Object Detection and Near-Miss Analysis in Traffic Videos of High and Low Quality Based on the YOLO Model

Abstract

The development of intelligent transportation systems has made traffic detection technology crucial for preventing near-miss events and enhancing road safety. You Only Look Once (YOLO), a popular target detection model, is widely used in traffic target detection. However, the performance of different YOLO models varies significantly on the same quality video data. Furthermore, different video qualities affect near-miss event detection differently. This paper conducts experiments on different YOLO models using high-quality and low-quality video data to assess their detection performance. Considering the expressway scenario in this study, we use manual reporting (based on human detection), the bird’s-eye view method (based on image processing), and the distance neighbor method (based on data analysis) for near-miss detection. Results show that YOLOv5m has the highest precision, recall, and mAP@0.5 (0.990, 0.977, and 0.994) for high-quality video detection. YOLOv8m has the highest mAP@0.5 - 0.95 (0.908) for low-quality video detection. In high-quality video near-miss detection, manual reporting, the bird’s-eye view method, and the distance neighbor method show high consistency, indicating that the latter two can replace manual calculation in specific scenarios. In low-quality video near-miss detection, all three methods yield identical results, highlighting the significant impact of clarity and fluency on near-miss detection.