuators. With the continuing increase in volume of
road traffic, such manual techniques have proved impmctid. Improvements in electronic seneor
technology ham allowed for the development of automatic vehicle identification (AVI) equipment.
Conventional AVI system empioy RF., microwave and inductive loop technologies[2]. In such
systems, a unique code identifying each vehicle is utored on a vehicle mounted transponder. As
the vehicle passea specific points in the roadway, a roadside reader unit detects vehicle praence
and reads the coded inhrmation on the vehicle mounted transponder unit. Most AVI installations
additionally use a video/CCTV subsystem for photo/videologging vehicles that are unequipped
with transponders. These video recordings are often manually reviewed, off-line, for extracting
the registration number platen of each defaulting vehide.
Although coventional AVI installations are b m i n g widespread, the practical implementation
of such system on a large sede requires mounting tramponders on a very large number of vehicles.
They also have the need to ensure the reliable operation and maintenance of these transpondem.
The latter requires “e means of performing automatic ch& to detect any malfunction in the
vehicle mounted transponder. The alternative involving the manual interpretation of vehicle reg-
istration number plates from video recordings of t r a c ie tedious and expensive, especially for
heavily used AV1 installations.

AVI Using Computer Vibn: Rzcent advances in computer vision technology and the falling
prices of related devices ham extended the use of video/CCTV by making it practical to automat-
ically identify vehicles visually either on-line or off-lie.
The objective of our research is to develop a Computer-Vision based AV1 system to achieve
vehicle identification using optical character recognition (OCR)techniques. Our work involves

6/1

Authorized licensed use limited to: Indraprastha Institute of Information Technology. Downloaded on November 23,2021 at 06:12:52 UTC from IEEE Xplore. Restrictions apply.
 investigations into real-time automatic Number-Plate Recognition (NPR) and its extension to
other apecb of road trafEc monitoring and control.
Automatic Number-Plate Recognition avoids the need to equip vehicles with a special transpon-
der, since all vehidea already have a unique registration number plate. Furthermore, most of the
conventional AVI systema currently in operation already incorporate llom form of video/CCTV
subsystem. T h i means that the only additional ccst to exisring inetallations will be that of the
automatic image analysis subsystem.

2 The Number-Plate Recognition System

Design Requirement.
The b a i c requirement for any Number Plate Recognition system, for practical implementation,
is real-time performance with a high recognition accuracy. Thus, the following basic performance
requirements can be identified:
1. High recognition speed (approx. a eeeond per vehicle);

2. High recognition accuracy (typically more than 95%);

3. High level of consistency

The specification for a number plate image consists of a clean picture with the number plate
occupying at least 25% of the width of the image and with a minimumcontrast factor of 4 between
the characters and background.
Configuration
A typical application ~ o n s i s bof a configuration where a camera is mounted so as to capture
an image of the fmnt number plate of a vehicle entering a car park. When a vehicle arrives at
the barrier, an inductive loop detector triggers the number plate recognition system. The number
plsk is extracted and decoded kom the image, which can be compared to a data base, or stored
for later analysia. In moet applieatio~,the requirement of real-time operation is fundamental:
the number plate must be decoded within couple of seconda from the capture time. This impoees
a large demand on the speed of image processing operations.
System Architecture
The number plate recognition system can be divided in two main parte: image acquisition and
image analysis. The image acquisition part ~onsisteof the vehicle sensor, the camera and the video
digitiser. The image a n d y s i ~is divided in three further sections: the location of the number plate
in the captured image, character feature extraction and character claaaification.
The succeeafull operation of the number plata reader depends on both parts. The image
acquisition must deliver a clean picture of the number plate with good illumination, minimum
blur and reaaonable resolution. The image analysis must identify the number plate and correctly
decode it within Ileconds.
Hardwam Architecture
The system COM~S~Sof four basic camponente: vehicle seneor input, video interface input, video
interface output and transputer framestore. The functions of each component are as follows.
vehicle sensor input: a digital input port, addressable by the processor to detect the right
time to acquire a picture.
video interface input: analog video input is digitised to 8 bits, passed to a lookup table
and output to buffer memory. The interface can be synchronised externaly with the video
synchronism extractor and frequency lock circuit. Thie u e a a gated oscillator with good
start-up characteristi- rather than the common phaee-locked loop solution which can often
require a t h b a s e corrector with VCR signals. Thia enables the grabbing of pictures both
from a Video caasete Recorder and TV camera.

