Special Issue - 2021 International Journal of Engineering Research & Technology (IJERT)

ISSN: 2278-0181
NCREIS - 2021 Conference Proceedings

Theft Vehicle Detection using Automatic License

Plate Recognition
Aromal A, Brinda H Shilpa J, Sreebala S
Scholar (CSE), College of Engineering Perumon Scholar (CSE), College of Engineering Perumon
APJ Abdul Kalam Technological University, APJ Abdul Kalam Technological University
Kerala, India Kerala, India

Soumya K S
Assistant Professor (CSE), College of Engineering Perumon
APJ Abdul Kalam Technological University
Kerala, India

Abstract— A large number of automobiles causes a database. Typically, the identification system will be put at
variety of issues and disruptions in daily life. Vehicles the residential area's gate, factory entrances, parking
contribute to traffic congestion on the road and vehicle theft spaces, toll gates, university entrances, or other high-
in parking lots. Vehicle and transportation management is a security buildings like defence institutes and nuclear power
time-consuming and laborious undertaking. If everything is
done by hand, there will be a lot of mistakes and challenges.
plants. Camera reflects the presence of a vehicle. It takes a
As a result, it is important to design an automatic recognition picture of the car that will be utilised in future processing.
and detection system for vehicle number plates. This paper A vehicle registration plate is a plate that is normally
sheds some light on the issue at hand. With the YOLO V3 attached to the front or back of a vehicle and is composed
algorithm and Canny Edge Detection, the recognition system of either metal or plastic. Numbers and alphabetical letters
will automatically recognize the front number plate of make up the vehicle plate number. The car's initial two
automobiles. Following the detection of a license plate, the letters serve as the state location prefix, followed by two
following actions are taken: 1.To take a picture of the license number digits that represent the district to which the
plate. 2. To recognize and segment characters. 3. A graphical vehicle belongs. Number plate format of Indian vehicles
user interface displays the recognized license plate, which is
then saved in a database with the time and date for later use.
are LLNN LLNNNN where L is letter and N is the number.
4. If a stolen vehicle is discovered, a notification stating that Section II covers the extensive survey on the Literature.
matches have been found is displayed, and a full report Section III describes the methodology of implementing the
containing the vehicle and the time it appears is generated. . detection of theft vehicles using Automatic License Plate
The technology can be utilized for security and accuracy Recognition. Followed by the result of the implementation
purposes. The license plate number is displayed on the in Section IV. Finally the report is wrapped up with
graphical user interface and saved in a database with the time conclusion in Section V and the list of referenced papers as
and date for later use. Section VI.
Keywords—Automatic License Plate Recognition; Theft
Detection; Character Recognition; Canny Edge Detection;
II. LITERATURE SURVEY
YOLO V3 Algorithm.
I. INTRODUCTION A. The Recognition And Detection System For Theft
In today's world, the number of vehicles is rapidly Vehicle By Number Plates
expanding. It is important to get from one location to The system detect and recognize license number
another within a certain amount of time. We can see that of captured front view image of any vehicle by camera. It
there are a lot of vehicles around us. Vehicles are required contains main three processes: plate extraction, character
by everyone for various purposes. Vehicles have increased detection and character recognition. The stolen vehicle is
dramatically in population proportion over the last two detected by comparing it with database of stolen vehicle
decades. However, it causes problems and difficulties in provided by the police station. Afterword, it rings an alarm
human life. It leads to issues such as heavy traffic, loud and sends a SMS to the police station. It can be used in
noise, and criminal activity such as vehicle theft and many security purposes like speed detection, detection of
accidents, among other things. As a result, vehicle traffic violation, toll collection, parking system. It can also
management and administration are critical to avoiding the beneficial to secure areas like the gate of residential area,
aforementioned societal issues. As a result, several efforts factory gates, parking space, toll plazas, university entrance
are being made to improve the challenges in vehicle or other high-secured building such as defense institutes
transportation. Vehicle Plate Recognition System is the and nuclear factories. Block diagram of entire system is
most appealing study topic to the most recent researcher. A shown in Fig 2.1
VPR system is a vehicle tracking system that identifies the
vehicle so that it may be tracked down using an existing

Volume 9, Issue 13 Published by, www.ijert.org 77

Special Issue - 2021 International Journal of Engineering Research & Technology (IJERT)
ISSN: 2278-0181
NCREIS - 2021 Conference Proceedings

Vehicle Database in 0.2 second. The recognizable vehicle

speed is from 0 to 100 KM/H.

