Research On Multiple Object Tracking Method Based On Feature Enhancement And Data Association

With the popularization of the Internet of Things and the development of 5G applications,more and more image data is generated and collected.In order to effectively process these data,researchers have conducted in-depth research and exploration in various sub-fields such as image classification,object detection,image segmentation and object tracking under computer vision.Multi-target tracking is to derive the trajectories of all objects of interest from consecutive image frames.It is widely used in autonomous driving,human-computer interaction and robot vision navigation,and has received more and more attention in recent years.Designing an accurate and efficient multi-target tracking algorithm is still a challenging task due to the randomness of the number of targets in multi-target tracking,the complexity of the tracking background and the occlusion between targets.In order to improve the accuracy of the multi-target tracking algorithm,this dissertation focuses on the feature extraction and data association steps in multi-target tracking and proposes a multi-target tracking algorithm based on feature enhancement and data association,which effectively improves the missed detection and false detection.And the wrong association problem caused by occlusion.The main research contents are as follows:(1)Aiming at the problem of missed detection and false detection of targets,a multi-target tracking algorithm based on motion salient features and cross-correlation attention is proposed.Firstly,the motion enhancement module is used to enhance the motion-related features in the channel and suppress irrelevant information by using the temporal difference operation of the adjacent frame features.Then a cross-correlated attention module is adopted to decouple the detection task and the re-identification task to reduce the competition between these two tasks.Finally,object detection,feature extraction and data association are unified into one framework to achieve end-to-end optimization.(2)Aiming at the problems of target trajectory interruption and trajectory ID transformation caused by occlusion,a multi-target tracking algorithm based on quadratic correlation low-scoring detection frame is proposed.By using the quadratic data association method based on the proposed way to use motion enhancement and cross-correlated attention.This method is different from the previous practice of only retaining the high-scoring detection frame,but processing the high-scoring frame and the low-scoring frame separately,using the similarity between the low-scoring detection frame and the trajectory to find the real target and realize the occlusion situation.correct association.Experiments were conducted on the MOT 16/17 dataset and MOT20 dataset provided on the MOT Challenge and 8 standard evaluation indicators(MOTA,MOTP,IDF1,MT,ML,FP,FN,IDs)were used to comprehensively measure the algorithm.The experimental results on the MOT 17 dataset show that the application of the motion enhancement module and the cross-correlation attention module improves MOTA from 66.6%to 67.5%and reduces FP by 12.1%.On this basis,the secondary data association is superimposed and the final MOTA is 68.7%.The experimental results show that the improved chain tracking algorithm increase the correlation between two adjacent frames,can effectively solve the occlusion problem and improve the accuracy of the multi-target tracking algorithm.