Title:COLLIDE-PRED: Prediction of On-Road Collision From Surveillance Videos

Abstract:Predicting on-road abnormalities such as road accidents or traffic violations is a challenging task in traffic surveillance. If such predictions can be done in advance, many damages can be controlled. Here in our wok, we tried to formulate a solution for automated collision prediction in traffic surveillance videos with computer vision and deep networks. It involves object detection, tracking, trajectory estimation, and collision prediction. We propose an end-to-end collision prediction system, named as COLLIDE-PRED, that intelligently integrates the information of past and future trajectories of moving objects to predict collisions in videos. It is a pipeline that starts with object detection, which is used for object tracking, and then trajectory prediction is performed which concludes by collision detection. The probable place of collision, and the objects those may cause the collision, both can be identified correctly with COLLIDE-PRED. The proposed method is experimentally validated with a number of different videos and proves to be effective in identifying accident in advance.