Abstract:

The goal of this study is to examine the issue of automatically and effectively identifying and analysing traffic accidents using surveillance cameras, and to implement the whole framework on an AI demo board. First, the motion interaction field (MIF) approach, which has the capacity to identify collisions in video, is used to find the wrecked cars based on the interactions of various moving objects. Second, the YOLO v3 model is used to pinpoint the position of the wrecked automobiles. A hierarchical clustering technique is employed to recover the vehicle trajectories prior to the collision, and the associated trajectories are recovered. Third, the trajectory is projected to a vertical view using a perspective transformation to aid traffic officers' judgement. The unbiased finite impulse response (UFIR) technique is used to estimate the vehicle velocity, which does not need statistical knowledge of the external noise. The estimated velocity and collision angle acquired from the vertical view may then be used to investigate the traffic accident. Finally, to demonstrate the usefulness and implementation performance of the suggested technique, an experiment is carried out using a Huawei AI demo board dubbed HiKey970, which is utilised to code all of the aforementioned algorithms. Several accident surveillance movies serve as the demo board's input. Accidents are effectively recognised, and the relevant vehicle trajectories are retrieved

Description:

Accident detection, speed estimation, target tracking, unbiased finite impulse response (UFIR) filter, vehicles.