Volume 8, Issue 4, April – 2023 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

License Plate Detection, Recognition and Speed

Estimation of Vehicles
Shivangi Gupta1, Dhruv Kumar Sharma2,
1 2
Student, Student,
Dept. of ECE, Jaypee Institute of Information Technology, Dept. of ECE, Jaypee Institute of Information Technology,
Noida, India Noida, India

Mandeep Singh Narula3

3
Assistant Professor,
Dept. of ECE, Jaypee Institute of Information Technology,
Noida, India

Abstract:- The goal of this research is to develop a model The paper explained the use of wavelet analysis for
for detecting and recognizing number plates as well as license plate location and suggested a type of location
detecting speed of vehicles. Maintaining security in the method based on the unique texture character of the given
neighborhood is a tough task in the current times. To license plate [2]. The authors tried to improve the algorithm
address this issue, a surveillance system that uses license for low-contrast images for accurate detection of the license
plate recognition to monitor the visiting vehicles and plate.
speed detection to monitor the over speeding vehicles is
being presented as the solution. We have compared The paper examines the effectiveness of employing
different model’s performance for detecting and license plate recognition to obtain all vehicle-related
recognizing license plates. The best approach has been information [3]. To raise the standard of the vehicle pictures
chosen and implemented to get results. fusion technique is applied to extract the number plate,
separate the characters that are present on the number plate,
Keywords: Vehicle Detection, Number Plate Detection, and then use an ANN to determine the characters on the
recognition, Speed Estimation, Yolov5, Tesseract. number plate. The experiment concluded that even in the
presence of noise with a 50% density, a neural network can
I. INTRODUCTION properly identify the characters on a license plate with a
probability of 95%.
As the need for monitoring and tracking of vehicles on
roads is increasing, we aim to develop a surveillance system The author has evaluated the efficacy of three methods
that will detect the number plates of vehicles and their speed for detecting license plates in India [4]. In order to test their
to enhance the safety and security of roads. In the current accuracy on both still photos and live stream videos,
times, the amount of traffic on the road is increasing steadily OpenALPR, k-NN, and CNN based approaches are chosen
every year. Increase in the number of vehicles on road make and customized for the local environment, and used.
it important to monitor the vehicle number plates and speed
to ensure the safety and security on roads. As a result, deep Chomtip Pornpanomchai and Nuchakarn
learning models can be used to detect and track a vehicle Anawatmongkonhave proposed a study to assess the
using its license plate and also estimate its speed for safety efficacy of a simple algorithm used to locate license plates
purposes. in a frame of video [5].

 Related Work The research aims to identify a license plate on the

An algorithm for license plate detection based on top- front or back of a vehicle. The application allows users to
hat transform and wavelet transform has been presented [1]. snap screenshots from the video file, which will then result
The approach followed the following steps: the vertical in the screenshot. The application will also pinpoint the
gradient of the gray image is calculated and the region of location of the license plate in that screenshot. The
license plate is distinguished by using top-hat application stores the image it captures as a jpg file.
transformation, then the horizontal projection of gradient
variance is obtained and multiresolution features of A method for detecting license plates using a Support
candidates are extracted by using wavelet transformation. Vector Machine classifier with Histogram of Oriented
Lastly, the license plate's precise vertical position is Gradients features is discussed [6]. The system runs window
recognised and recovered from the image. 360-degree searches at various scales, analyzes the HOG feature using
photos from various landscapes and environments were used an SVM, and finds the bounding boxes of those features
in experiments. The system immediately identified the using the Mean Shift technique. The lengthy scanning
license plate region, and the success percentage for procedure is sped up using edge information.
successfully detecting license plates is 98.61%.

IJISRT23APR1936 www.ijisrt.com 1608

Volume 8, Issue 4, April – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
The results on license plate detection indicate that this  Proposed Solution
method is mostly unaffected by changes in lighting, license
plate patterns, camera angle, and background. Alignment is
then done on the plate candidates after license plate
detection. It is suggested to segment the license plates. The
ideal segmentation—where the characters are all segmented
but not broken up—is what this model looks for. Finally,
another SVM classifier including raw features, vertical and
horizontal scanning features, recognises the characters. The
results indicated that 99% of the digits were correctly
identified, while 95% of the letters were.

