A VISION-BASED SYSTEM DESIGN AND IMPLEMENTATION FOR ACCIDENT DETECTION AND ANALYSIS VIA TRAFFIC SURVEILLANCE VIDEO

Abstract

This study aims to test the entire framework on an AI demo board and investigate the issue of automatically and effectively recognizing and analyzing traffic incidents captured by surveillance cameras. To begin, the motion interaction field (MIF) method, which is capable of detecting collisions in video, is used to locate damaged automobiles based on the interactions of various moving objects. Second, the location of the destroyed vehicles is determined using the YOLO v3 model. Using a hierarchical clustering method, the vehicle trajectories prior to the collision are recovered, and the related trajectories are reconstructed. Finally, a perspective transformation is used to project the trajectory into a vertical view to help traffic cops make better decisions. The unbiased finite impulse response (UFIR) method, which does not require statistical information about the external noise, is used to estimate the vehicle's velocity. The estimated velocity and impact angle from the vertical view can then be used to investigate the traffic accident. Finally, a Huawei AI demo board called HiKey970, which was used to code all of the aforementioned algorithms, is used in an experiment to show how useful and effective the proposed method is in practice. A few mishap observation recordings are sent onto the demo board. Mishaps are recognized and the proper vehicle directions are gathered.