Research On Multi-Target Tracking Algorithm Based On Deep Learning

In the field of computer vision,multi-target tracking algorithms have always been a hot spot for researchers and scientific research institutions.In the era of big data,how to efficiently analyze and use these data and dig out valuable information from them is a difficult problem to be solved urgently.The multi-target tracking algorithm can determine the position of all the interested target in each video frame of the video sequence,and ensure that the identity information of the same object does not change in the continuous video frame,and finally obtain the movement trajectory of all the interested target.Therefore,in the field of computer vision,the research on multi-target tracking algorithm is of very high value,but the distance to completely use machines to replace humans to recognize and perceive the world still needs to make more efforts.In this paper,the problems and main challenges in the current multi-target tracking algorithm have been studied comprehensively and deeply,and some research achievements have been made.The specific content is shown below:(1)In view of the problem that the existing data association algorithm is too dependent on the performance of the target detector,and when the single target tracking algorithm is directly extended to the multi-target tracking task,the performance of the target tracking algorithm will drop sharply as the number of tracked targets increases,this paper proposes a joint data association algorithm of single target tracking.Data interaction can be carried out between the data association module and the single target tracking module,and the two modules can share the output results of the feature convolutional network.In the single target tracking module,it is divided into two stages: correlation tracking and position adjustment.It can achieve the feature expression of the target object detection result in the current video frame and all target objects in a forward propagation process of the network model.(2)The current multi-target tracking algorithms often ignore the connection between target detection and target tracking,and regard them as two mutually independent modules,so they cannot effectively use the data between them to interact.This paper proposes a multi-target tracking algorithm based on joint target detection.The network model is trained offline through multi-task learning.The target detection module and the target tracking module can share the output results of the feature convolutional network,thus greatly reducing the parameters and calculation amount in the network model.In the process of target tracking,the tracking results can be used to fine-tune the detection results of the target detector online,which can not only improve the performance of the target detector in the current video sequence,but also further improve the performance of the target tracking.