© 2025 IJRAR APR 2025, Volume 5, Issue www.ijrar.

org (E-ISSN 2348-1269, P- ISSN 2349-5138)

3

Helmet Detection And Number Plate Detection

Using Deep Learning
1
Payal Kshirsagar, 2Vaishnavi Patil, 3Dhammabhushan Wale, 4Divya Choudhari

1
Student, 2Student, 3Student, 4Professor

Department of computer science and engineering,

1
Padm.Dr.V.B.Kolte College of engineering, Malkapur, India

Abstract : In India, road accidents claim the lives of six two-wheeler riders every hour. During the COVID-19 pandemic, it was
observed that while many individuals wore face masks, they often neglected to wear helmets—mainly due to discomfort or the
desire to avoid congestion. This concerning trend inspired the development of our project, which focuses on identifying and
penalizing motorcyclists who ride without helmets and thereby violate traffic regulations.

To determine whether a rider is wearing a helmet, we employed two key techniques:

 Measuring the overlap between the bounding boxes of the rider and the helmet.
 Detecting the presence of a helmet in a defined region above the motorcycle.

Our model achieved a mean Average Precision (mAP) of 0.995. To the best of our knowledge, this approach is among the first to
use bounding box overlap as a method to associate helmet use with specific riders. For automatic license plate recognition, we
utilized EasyOCR.

Keywords: YOLO, Helmet Detection, Number Plate Recognition, ANPR, OCR

I. INTRODUCTION
Many motorcyclists suffer serious injuries or even lose their lives simply because they don’t wear helmets. In India, six two-
wheeler riders die every hour in road accidents. Even with strict traffic laws, many still ignore helmet safety. Existing surveillance
systems rely heavily on humans, which can lead to errors—so an automated solution is needed.

To address this, we used YOLOv5, a deep learning object detection model, to detect motorcyclists and check if they are wearing
helmets. Since riders wear different clothes, helmets, and appear at different angles, we used a deep neural network to improve
accuracy.

Our goal is to build a system that collects data on those who break helmet laws. The main challenge was identifying whether a
person is wearing a helmet and separating bikers from pedestrians. We solved this in two ways:

1. By checking if the helmet and rider bounding boxes overlap and linking them to the bike’s number plate.
2. By checking if a helmet appears at a certain height above the bike’s bounding box.

In the end, the system detects riders without helmets and captures their number plates for action.

IJRAR1601009 International Journal of Research and Analytical Reviews (IJRAR) www.ijrar.org 1

© 2025 IJRAR APR 2025, Volume 5, Issue www.ijrar.org (E-ISSN 2348-1269, P- ISSN 2349-5138)
3
II. MOTIVATION
Choosing a capstone project topic like "Helmet detection and number plate detection using deep learning" is not only timely
but also impactful. It is driven by both environmental and technological factors, each contributing to the overall goal of creating
sustainable and efficient solutions for traffic management service.

III. OBJECTIVES

To develop an AI-based automated surveillance system that:

1. Detects whether a two-wheeler rider is wearing a helmet or not in real-time using CCTV or traffic camera feeds.
2. Recognizes and extracts the vehicle number plate of riders who are found to be violating the helmet rule.
3. Generates a report or alert (e.g., image evidence + number plate info) to assist in issuing fines or warnings.

IV.LITERATURE SURVEY

A. Helmet Detection
Over the years, various algorithms have been used for helmet detection. Early methods like background subtraction and Gaussian
mixture models had issues with changing lighting conditions and background noise, which led to low accuracy. As technology
advanced, deep learning models, particularly CNN-based methods like Faster R-CNN and Detectron2, showed better results.
However, they still faced problems like misidentifying round objects (e.g., bald heads) as helmets. YOLO (You Only Look Once)
became popular for real-time object detection due to its efficiency in processing images. YOLOv5, in particular, has shown
strong performance in helmet detection, outperforming traditional methods in terms of speed and accuracy.

B. Number Plate Recognition

Automatic Number Plate Recognition (ANPR) systems have also evolved, with methods like optical character recognition (OCR)
used for character detection. Early systems used sliding window approaches but suffered from slow speeds and errors due to poor
lighting. YOLOv2 was used for number plate detection, but it had computational limitations. For character recognition, OCR tools
like Tesseract were used, but EasyOCR has proven to be faster and more accurate, making it a great choice for number plate
recognition in various real-world conditions.

V. SUMMARY AND DISCUSSION

These projets collectively highlight the growing role of artificial intelligence and computer vision in enhancing road safety and
engorcing traffic regulation. By automationg helmet detection and number plate detection recognition, these system contribute to
more efficient law enforcement , accident prevention , and improved public compliance with safety norms.

Here’s summary of the key component and functionalities:

 Real-Time Monitoring: The integration of video surveillance with Ai-Based models allow for real- time monitoring of
traffic violations, particularly identifying riders without helmet and capturing vehicle number plated instantly.
 Automated Enforcement: The system automates the process of detecting violations and logging vehicle details,
reducing the dependency on manual monitoring and enabling quicker and more consistent enforcement of traffic laws.
 Data Logging and Reporting: Detected violations are stored in a structured database with time, location, and image
evidence, supporting streamlined generation of reports and integration with traffic management systems.
 Scalability and Flexibility: Designed to be deployed on both fixed surveillance systems and mobile platforms, the
system can be scaled across urban and rural settings, enhancing its reach and adaptability.
 Public Safety Enhancement: By acting as a deterrent to non-compliance and helping authorities take swift action, these
technologies contribute to a culture of road safety and responsibility.

These AI-powered solutions showcase how technology can be leveraged not just for convenience, but also for societal benefits,
creating safer roads and more efficient administrative processes.

