Efficient traffic surveillance: Automatic accident detection through object tracking and trajectory conflict analysis

This work introduces an efficient framework for automatic traffic accident detection at intersections using surveillance systems and computer vision techniques. The framework comprises three hierarchical steps: object detection utilizing the state-of-the-art YOLOv4 method, object tracking based on a Kalman filter coupled with the Hungarian algorithm for accurate association, and accident detection through trajectory conflict analysis. A new cost function is implemented to manage challenges like occlusion, overlapping objects, and shape changes during tracking. Object trajectories are analyzed in terms of speed, angle, and distance to detect various trajectory conflicts involving vehicles, pedestrians, and bicycles. Experimental results using real traffic video data demonstrate the feasibility of the method for real-time traffic surveillance applications, with a low false alarm rate and high detection accuracy in identifying near-accidents and collisions at urban intersections, even under varying lighting conditions.