A Deep Learning-Based Traffic Violation Detection System with Conditional Number Plate Recognition for Helmet Compliance Monitoring

Abstract

One of the issues of concern is the enforcement of traffic safety because of the rising rate of accidents related to helmet non-compliance among the riders of the two wheelers with the systems of manual monitoring being inefficient and prone to errors. To overcome this problem, an automated system based on deep learning is suggested to observe the violation of the helmet and the intelligent monitoring of the traffic. The system uses a YOLOv8 object detection model to detect riders, use of helmet and number plates on traffic images in addition to conditional Optical Character Recognition (OCR) mechanism such that number plate information is only extracted when one of the violations (absence of helmet) are identified thus lowering computational cost and enhancing efficiency. The model is trained on a dataset consisting of four classes, with helmet, without helmet, rider and number plate, and fared off with 0.89 precisions, 0.885 recalls and 0.913 mAP at 0.5 which is a good performance. In the case of the OCR module, the accuracy of exactly matching is 33% and the accuracy of characters at the level of 78% is achieved, which indicates the practical issues like motion blur and low-resolution images in the real world. The system also records details of violations, such as rider present, helmet status and extracted number plate into an organized format, which facilitates its use in computerized traffic enforcing systems.