Fig.1 Block diagram of number plate detection system in toll plaza.

Fig.2 Architecture diagram of National Detecting Stolen Vehicles
When, a vehicle appears in the toll plaza. It will be Network System
stopped by the gate and the image of that vehicle will
captured by the camera. The camera sends the image of C. An Efficient License Plate Recognition System Using
appeared vehicle to microcomputer unit. It extracts number Convolution Neural Networks
plate and segmented characters. It also recognizes the This paper propose an efficient license plate
getting information with help of MATLAB software. If recognition system that first detects vehicles and then
appeared that the vehicle is stolen according to database retrieves license plates from vehicles to reduce false
,the gate will not open for that vehicle. At the same time, positives on plate detection. Then, we apply convolution
the alarm rings and sends a message to police station as neural networks to improve the character recognition of
well as it also displays on LCD of toll plaza. In the blurred and obscure images. To accommodate the
contrast, if the captured image of vehicle will not match complexity of images taken from cameras in crossroads,
and exist in the database, the door will open and buzzer this paper proposes to first detect vehicles and then detect
does not make sound or ringing. In this way, vehicle after plate on the vehicles. This method can avoid misidentifying
vehicle successfully detection is going 24 hours on toll traffic signs or advertisements as license plates. In this
plaza. paper, YOLOv2 is used to detect vehicles. The Darknet-19
model adopted by YOLOv2 has 19 convolution layers and
B. National Detection Of Stolen Vehicles
5 maxpooling layers. YOLOv2 first extracts features,
This paper deals with the design and deployment
reduces dimensions, and performs compression for an
of a National Detecting Stolen Vehicles Network System
entire image using the 19 convolution layers and 5
(NDSVNS) for a country and a police management
maxpooling layers. The original images are reduced to 7*7
authority. NDSVNS mainly uses the pattern recognition
or 13*13. Then, YOLOv2 directly performs object
techniques to setup a Stolen Vehicle Identification System
recognition and predicts object positions on the reduced
(SVIS), and integrates multiple heterogeneous databases of
images.SVM is a supervised learning method used for
different city and country police departments. Each
classification and regression analysis. The process of
database functions as a stand-alone entity, but they are also
character segmentation consists of several steps. First, the
connected by Police VPN. To share the valuable
captured image is converted to grayscale and then binarized
information, firstly Criminal Investigation Bureau (CIB)
to eliminate noise. Then, perform a horizontal projection of
unified the data format of license plate recognition records
the license plate image to determine the position of the
and exchange formats on Police VPN. Secondly, some
characters arranged on the license plate. The upper and
appropriate locations are selected in all cities and counties
lower borders are removed by horizontal projection.
and toll stations of highways to install SVIS.
Finally, perform a vertical projection on the license plate
Cameras installed on poles along major roads in
image to determine the position of each character and then
all cities and counties are responsible for capturing image
divide it into single characters. In the final stage, a
data of vehicle license plates and sending data to the
LPRCNN model to identify blurred and skewed characters.
License Plate Recognition Workstation. NDSVNS server is
The proposed LPRCNN model is composed of two
installed in the CIB Computer Center and provides police
convolutional layers, two maxpooling layers, two fully
officers of all police agencies with capability to add, query,
connected layers, and one output layer. The output layer
and modify license plate information and other functions.
contains 34 neurons to correspond to the 34 plate
The License Plate Recognition Workstation is responsible
characters.
for license plate pattern recognition, and comparing
D. Vehicle Number Plate Recognition System For Theft
suspected stolen vehicle information to the Stolen Vehicle
Detection
Database (SVD). If the vehicle is reported as stolen in
The presence of vehicle is detected using IR
SVD, the system will display a warning message on the
sensor after camera will capture image of vehicle which
screen and automatically sound the alarm to inform the
will be used for further processing. If vehicles are
police on duty. The police may proceed to arrest the thief.
recognized manually then will be more mistakes with less
License plate number is recognized and compared to Stolen
efficiency and slow. If the described system will be

Volume 9, Issue 13 Published by, www.ijert.org 78

Special Issue - 2021 International Journal of Engineering Research & Technology (IJERT)
ISSN: 2278-0181
NCREIS - 2021 Conference Proceedings

implemented artificially by using machines as shown in fig into four subsections, one for each of the ALPR stages (i.e.,
2.4 vehicle and LP detection, character segmentation and
character recognition) and one for temporal redundancy.
Fig.3.1 illustrates the ALPR pipeline. We use specific
CNNs for each ALPR stage. Thus, we can tune the
parameters separately in order to improve the performance
for each task. The models used are: Dark Flow, YOLO V3,
NumPy, Tensor Flow, OpenCV and Keras.