Authorized licensed use limited to: Indraprastha Institute of Information Technology. Downloaded on November 23,2021 at 06:12:52 UTC from IEEE Xplore. Restrictions apply.
 0 video interface output: converb digitised video data from the video bus to RGB analog W
dgnaln suitable for displaying in a colour monitor. The colour map function L implemented
using three 6 bit by 256 lookup table.
0 tramputer framaetore: the digitised video data is stored in a dual memory static RAM
addressable by the transputer. The video memory L organised aa a double buffer allowing
video data to be grabbed in one buffer concurrently with the processing of the previously
captured frame. The transputer framestorehaa also a digitised video output port to transmit
video data to the video interface output. The processor is b d on the 20MHo T414 or T800
32 bit transputer with 512 Kbyte of faat static RAM. Four INMOS serial links are provided
together with the standard network sign& Error, Reeet and Analyse. This interface enables
the expansion of the CPU power capability by an indefinite number of transputers.
0 hoot: a PC IBM compatible fitted with a tramputer card or an I"link interface adapter.
A prototype system based on three different board types haa been designed, involving:
0 a video interface card providing A/D and D/A conversion of video signals for image capture
and display, and vehicle sensor interface,
0 a processor card, baaed on the T800 32 bite floating point transputer, and
0 a framestore card consisting of a dual 256x256~8double buffer or a single 512x512~8buffer
memory accesled independently from the video interface card and from the processor card.
The proeeseor card aud the frameetore card make the transputer framestore component of the
system. Each of the cards uea a double Eurocard form factor so they may be racked together
with a variety of commercially available boards.
Image A n a l y h
The Number Plate Reading algorithm is divided into three main tasks: number plate location,
character detection and character cladication:
0 The number plate location task sann the image looking for the area of the number plate
and outpub the coordinates of the number plate.
0 The character detection task operates on the the number plate region, identifying each
individual character and extracting its relevant features.
0 The character claasiikation task gathem the features extracted for each character and das-
aifies them using a pre-stored data base.
For number plates with a known structure, such aa British number platea, where there are
rules for the position of letters and numbers, a further step can be used: a syntax checker and
corrector which uses the alphanumeric rules and probabilities from the character classification task
to c o n h n or correct the selected characters.
The algorithm
The images acquired are onen visually noisy and invariably the scene is unevenly illuminated.
For thia retwon, the system first executes an adaptive thresholding algorithm on the entire scene.
This aids in properly segmenting the vehicle number plate and ita constituent characters from the
rent of the scene.
A boundary following algorithm is then applied to the resulting binary image to locate the
number plate. This is achieved by a spatial scanning procedure which searches for closed contours
that have an appropriate size and aspect. If a number of these closed contours are found then it
in wumed that their position in in the area of the number plate.
Once the location of the number plate haa been found, a second threshold is computed and
applied to the number plate area to properly segment the eharactem from the rest of the plate
background. A similar boundary tracking technique is then performed, within the plate area, to

613

Authorized licensed use limited to: Indraprastha Institute of Information Technology. Downloaded on November 23,2021 at 06:12:52 UTC from IEEE Xplore. Restrictions apply.
 Figure 1: Original image and threshold image showing the boundary tracking in the region of the
number plate.

extract the character features. The figure above shows thia scanning process. Typical features
include number of holea, comers, boundary chain codea, bounding box, etc.
The next stage performa a detailed matching process between the extracted character features
and the reference features (contained in a database) using a statistical nearest neighbour classifier.
This process uses stored templates of each character in each font to compare with the character
image. Likelihood/confidence measurea are determined tiom each comparison and the characters
with the highest confidence leveb are stored as potentid candidah for each character. The
character having the highest confidence is automatically selected 8~ the mast likely candidate.
Having obtained an identity for each character, the format of the extracted characters is checked
against the standard number plate format to validate whether in fact thin string of characters con-
stitutes a valid vehicle number plate. Thia is achieved using a syntax checkmg/forcing algorithm.
Thia examines the type and positional validity of each character for each syntax type (only the
four most prolific syntax types have been examined so far) and performa syntax forcing if the
wrong type is dekcted. This is achieved by replacing the wrong character with the appropriate
type from the other potential candidater recorded earlier in order of likelihood.

3 System Performance and Evaluation

There are two basic figures relevant to number plate reading: the reaponae time and the confidence
of the extracted number plate. Thew figures are related and a trade-off is required to achieve the
desired confidence and response time. The hardware architecture enables the expansion to parallel
processing in order to speed-up the intermediate t a b simply by adding standard off-thashelf
transputer products. Most of the t a a b allow natural parallel decomposition and could benefit
from this pardel proceasing capability.
The overall system can be seen as a pipelind structure, having at the initial stage, the image
input and at the output the decoded number plate. As a pipeline structure, all components play
a crucial part in the overall performance of the system. ksauming the confidence factors of each
part arc statistically independent, the total system confidence is the product of the confidence
factors of its parts.
Evaluation
An important aspect of the project is the method used to evaluate the system. An appro-