Mangala A.G and Dr. Balasubramani R (2019)

reviewed various methods for speed detection of vehicle
detection. Various methods used included doing background
image subtraction, object tracking, motion vector
methodology, absolute, and centroid method. They used
MATLAB to process the data [7].

The author introduces a new approach called Speed

Detection Camera System (SDCS), a Radar substitute [8]. It
can calculate a vehicle's speed by using a variety of image
Fig 1 Proposed Solution
processing techniques on a video stream that is recorded
from a single camera. The limitations of radar are solved by
 Number Plate Detection
providing a cheaper and better solution with equal or greater
For the purpose of number plate detection, we used
precision.
two approaches:
A non-intrusive video-based method for assessing
 Tensorflow
vehicle speed on city streets is suggested by the author in
 Yolov5
[9].The technique effectively locates vehicle license plates
in moving image areas by integrating a novel text detector to
Our dataset consisted of images of number plates of
an upgraded motion detector. Later, selected features on the
cars. The images had number plates of various colors and in
license plate regions are tracked over numerous frames and
different orientations.
perspective distortion is adjusted. Vehicle speed is
determined by comparing the tracked features' trajectories to
established outside observations. For more than 96.0% of
the instances, the speeds were inside the [-3, +2] km/h limit
specified by regulatory bodies in many countries, with an
average error of -0.5 km/h. There was precision of 0.93 and
recall of 0.87.

The author states the use of the Haar cascade classifier

for object detection as a possibility [10]. In this work, fake
drivers in the flow of traffic are found and tracked using
image processing techniques. To find automobiles, the Haar
Cascade Classifier is employed. The vehicle recognised in
the image's axis coordinates are assessed, and an effort is
made to pinpoint the vehicle using these axis coordinates.
The system's accuracy rate reaches a very high value. The Fig 2 Object Detector Anatomy from [11]
Haar cascade classifier is a reasonable choice for object
detection, according to this study. In our first approach, we used tensorflow object
detection api to detect, locate, and trace an object from an
 Problem Statement image or a video. The api was used to locate the number
The need for monitoring and tracking vehicles is plate in an image using different car images and annotations
increasing due to safety and security issues. The objective of to verify the results.
this study is to develop a dependable and effective
surveillance system that will give authorities the information In our second approach, we used YOLO (You Only
they need in real-time to ensure public safety and efficient Look Once) [17] for object detection. A use case for which
traffic management. YOLO is intended involves the creation of features from
input photos for object detection. Following that, a

IJISRT23APR1936 www.ijisrt.com 1609

Volume 8, Issue 4, April – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
prediction system is given these features in order to create characters in the number plate. Hough transform solves this
boxes around objects and determine their classes. issue, making recognizing characters much easier and
accurate.
The YOLO model was the first object detector to
include class labels into the end-to-end differentiable We used two methods for optical character recognition
network process of predicting bounding boxes. There are to read the characters on the license plates:
three key components to the YOLO network.
 EasyOCR
 Backbone: A convolutional neural network that gathers  Tesseract
and creates visual features at various granularities.-
 Neck: A number of layers that integrate and mix visual Developers of computer vision software can easily
features before sending them on to prediction. perform optical character recognition using the Python tool
 Head: Consumes neck-related features and performs known as EasyOCR. Jaided AI, a business that specializes in
class and box prediction procedures. optical character recognition services, developed and
maintains the EasyOCR package. The PyTorch library and
Python are used to implement EasyOCR.

Fig 3 Representation of YOLO Process from [12]

Fig 5 EasyOCR Pipeline from [14]
We fine-tuned the object detection model to detect a
type of object over different types of objects and then fine- EasyOCR is a simple light-weight package that
tuned it on a dataset of number plates so that it would detect performs simple, straightforward OCR. It performs great
number plates in videos and images. when using a clean document or image. Camera captured
images tend to have some distortions hence it was a good
but not the best technology for optical character recognition.