Fig.3 Automatic Vehicle License Recognition System

The crucial and initial step in ALPR system is to

extract the characters of number plate from the vehicle
image. The procedure for detecting is done in several
stages. The number plate extraction is started with the
horizontal and vertical edge detection techniques that are Fig.4An usual ALPR pipeline having temporal redundancy at the end
based on the characteristics of the edge displayed by the
edges of the character on the vehicle’s number plate. A. License Plate Detection
The process of character recognition involves The video from the cctv footages is given as input to the
several steps like feature extraction and classification. license plate recognition system. The resolution of camera
Before recognition algorithm, the characters are needs to be good so that the captured video can be further
normalized. Normalization is to refine the characters into a utilized for processing. Using Dark flow, we trained a
block containing no extra white spaces (pixels) in all the YOLO (You Only Look Once) model with 1900 images of
four sides of characters. Then each character is fit to equal car with annotated plate. Labeling is a graphical image
size. Fitting technique is necessary for template matching. annotation tool. It is written in Python and uses Qt for its
For comparing the characters with the database, input graphical interface. Annotations are saved as XML files in
images must be same-sized with the database characters PASCAL VOC format, the format used by ImageNet.
Template matching is a proper algorithm for recognition of Dataset was composed of car images found online. YOLO
characters. The character image is matched with the ones in v3 should be able to detect the vehicles and their LPs
the database and the best similarity is calculated. To correctly in much less time. YOLO uses A anchor boxes to
measure the similarity and find the best match, a predict bounding boxes (we use A = 5) each with four
correlation function is used. coordinates (x, y, w, h), confidence and C class
For proper user-friendly environments various graphical probabilities, so the number of filters is given by
user interfaces are created so that user can process step by filters = (C + 5) × A
step. Every recognized number plate is compared with
database of stolen vehicle if math founds then that vehicle
is declared as stolen vehicle. The data base of such vehicle
maintained for security purpose so that stolen vehicle will
be detected easily. After detecting stolen vehicle system
will generate alarm and it will close door so that vehicle
will not pass from that place. After message will be send to
the traffic police for detecting theft. Fig.5: Plate recognition

III . METHODOLOGY Plate Detection is done in 2 stages using YOLO model and
Automatic License Plate Recognition (ALPR) has been a OpenCV functions. First Crop function use best prediction
frequent topic of research due to many practical from YOLO model and return the license plate. Second
applications. However, many of the current solutions are Crop Function use OpenCV functions to crop a little more
still not robust in real-world situations, commonly of the plate to avoid noise in the background.
depending on many constraints. This paper presents a
robust and efficient ALPR system based on the state-of-
the-art YOLO object detector. The Convolutional Neural
Networks (CNNs) are trained and fine-tuned for each
ALPR stage so that they are robust under different
conditions (e.g., variations in camera, lighting, and
background).
Specially for character segmentation and recognition, we
design a two-stage approach employing simple data Fig.6: Sample dataset image
augmentation tricks such as inverted License Plates (LPs)
and flipped characters. The proposed approach is divided

Volume 9, Issue 13 Published by, www.ijert.org 79

Special Issue - 2021 International Journal of Engineering Research & Technology (IJERT)
ISSN: 2278-0181
NCREIS - 2021 Conference Proceedings

B. Character Segmentation The recognized plate is compared with the number given
Thresholding method is used to convert the grey scale as input. A GUI is created where the wanted number is
image given as input. An error message will be displayed to enter
to monochrome. This method reduces the complexity of the number if number is not given as input. Only after
captured image (input). Once the LP has been detected, we entering the number can then the video be uploaded.
employ the CNN for character segmentation and
recognition. However, instead of performing both stages at IV. RESULT
the same time through an architecture with 35 classes (0-9, We have used number of photos of cars by using
A-Z, where the letter O is detected jointly with the digit 0), pre-trained models. A number of images of number plates
we chose to first use a network to segment the characters were used to correctly recognize plates. GUI was created in
and then another two to recognize them. The character which the vehicle ID i.e., number plate is given as input
segmentation CNN is trained using the LP patch (with a along with which a video is uploaded for detection. Only
margin) and the characters coordinates as inputs. As in the after entering the number can then the video be uploaded.
previous stage, this margin is defined based on the An error message will be displayed else to enter the
validation set to ensure that all characters are completely required number. The speed of the working depends on the
within its predicted LP. YOLO model trained with images hardware of the system. A system with good hardware can
of license plates where characters have been annotated. help in the smooth working of the model. The model is
There is only one label character around 14000 characters. trained to work with more accuracy compared to previous
OpenCV is used for image processing such as image models.
cropping etc.