Authorized licensed use limited to: Indraprastha Institute of Information Technology. Downloaded on November 23,2021 at 06:12:52 UTC from IEEE Xplore. Restrictions apply.
 priak muwrement of the confidence and respotme time pcrrameten of the systems requim the
application of a very large number of typical imagea to the number plats reader. The evaluation
procedure must be repeated to aimam the resulh of a mDdification to the recognition algorithm.
To continuously iwaeao the system during development, an automatic evaluation procedure was
designed:
Afkr recording scenesof real situations onto a video caascte, a sound burst signal wae dubbed
onto the audio channel to simulate the vehicle sensor input. A referenee file is created with the
correct number plate sequence recorded in the tape. The number plats reader can autonomously
procesr, the video tape digitsing the image at each sound trigger, reporting the processing results
into a "log" tile. At the end, an evaluation program, reading the "log" file and the reference file
compuh the performance figures of the system.
In order to analyse and quantify the various sourcea of error in the overall system, a simplified
pipeline model can be used whose components are described below along with the contributing
sourcea of error:
acqukitbn This stage considers the details of the camera position and pwibly the arrangement
of appropriate illumination to provide a clear view of the number plate that is free from
shadom. Our teat data suffus from a number of the fohwing problems,
broken number plate, plak partidy visible, poor illumination and contrast, number plate
not within the legal specification, low reeolution of the characters.
number-plate location This stage of the algorithm attempts to find the paition of the plate in
the image by scanning for c l d contours with the appropriate character shape. This is the
mrat robust stage in the algorithm and only faib when an obvious fault in the acquisition
stage has occured. Other sou~cesof error at the moment are,
more than one number plate in the image, difference in background/foreground colours (old
licence plates).
character feature extraction Thi.~stage extracb the c l 4 contour around the character to-
gether with additional information such an 'holed etc. The main sources of error here are
due to poor maintenace of the vehicle plate,
i.e. character corrupted by bolt, broken character contour, joined characters.
character feature matching The basic pattern matching process has a fundamental problem
with the similarity between certain charactem, namely, 0 and D, 5 and S,8 and B (even for
human recognition). The syntax che&g/forcing stage is capable of distinguishing between
similar numbers or lettem but the OD probkm r e d and to a lesser extent the CG
problem.
In addition, the noise introduced by the camera and electronic circuits can affect each of the
last three stages becauee each stage procesaea the raw image data.
It is desirable to mulllure the performance of each stage to assean its impact on the overdl
performance. Using thin simplistic model of a pipeiiined structure and aesUming the success ratio of
each part is not correlated with the others, the final success ratio is the product of the intermediate
ratira.
If it ia aeeumed that the acquisition problems can be avoided by proper operation of the sys-
tem then the number plate location stage is nearly perfect and can be overlooked for the moment.
Indeed, the plate location stage could be considesed as a reliable filter for poorly maintained
number plates. The two remaining stages of character feature extraction and matching are, corn-
paratively, the least reliable and require further analysis. The overall performance measure for
the system specifies the percentage of correctly claseified number plates. This implies the correct
extraction and classification of, on average, seven characters. However, an error is usually due to
the migclaseification of a single character and it is thus more informative to quote claseification
percentages for individual characters to identify the problems. An idea of the problem can be
gained from initially assuming that the success ratio in processing each character is the same,

Authorized licensed use limited to: Indraprastha Institute of Information Technology. Downloaded on November 23,2021 at 06:12:52 UTC from IEEE Xplore. Restrictions apply.
 then the overall success percentage is the the success percentage of processing a single charater to
the power of the number of charactem in the number plate. That is, a very steep curve, i.e. 95%
character succeae equates to 70% plate succeae and 99% character success equates to 93% plate
success.

4 Results and Conclusions

The results of running the evaluation system on more than one thousand cars, considering only
those number plate which are within the specification are as follows:
0 95% success in the feature extraction ( 99% at each character )
0 80% succesll in the feature matching ( 97% at each character)

given a resultant 76
The main reason of the poor performance on the feature matching can be explained by the fact
the matching process does not include yet a good diecrimination function between the character
pain D and 0 and C and G.
From first samples of recorded taped, there are a large percentage of number plates which are
not within the specifications. Although part of that waa due to inexperience in the operation and
set-up of the camera, there exists a number of consideratiom in the acquisition of the image:

To get a large number plate within the camera angle of view, it is necessary to restrict in
some way the parition of the car during the image grabbing.
0 The illumination can be very poor depending on weather conditions, especially in number
plaks fixed below the bumper of the vehicle.
0 The use of pre-recorded video tape introducea a significant amount of noise, mainly because
of the poor line synchronisation. On many O C C ~ ~ ~ Othis
M was found to distort the character
image and thus affect the claaeification stage. In future, a 'time base corrector w i l l be used
to overcome this problem.

5 Acknowledgements
Thin research is funded by Golden River (Remote Data Collection Systems). The authors would
l i e to acknowledgethe advice and encouragement given by Professor E.L.Dagless, Dr.B.T.Thomas
and Dr D.J.Milford and the assistance of Dr.RDanbury and Mr P.Clarke. Mr A.S.Johnson would
like to thank the British Council for their financial support

6 References
1. Salter, D. R. (19841"The Potentialof Automatic Vehicle Identification", IEE Conf. on Road
Traffic Data Collection, 242, pp 79-82.
2. Ayland, N. and Daviea, P. [t989] "Automatic Vehicle Identitication for Heavy Vehicle Mon-
itoring", IEE 2nd Int. Conf. on Road Traffic Monitoring, 299, pp 152-155.

Authorized licensed use limited to: Indraprastha Institute of Information Technology. Downloaded on November 23,2021 at 06:12:52 UTC from IEEE Xplore. Restrictions apply.