Fig 4 The Structure of YOLOv5 from [13]

We used YOLOv5 for detection of the license plates. It Fig 6 Tesseract OCR Engine Architecture from [15]
is responsible for the bounding boxes around the object to
be detected, here the object being license plates. The Tesseract is an open-source optical character
YOLOv5 model does not have a particular feature like recognition (OCR) engine. It is acknowledged as one of the
accuracy. The performance can be measured using the most well-known, reliable and free-of-cost open-source
confidence of every object that is being detected in the OCR engines. It is currently one of the best technologies to
frame. IOU is another metric to evaluate its performance implement character recognition. It is lightweight and easy
which compares the true bounding box to the predicted to use.
bounding box.
We used tesseract OCR for optical character
 Number Plate Recognition recognition and it performed well in reading number plates.
In image processing, the Hough Transform is a feature Even in video captures, it was able to extract the correct
extraction method used to locate instances of objects within number plate. It took different frames from the video and
a particular class of forms. We have used it for skew applied the OCR to extract number plates even on moving
correction of the license plates. When a license plate is vehicles. The text read from the license plate was stored in a
captured, its orientation may vary and there might be csv to create a database of all the vehicles captured by the
skewness that will make it difficult to recognize the camera.

IJISRT23APR1936 www.ijisrt.com 1610

Volume 8, Issue 4, April – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Speed Estimation II. RESULTS
We have used open cv to track the speeds of various
cars on multi lane roadways. For every nth frame, it makes We have designed a system to detect license plates and
use of the Haar Cascade Classifier to find automobiles. For speed with good accuracy. It detects the license plate using
faster processing, it removes superfluous portions from the YOLOv5 and reading the characters of the number plate is
image. done using Tesseract OCR. The performance of the system
The process involves setting up two reference lines that is as expected.
have been specified for vehicle entry and exit. Any vehicle
in any lane that crosses the entrance point has its time and
location recorded. Utilizing centroid tracking techniques,
tracking is performed. The moment the vehicles pass the
exit line, time is taken. The projected speed of the vehicle is
based on the time difference. Correlation tracker from the
dlib library has been used for tracking vehicles on the road.

 Implementation using Raspberry Pi

The Raspberry Pi is a low-cost, "credit card-sized
computer" that plugs into any display device. It is a very
user-friendly tool, capable of everything a desktop computer
is supposed to be able to accomplish [16].

Fig 8 License Plate Detection and Recognition

Fig 7 Raspberry Pi Hardware Architecture from [18]

The footage is then sent from the camera mounted on

the Raspberry Pi to the client device, which is our own
server, over the internet. Every image in the video is taken
Fig 9 Speed Estimation with Speed Limit 30kmph
individually and saved as a JPEG file before being added to
a buffer. Because the video is handled on the computer
III. CONCLUSION
frame- by-frame and by many threads, the streaming of
videos is susceptible to asynchronization issues. As a result,
The main purpose of this project is to create an
the frame will be processed ahead of the others if one thread
inexpensive surveillance system using readily available IOT
has less work to do than the others. Race conditions may
devices employing license plate detection and speed
result from this, which will cause the frames to be out of
estimation for a secure and safe surrounding.
sequence and the video to not be correctly stitched.
In this work, we present a method for detecting and
We need to employ a video buffer to hold some frames
recognising the number plate details on different license
for transmission in order to get around this issue. Keeping
plates. We used an object detection model (YOLO) to detect
this buffer guarantees that the frames from the Raspberry Pi
and extract the license plate for recognition. Then we used
server reach the client in the proper order and at the proper
an optical recognition technique, which is in charge of
time, even if there will be a slight delay between the
segmenting and recognising the number plate details The
captured video and the streamed video.
data is then stored in a csv file to maintain record for
tracking vehicles for security purposes. We have also used
Additionally, the video is compressed to reduce
speed detection techniques to detect speed violations. A
network traffic and to make it simpler for our computers to speed limit can be set manually to detect the cars that violate
read and decode since raw video is more challenging for the speed regulations. Hence, the system serves its purpose
computers to process due to its larger size and incompatible
to ensure safety in the area.
file formats.