IJRAR1601009 International Journal of Research and Analytical Reviews (IJRAR) www.ijrar.org 2

© 2025 IJRAR APR 2025, Volume 5, Issue www.ijrar.org (E-ISSN 2348-1269, P- ISSN 2349-5138)
3
V. METHODOLOGY

IN OUR PROJECT, WE TRAINED TWO YOLOV5 MODELS USING TRANSFER LEARNING—ONE FOR DETECTING HELMET AND
PERSON, AND THE OTHER FOR MOTORCYCLE AND NUMBER PLATE. AS SHOWN IN FIG. 1, FRAMES FROM THE CAMERA FEED ARE
SAVED AND PASSED THROUGH THE FIRST MODEL. THE RESULTS ARE THEN PROCESSED BY THE SECOND MODEL. BOUNDING BOX
COORDINATES FROM BOTH MODELS ARE USED TO CHECK OVERLAPS BETWEEN ALL FOUR CLASSES FOR ACCURATE HELMET
VIOLATION DETECTION.

Fig.1. Flowchart for detecting person not wearing a helmet

In Method 1, we used YOLOv5 to detect four classes: helmet, person, motorcycle, and number plate. Due to GPU limits, we set
the batch size to 8 and used a low learning rate of 0.001 for better accuracy. We found that training with two classes at a time gave
higher mAP than training all four together. So, we trained two separate models: one for helmet and person (mAP: 0.945) and
another for motorcycle and number plate (mAP: 0.943), both using YOLOv5s weights. After detection, YOLOv5 outputs
bounding boxes in YOLO format. For easier processing, we convert them to PASCAL VOC format.

Fig 2. Flowchart for detecting if a person is wearing a helmet (Method 1)

After getting all bounding box coordinates, we identify bike riders without helmets using the overlap method. First, we
check if the Person and Helmet boxes overlap—if yes, the rider is wearing a helmet, so we skip them (Figure 3.a). If there's no
overlap, the person might be a pedestrian or a rider without a helmet. To confirm, we check if the Person and Motorcycle boxes
overlap. If not, it's just a pedestrian. If they do overlap, it’s a biker without a helmet (Figure 3.b). Finally, we check overlap
between Motorcycle and Number Plate boxes to extract their number plate (Figure 3.c). This way, we detect helmetless riders and
get their number plates.

Fig 3.a Checking Overlap

between helmet and person class
Method 1
IJRAR1601009 International Journal of Research and Analytical Reviews (IJRAR) www.ijrar.org 3
© 2025 IJRAR APR 2025, Volume 5, Issue www.ijrar.org (E-ISSN 2348-1269, P- ISSN 2349-5138)
3

Fig 3.b Checking Overlap Fig 3.c Checking Overlap between

between person and motorcycle the motorcycle and number plate-
Method 1 Method 1

In Method 2, we used only 3 classes: helmet, motorcycle, and number plate, and achieved an mAP of 0.847. To check if a
biker is wearing a helmet, we focus on the motorcycle's bounding box. We define a specific range based on its height and width. If
the helmet’s x-coordinates fall within the motorcycle’s x-range, and its y-coordinates are around 40% above the motorcycle (as
seen in our dataset), we consider the helmet to be present (Figure 3.d). If the helmet falls in this zone, the system outputs ―Biker
is wearing a helmet.‖ If not, it shows ―Biker is not wearing a helmet,‖ and we then extract the number plate by checking overlap
between the motorcycle and number plate boxes.

Fig 3.d Checking if the helmet is

present 40% above the height of
the motorcycle - Method 2

VI. CONCLUSION

We used the YOLOv5 algorithm, known for its high accuracy in real-time object detection. Our system works in two steps:
first, it checks if a motorcyclist is wearing a helmet; if not, it then detects and reads the number plate. The number plate characters
are converted into text and automatically saved in a record sheet for traffic law violations. We achieved an impressive mAP of
0.995, which highlights the accuracy and reliability of our model. What makes our project stand out is its ability to handle
overlapping objects effectively. This system is fully automated, helping authorities easily identify and penalize those not wearing
helmets.

IJRAR1601009 International Journal of Research and Analytical Reviews (IJRAR) www.ijrar.org 4

© 2025 IJRAR APR 2025, Volume 5, Issue www.ijrar.org (E-ISSN 2348-1269, P- ISSN 2349-5138)
3
VII. REFERENCES

Android Application Development Available at: Android Developer

 J. Chiverton – Motorcycle and helmet detection using tracking (IET, 2012).

 Z. Chen et al. – Vehicle detection and tracking in traffic (IEEE ITS, 2012).

 C. Patel et al. – Automatic number plate recognition (IJCA, 2013).

 S. Ren et al. – Faster R-CNN: Object detection with deep learning (2015).

 K. Dahiya et al. – Real-time helmet violation detection using videos (IJCNN, 2016).

 F. Wu, G. Jin – Helmet detection using improved YOLOv3 (IEEE ICNSC, 2019).

 F. Siebert, H. Lin – Helmet detection with CNNs (Germany, 2019).

 Y. Jonten et al. – Real-time number plate and helmet detection using YOLO (2020).

IEEE 5th International Conference on Digital Object Identifier: 10.1109/ANTS.2011.636865.

Publication Year: 2011.

N.K. Mishra
"FAO/AFMA/Myanmar: Enhancing Agricultural Marketing," Journal of Agricultural Marketing Information Systems, 2003,
Volume 15, Issue 4, pp. 2-4.

Yan Bo and Bu Yibi

"Agricultural Marketing Systems in China," Journal of Agricultural Marketing Information Systems, 2003, Volume 15, Issue 4,
pp. 33-37.

IJRAR1601009 International Journal of Research and Analytical Reviews (IJRAR) www.ijrar.org 5