Research On Related Technologies Of Trajectory Analysis In Video Multi-target Tracking

Video multi-target tracking,as a research hotspot in the field of computer vision,has good application prospects in areas such as video surveillance and autonomous driving.Due to occlusion,lighting,camera movement,etc.,the trajectory data obtained by the tracking algorithm may be wrong.Researchers need to analyze these erroneous trajectory data to improve the tracking algorithm,and finally compare the final trajectory with the ground truth to evaluate the performance of the algorithm.However,there is currently no relatively complete trajectory analysis technology system.Therefore,it is necessary and urgent to study the technology related to trajectory analysis in video multi-target tracking.The research contents of this thesis are as follows:1)Trajectory visualization analysis technology in a single perspective.First,the multi-angle display function of the trajectory is designed(conventional display,zoom display,target-to-original frame jump,dynamic display,forward and backward prediction),this solves the problem that the spatial attributes,appearance attributes,and temporal attributes of the trajectory data are obscure and invisible;Then,the visual editing(cut,paste,delete)and save functions of the trajectory are designed.These two functions can save the trajectory segments that are of interest to the user for targeted analysis;Finally,the function of loading additional information of the trajectory data is designed,which can load some intermediate data of the tracking algorithm into the output window of the software,such as the appearance characteristic value of the target and the similarity measure between trajectories.2)Comparative analysis of trajectories in a single perspective.A mathematical model of trajectory pairing is established,and a formula method for calculating the similarity of trajectory sets is defined to assist in comparing the paired trajectories.3)Coupling visual analysis technology from multiple perspectives.The segmentation marking function of the trajectory segments in the coupling is designed to facilitate the observation and analysis of the composition of the coupling,and a formula method for calculating the coupling quality is defined to assist the screening of lowquality couplings for targeted visual analysis.4)Trajectory/coupling fast loading technology.A multi-threaded technology was used to design a fast loading model for the trajectory visualization analysis in a single perspective,the comparative trajectory analysis in a single perspective,and the coupled visual analysis in multiple perspectives through the producer-consumer model.The test results show that the fast loading model largely solves the problem of inefficient conversion of a large number of trajectory/coupling raw data into trajectory/coupling images.5)Trajectory efficient labeling technology.An efficient trajectory labeling model is designed,which can quickly obtain reference standards similar to ground truth,so as to provide comparison objects for tracking results.The test results show that the labeling model improves the labeling efficiency while ensuring the labeling effect.The labeling time consumed by this labeling model is about 12% of pure manual labeling and 50% of CVAT,and the overall similarity between the labeled trajectory data and purely manually labeled trajectory data is about 98%.6)Trajectory analysis software for video multi-target tracking.This thesis uses C ++ to write code,combined with the QT framework to implement the above technology in application software.