IJISRT23APR1936 www.ijisrt.com 1611

Volume 8, Issue 4, April – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
REFERENCES [15]. GitHub-JaidedAI/EasyOCR: Ready-to-Use OCR
with 80+ Supported Languages and all Popular
[1]. Feng Yang and Fan Yang, "Detecting license plate Writing Scripts Including Latin, Chinese, Arabic,
based on top-hat transform and wavelet transform," Devanagari, Cyrillic, etc.Available online:
2008 International Conference on Audio, Language https://github.com/JaidedAI/EasyOCR.
and Image Processing, Shanghai, 2008, pp. 998- [16]. Gjoreski, Martin & Zajkovski, Gorjan & Bogatinov,
1003; Aleksandar & Madjarov, Gjorgji & Gjorgjevikj,
[2]. Ye Yunyang, Chen Wenming, Hu Junda and Zou Dejan & Gjoreski, Hristijan. (2014). Optical
Wenbin, "Research on localization method of vehicle character recognition applied on receipts printed in
license plate based on wavelet analysis," 2010 macedonian language. 10.13140/2.1.1632.4489.
International Conference on Networking and Digital [17]. Khalil, Faisal. (2015). Raspberry Pi Personal Cloud
Society, Wenzhou, China, 2010, pp. 672-675; Storage. 10.13140/RG.2.2.18628.88960.
[3]. Koval, Vasyl & Turchenko, Volodymyr & Kochan, [18]. You Only Look Once: Unified, Real-Time Object
V. & Sachenko, A. & Markowsky, George. (2003). Detection Joseph Redmon, Santosh Divvala, Ross
Smart license plate recognition system based on Girshick , Ali Farhadi.
image processing using neural networks. 123 - 127. [19]. Pathinettampadian, Karthikeyan & Sivamurugesan,
10.1109/IDAACS.2003.1249531. Vasuki & Karthik, K. & Murugesan, Sakthivel.
[4]. B. S. Prabhu, S. Kalambur and D. Sitaram, (2018). Design and Development of Laplacian
"Recognition of Indian license plate number from Pyramid Combined with Bilateral Filtering Based
live stream videos," 2017 International Conference Image Denoising: Second International Conference,
on Advances in Computing, Communications and ICSCS 2018, Kollam, India, April 19–20, 2018,
Informatics (ICACCI), Udupi, India, 2017, pp. 2359- Revised Selected Papers. 10.1007/978-981-13-1936-
2365; 5_24.
[5]. C. Pornpanomchai and N. Anawatmongkon, "Thai
License Plate Detection from a Video Frame," 2009
WRI Global Congress on Intelligent Systems,
Xiamen, China, 2009, pp. 533-537;
[6]. Xin Li, Dr. Zhihai He.(2010).Vehicle License Plate
Detection and Recognition, University of Missouri.
[7]. Ramasamy, Balasubramani. (2017). A Review on
Vehicle Speed Detection using Image Processing. 4.
23-28.
[8]. Ibrahim, Osman & ElGendy, Hazem & Elshafee,
Ahmed. (2011). Speed Detection Camera System
using Image Processing Techniques on Video
Streams. International Journal of Computer and
Electrical Engineering. 3. 771. 10.7763/IJCEE.
[9]. D. C. Luvizon, B. T. Nassu and R. Minetto, "A
Video-Based System for Vehicle Speed
Measurement in Urban Roadways," in IEEE
Transactions on Intelligent Transportation Systems,
vol. 18, no. 6, pp. 1393-1404, June 2017, doi:
10.1109/TITS.2016.2606369.
[10]. Luvizon, Diogo & Nassu, Bogdan & Minetto,
Rodrigo. (2016). A Video-Based System for Vehicle
Speed Measurement in Urban Roadways. IEEE
Transactions on Intelligent Transportation Systems.
PP. 1-12. 10.1109/TITS.2016.2606369.
[11]. Gowsikraja P, Thevakumaresh T, Raveena M,
Santhiya.J, Vaishali.A.R.R. (2022). Object Detection
Uing Haar Cascade Machine Learning Algorithm.
[12]. Tan, Mingxing & Pang, Ruoming & Le, Quoc.
(2019). EfficientDet: Scalable and Efficient Object
Detection.
[13]. Bochkovskiy, Alexey & Wang, Chien-Yao & Liao,
Hong-yuan. (2020). YOLOv4: Optimal Speed and
Accuracy of Object Detection.
[14]. Z. Chen, L. Cao and Wang ,YOLOv5-Based Vehicle
Detection Method for High-Resolution UAV Images.

IJISRT23APR1936 www.ijisrt.com 1612