An Automated Traffic Violation Detection System Based on YOLOv5

Traffic violations are a major cause of road accidents, especially in urban areas with limited law enforcement. This study proposes an automated system to detect three common violations: red-light running, wrong-way driving, and helmet noncompliance. The system combines the YOLOv 5 object detection model with Deep SORT tracking to ensure consistent object identification and utilizes the ONNX format to optimize runtime performance for efficient inference. Trained on a custom dataset of 250 annotated images, the system was evaluated on real-world CCTV footage under both daytime and nighttime conditions. The evaluation revealed excellent performance for red-light violation detection, achieving a flawless 1.0 accuracy. Helmet violation detection showed varied results, with accuracy improving from 0.67 in daylight to 0.82 at night. While wrong-way driving detection was highly accurate with 1.0 accuracy during daytime and 0.96 at nighttime, the system was susceptible to false positives in low-light conditions. These findings demonstrate the system effectiveness as a practical solution for traffic monitoring. However, the current implementation relies on manual video file input. Therefore, a primary direction for future work is the realtime implementation of the system directly on live traffic surveillance camera.