Traffic Accident Reconstruction Method Integrating Surveillance Video and Scene Point Cloud

In order to overcome the problems such as ignoring the lack of depth information in the process of perspective projection, or sensitive to surveillance video quality that the existing vehicle motion state solution methods based on video image, this paper presents a methodology for reconstructing traffic accident based on surveillance video and scene point cloud. Firstly, the 2D-3D corresponding points from surveillance video image and scene point cloud are used to estimate the camera spatial pose, and then the Camshift algorithm is used to track the vehicle features and obtain the sequence of vehicle feature pixels. Secondly, the vehicle feature spatial position analysis model is constructed to analysis vehicle feature spatial position sequence, next the vehicle trajectory information is obtained by polynomial function fitting, and the vehicle speed information is obtained by feature spatial position Euclidean distance. Finally, simulation vehicle experiments are carried out under the two driving routes of left turn and straight travel respectively. The experimental results show that the maximum relative error of running speed under the condition of constant speed left turn running is 8.3% and the average relative error is less than 3%. The maximum relative error of running speed under the condition of constant speed straight running is 8.7%, and the average relative error is within 5%, which proves the feasibility and accuracy of this method, and completes the real scene reproduction of the experimental process in the scene point cloud.