- Research Article
- 10.26518/2071-7296-2025-22-6-916-927
Detection and classification of dangerous maneuvers based on traffic camera recordings
- Jan 12, 2026
- The Russian Automobile and Highway Industry Journal
- A N Novikov + 3 more +3
Introduction. Traffic accidents are a leading cause of death. In 2024, 19,940 people were killed in accidents in the EU, while in Russia this number amounts to 14,400 people. A significant proportion of accidents are associated with dangerous maneuvers: abrupt lane changes, overtaking, emergency braking, and running red lights. Traditional monitoring methods are limited by cost and scalability. The goal of this research is to develop a system for automatic detection and classification of dangerous maneuvers based on video data obtained with YOLOv8 and Deep SORT. Materials and Methods. A system consisting of four modules is proposed: a modified YOLOv8n (with a P2 layer, LW_C2f, Wise-IoU) for vehicle detection; an optimized Deep SORT for tracking; trajectory analysis with camera calibration; maneuver classification based on acceleration thresholds (0.35g for lane change, 0.30g for braking), lane marking crossing, and traffic light status (YOLOv8). The system was trained on 45,000 vehicle images and on 20,000 images for re-ID. Results . Testing of 150 hours of video (various conditions) has shown the following results: mAP vehicle detection – 92.7%, MOTA tracking – 86.3%, maneuver classification accuracy – 89.3% (F1: lane change – 89.4%, braking – 89.7%, red light – 85.2%) at 28 FPS on RTX 3070. Latency – 0.12 s. Discussion and conclusions. The system under investigation is superior to similar systems in speed and maneuver coverage, and is suitable for ITS applications. Limitations refer to decreased accuracy in foggy or rainy weather. Further research can lead to expanding the class range, edge computing, and risk prediction. The system implementation will reduce accidents and results in automated traffic enforcement.
Read more