Car-mounted (black box) camera-based prediction and avoidance of intersection collisions for advanced driver assistance systems

Abstract

This study analyzed video and quantitative data of 471 four-way intersection vehicle collisions obtained from Virginia Tech Transportation Institute near-accidents data and used the analysis results to determine the threshold value for each of the nine types of intersection collisions. The collision cases obtained for this study were categorized into nine groups based on the direction of the car that recorded the video and location of the other car estimated through video analysis. In obscure cases, the aspect rate was additionally used to assign a group. After the group it belongs to is identified, the change rate of aspect ratio and area change rate were used to determine the possibility and specific type of intersection collision. When a collision was imminent, avoidance possibility was calculated to avoid the collision completely, and if the collision was inevitable, partial collision maneuver method that causes the least damage was deduced. The suggested algorithms were verified using the black box video from 16 actual accident cases. With the exception of special cases such as when most of the vehicle was out of view, most of them showed high correspondence.