Joint Fringe Feature Extraction And Spatio-Temporal Constrained Trajectory Matching For Cross-View Vehicle Tracking

Video multi-target tracking technology can continuously track the state of multiple targets in the video sequence and generate the trajectory of the target.this technology has an important application prospect in the fields of intelligent video surveillance,robot vision,intelligent transportation and so on.In recent years,in order to extract the multi-target trajectory across the view,cross-view video multi-target tracking has gradually become an important research direction.However,due to the significant change of target characteristics caused by the change of visual angle,there are still many challenges to achieve effective cross-view multi-target tracking.This thesis focuses on three subtasks: single-camera multi-target tracking,target appearance feature extraction and cross-view trajectory matching in the cross-view tracking process.On the basis of combing the correlation between each sub-task and the existing difficult problems,this thesis focuses on the vehicle appearance feature extraction based on joint fringes and the trajectory correlation method considering space-time constraints.The main contents of this thesis are as follows:(1)A vehicle re-identification method based on joint fringe relation is proposed.In order to solve the problem of the loss of spatial information based on feature map segmentation and re-identification,this thesis constructs a relationship description module to make up for the loss of spatial information caused by feature map segmentation.the vehicle re-identification of joint stripe relationship is realized by combining multi-activation value module and batch normalization module.This method can extract robust appearance features for vehicles from different perspectives,which lays a foundation for cross-view vehicle trajectory matching.(2)A cross-view trajectory matching method based on trajectory spatio-temporal feature extraction is proposed.In view of a large number of identity switching problems due to long-term occlusion in the multi-target tracking algorithm based on YOLOv5 and Deep SORT,this thesis uses the vehicle re-identification algorithm with joint fringe relationship to extract the depth appearance features of the target,and integrates the appearance features into the calculation of similarity between vehicles to achieve multitarget tracking under a single camera.This method can maintain the consistency of the target identity when facing the target which is obscured for a long time,reduce the number of target identity switching,and provide accurate trajectory information for cross-view trajectory matching.In order to solve the problem of single semantic information of target trajectory feature extraction based on deep learning,this thesis constructs a trajectory spatio-temporal feature extraction model to extract the spatial and temporal features of the target track.according to the topological structure between the cameras and the view space in the camera,the target transfer rule is constructed,and the trajectory spatio-temporal feature extraction model is combined to realize the crossview matching of the target trajectory.This method uses the spatio-temporal feature extraction model to characterize the trajectory,takes into account both the spatial information and time information of the trajectory,and the establishment of spatiotemporal constraints between trajectory matching reduces a large number of unnecessary calculations and improves the matching accuracy.