��IEEE International Conference on Robotics, �Automation, Artificial-Intelligence and �Internet-of-Things 2025

Paper ID: 49 || Session: TS - 1 (AIML-1)

A Real-Time Automated Traffic Violations Detection System for Motorcycles Using YOLOR

​

Authored by:

Sadab Khan Prangon, Rabab Khan Rongon, K Das, F. I Badhon, M.R Momin, A.R Shakur

Presented By:

Sadab Khan Prangon

B.Sc. in Computer Science and Engineering

IUBAT - International University of Business Agriculture and Technology

Dhaka, Bangladesh

Introduction

• Motorcycle accidents are increasing in Bangladesh because many riders do not wear helmets and some carry more than two people on a motorcycle.
• Manual traffic monitoring is slow, tiring for officers, and often misses violations even when CCTV cameras are used.
• This study builds a real-time system that uses YOLOR to detect helmet violations, triple riding, and OCR to read Bangla license plates to help improve traffic safety.

​

2

Problem Statement

• Manual traffic monitoring often misses violations because it is slow, tiring, and depends on human attention.
• Existing detection systems struggle in low light, crowded scenes, and cannot reliably read Bangla license plates.
• Most previous studies do not detect triple riding and do not offer a complete real-time solution for motorcycle violations in Bangladesh.

​

3

Related Work

Helmet Detection Studies : Afzal et al. (2021)

• Contribution: Earlier research used HOG, SVM, LBP and CNN models to detect helmets in images and videos. These methods showed good detection accuracy in simple environments.
• Limitation: These models were slow, had trouble in low light, and could not handle crowded traffic or rider counting.
• Difference: Our system uses YOLOR for faster detection with better accuracy in complex urban traffic.

​

4

Related Work (Continued)

YOLO Based Approaches: Jamtsho et al. (2021)

• Contribution: YOLO models have been used for helmet detection and license plate detection because they can work in real time.
• Limitation: Older YOLO versions struggled in dim light, cluttered scenes and were not tested well in South Asian city traffic.
• Difference: We use YOLOR, which gives better accuracy and speed, and we evaluate it using real data from Dhaka.

​

5

Related Work (Continued)

Bangla License Plate Recognition: Haque et al. (2022)

• Contribution: Prior works used CNN and OCR tools to detect and read Bangla license plates. Some achieved good results on clean images.
• Limitation: Many systems failed in real situations such as motion blur, plate angle changes and low light.
• Difference: Our system uses Bengali OCR with preprocessing to improve reading accuracy in real traffic conditions.

​

6

Objectives

• To build a real-time system that can detect helmetless riders, triple riding and motorcycle violations using deep learning.
• To improve Bangla license plate reading by using OCR with better preprocessing for real traffic conditions.
• To create a low cost and scalable solution that can support traffic police and reduce the need for manual monitoring in Bangladesh.

​

7

Methodology

8

Fig 1. Flowchart of the Proposed Methodology

Dataset

• Dataset Setup:

• Locations: Rampura and Khilgaon areas in Dhaka�• Image conditions: Different lighting and weather situations

• Dataset size:�• Helmet and rider detection: 764 images�• License plate detection: 685 images

9

• Data split:�• Training: 70%�• Validation: 15%�• Testing: 15%

​

• Augmentations:�• Image flipping�• Brightness and contrast changes�• Noise and rotation adjustments

​

Model Architecture (Continued)

10

Training Setup:

• Input size: 416 × 416 pixels
• Batch size: 16
• Epochs: 100
• Optimizer: SGD with momentum 0.937
• Learning rate: 0.01 with cosine scheduling
• Loss functions:� - CIoU loss for bounding boxes� - Binary cross entropy for classification

​

Result (Detection Performance)

• Helmet detection reached 83.92% mAP.�• Triple riding detection has 92.6% accuracy.�• Rider and helmet detection improved steadily across training.

​

11

Fig 1. Helmet and Rider Detection Performance over Epochs

Table I: Tabulation of Average Precision And IoU Per

Class

​

Result (License Plate Detection and OCR Performance)

License Plate Detection:

• mAP reached 67.25%�• Accuracy improved across epochs�• Small plate size and occlusion make detection harder

OCR Performance:

• Bengali OCR achieved 85.3 % accuracy�• Preprocessing like resizing and contrast adjustment improved results

​

12

Fig 2. License Plate Detection and OCR Accuracy Trends

Result (Real Time Performance)

• Helmet detection, rider counting, plate detection and OCR all run with low inference time between 5 ms and 14 ms per task.

• The model maintains strong precision and recall across all tasks, as shown in the bar chart.

• The low latency of each detection step makes the system suitable for real-time use in surveillance video processing.

​

13

Fig 3. Precision, Recall & Inference Time

Result (Comparison and Robustness)

14

Table II — Comparison of YOLOR vs Other YOLO Models

Table III — Performance Under Noisy Conditions

Conclusion

• The system can successfully detect helmetless riders, triple riding and Bangla license plates with good accuracy, and it works in real time on normal hardware.
• OCR performance improves with preprocessing, but accuracy becomes lower in low light, rainy weather or when plates are blurred or tilted.
• License plate detection is harder than helmet detection because plates are small and often partly hidden, but the model still performs well for practical use.

15

Future Work

• Collect more data from night time, rainy weather and busy traffic to improve model robustness.
• Use advanced tracking methods such as Deep SORT or ByteTrack to follow motorcycles across video frames.
• Improve OCR performance for low quality or partly hidden Bangla license plates.
• Test the system in real deployments with traffic authorities and explore automated violation notification.

16

Reference

[1] Afzal, A. Afzal, H. U. Draz, M. Z. Khan, and M. U. G. Khan, "Automatic helmet violation detection of motorcyclists from surveillance videos using deep learning approaches of computer vision," in Proc. Int. Conf. Artif. Intell. (ICAI), 2021, pp. 61–66, doi: 10.1109/ICAI52203.2021.9445283.

​

[2] Y. Jamtsho, P. Riyamongkol, and R. Waranusast, "Real-time license plate detection for non-helmeted motorcyclist using YOLO," ICT Express, vol. 7, no. 2, pp. 166–171, 2021, doi: 10.1016/j.icte.2020.07.008.

​

[3] N. Haque, S. Islam, R. A. Tithy, and M. S. Uddin, “Automatic bangla license plate recognition system for low-resolution images,” in 2022 4th International Conference on Sustainable Technologies for Industry 4.0 (STI), pp. 1–6, IEEE, 2022

​

17

Thanks!

Any questions?

18