Research On The Prediction Method Of Pedestrian Crossing Behavior Based On Surveillance Video

Most of the current research on pedestrian crossing behavior prediction base on vehicle video.However,vehicle video has a low viewing angle and is easily occluded by factors such as vehicles.With the development of Internet of Vehicles technology,autonomous driving based on vehicle-road collaboration has become a research trend,and the autonomous driving of vehicles assisted by edge computing will play an important role in the future.Compared with vehicle video,surveillance video has a wider field of vision and can more fully display road element information.This paper mainly studies the prediction method of pedestrian crossing behavior based on surveillance video.The research content mainly includes three aspects: vehicle pedestrian detection,vehicle pedestrian tracking and pedestrian crossing behavior prediction.(1)An improved vehicle and pedestrian detection network based on YOLOv3 is proposed.Aiming at the difficulty of target detection in the case of small target and occlusion,YOLOv3 is improved by adding detection layer,introducing attention mechanism,introducing spatial pyramid structure and improving Io U.Finally,experiments are performed on the Vis Drone2021 target detection dataset.The results show that the proposed algorithm can reduce the missed detection and false detection of the original YOLOv3 algorithm and improve the detection performance,especially for small or occlusion targets.(2)A vehicle-pedestrian tracking algorithm based on Deep SORT is proposed.According to the difference of detection and tracking difficulty between vehicles and pedestrians,different matching and tracking methods are adopted for vehicles and pedestrians.For vehicle target,only the motion features in the SORT method are used to perform target matching between two frames.For pedestrian targets,the combination of appearance features and motion features is used to perform target matching between two frames.The algorithm proposed in this paper achieves better tracking effect on BDD video data.(3)On the basis of BDD surveillance video data,the location of vehicles and pedestrians and the pedestrian crossing behavior of target pedestrians are marked,and a pedestrian crossing behavior prediction dataset based on surveillance video is produced.A multi-source feature fusion method is proposed,which integrates pedestrian pose features,vehicle and pedestrian trajectories,pedestrian surrounding environment features and global environment features in a hierarchical stacking manner to predict pedestrian crossing behavior.Compared with several baseline methods,the fusion method proposed in this paper has achieved the best results in the evaluation indicators such as accuracy and F1 parameters.In addition,this paper conducts further comparative experiments under different prediction time and pedestrian speed to study the impact of prediction time and pedestrian speed on the prediction performance of each model.