LAUSR

Road accidents impose serious problems on society. Possible collisions between vehicles and pedestrians must be detected before they occur so that a timely warning may be issued. By using the vision-based approach, this study presents an effective and efficient algorithm to estimate the vehicle-pedestrian collision probability at intersections. The real-time trajectories and movement parameters (position, speed, acceleration or direction) of vehicles and pedestrians are obtained based on state-of-the-art detection and tracking algorithm which include background subtraction method, faster regions with convolutional neural networks and optical flow method. To find the appropriate time to identify the latent collision risk for calculating the collision probability, this study defines the critical time based on different collision patterns of perception-reaction failure and evasive action failure. In addition, based on discrete acceleration and discrete angle, the authors get different extended trajectories which can include most situation when the conflict happened. Trajectories generation probability are given by the discrete choice probability model based on the Logit model to get the accurate collision probability. Real-world video data is implemented to demonstrate the approach. This proposed collision prediction method can provide some important results for designing the intelligent pedestrian signal timing schemes at intersections.