YOLOV4 Deepsort ANN for Traffic Collision Detection

Abstract

Every collision must be handled right away to prevent further harm, damage, and traffic bottlenecks. Hence, the implementation of a systematic approach for accident detection becomes imperative to expedite response mechanisms. Our proposed accident detection system operates in three stages, encompassing vehicle object detection, multiple object tracking, and vehicle interaction analysis. YOLOv4 is employed for object detection, while DeepSort is utilized to the tracking of multiple vehicle objects. Subsequently, the positional and interactional data of each object within the video frame undergo thorough analysis to identify collisions, utilizing an Artificial Neural Network (ANN). Notably, collisions involving a single vehicle and not affecting other road users are excluded from the scope of this study. The evaluation of our approach reveals that the ANN model achieves a commendable F-Measure of 0.97 for detecting objects without collisions and 0.88 for objects involved in collisions, based on the conducted tests.