Real-Time Traffic Accident Detection via Multi-Criteria Trajectory Interaction Analysis with YOLOv11 and DeepSORT
Abstract
Automatic traffic accident detection in surveillance videos remains a challenging task for intelligent transportation systems, particularly in dense urban environments where severe occlusion and complex vehicle interactions frequently occur. This study proposes a hybrid framework that integrates deep learning based vehicle detection with trajectory interaction analysis to improve the reliability of accident identification. Vehicle detection is performed using YOLOv8 and YOLOv11 models, while DeepSORT and ByteTrack are employed to maintain temporal identity consistency across consecutive frames. A multi parameter interaction mechanism is applied to identify potential collision events by analyzing spatial proximity, motion variation, and directional conflicts over time. To reduce false detections, candidate collision events are further verified through an image level accident classification stage. Experimental results show that the YOLOv11 DeepSORT configuration provides the most stable tracking performance, achieving a MOTA of 0.89 and an IDF1 score of 0.91. The integrated system achieves an event level precision of 0.90 and an F1 score of 0.85 on real world accident videos, demonstrating its potential for supporting automated traffic monitoring and accident analysis in intelligent transportation systems.