Fig.9: Accuracy on training

Fig.10: GUI of model

Fig.7: Segmentation by converting to grayscale

C. Character Recognition
It is employed for the purpose of conversion of images
of
text into characters. For Character recognition trained a
convolutional neural network with Tensorflow and Keras
libraries. Tensor flow is an End-to-end open-source
platform for machine learning. Comprehensive, flexible Fig.11: Error message to enter number
ecosystem of tools, libraries and community resources that
let researchers push the state-of-the-art in machine learning V. CONCLUSION
and developers easily build and deploy machine learning Our proposed model is aimed to perform at higher
powered applications. There are 35 classes (10 for numbers accuracy than pre-existing models. YOLO Model and
and 25 for alphabet without “O”). Approximately 1000 OpenCV are used for more accuracy. Vehicles can be
images used for each classes. identified automatically rather than manual ways prevailing
today. Alert messages are sent immediately to the
authorities. Manipulating traffic signals to help recover the
wanted vehicles can be done with a few future
enhancements. The proposed system of Automatic Vehicle
License Plate Recognition can be implemented using above
discussed method which involves fundamental Image
processing steps. We can detect and recognize license
number from a captured front view image of any vehicle
from camera. It contains main three processes: plate
detection, character segmentation and character
recognition. Pre-trained models such as darkflow, keras and
Fig’8: Recognition - Data Set Sample. numpy are used. Further we can detect stolen vehicle by

Volume 9, Issue 13 Published by, www.ijert.org 80

Special Issue - 2021 International Journal of Engineering Research & Technology (IJERT)
ISSN: 2278-0181
NCREIS - 2021 Conference Proceedings

comparing it with database and a message is displayed such

that matches found and along with a detailed report is
generated. The report indicates the image of the vehicle, the
number plate and the duration in which the theft vehicle
was detected. Hence the location of the theft vehicle can be
traced. The proposed system can be effectively applied for
the identification and detection of theft vehicles. With
increased number of datasets, the accuracy of the model
has been increased.

REFERENCES
[1] Christos Nikolaos E Vassili Loumos, Anagnostopoulos, Ioannis
E. Anagnostopoulos, ,and Eleftherios Kayafas A License Plate
Recognition. Algorithm for Intelligent Transportation System
Applications IEEE transaction on Intelligent Transportation
System, Vol. 8, No. 3, Sept 2006
[2] ChoudhuryA .Rahman, Wael Badawy,AhmadRadmanesh A Real
Time Vehicles License PlateRecognition System Proceedings of
the IEEE Conferenceon Advanced Video and Signal Based
Surveillance,0-7695-1971.June 2010
[3] Banshidhar Majhi Heuristics for license plate localizationand
hardware implementation of Automatic License
PlateRecognition (ALPR) systemDepartment of Computer
Science and Engg, National Institute of TechnologyRourkela,
2012
[4] Weiping Chang1 and ChingweiSu Design and Deployment of a
National Detecting Stolen Vehicles Network System by 2012 Feb
[5] Cheng-Hung Lin1, Yong-Sin Lin1, and Wei-Chen Liu An
Efficient License Plate Recognition System Using Convolution
Neural NetworksProceedings of IEEE International Conference
on Applied System Innovation 2018
[6] Sham Madhukar Lomte, Hajiali Sayyed The Detection and
Recognition System for Theft Vehicle by Number Plates
International Journal of Science and Research 2015
[7] Hendry, Rung-Ching ChenAutomatic License Plate Recognition
via sliding-window darknet-YOLOdeep learningDepartment of
Information Management, Chaoyang University of Technology,
Taichung, Taiwan Elsevier May 2019

Volume 9, Issue 13 Published by, www.ijert.org 81