Multi-object Tracking Research Based On Transformer

Multi-target tracking tasks,generally understood as inferences about the location and identity of multiple people simultaneously in a visual scene recorded by one or more cameras,have become a central research problem in the field of computer vision over the past few years.Multi-target tracking is mainly used in transportation,national defense construction,military control,automatic driving and so on.In recent research,Transformer architecture for natural language processing has been applied to multi-target tracking with good results.Transformer,as a special query-key mechanism,relies heavily on the attention mechanism to process extracted features.However,the query target is continuously updated in the tracking process and gradually polluted by the background information.The inaccurate target information will lead to the drift of the target box and the loss of the target.In the whole tracking process,the apparent features and motion features of the target are ignored,and the problem of identity continuity and target recurrence of the blocked target cannot be well dealt with.Therefore,the multi-target tracking algorithm proposed in this paper adopts Track Former algorithm based on Transformer structure to jointly detect and track items,deal with the identity problem of blocked targets,track and predict the blocked targets by using the motion features of the targets,and retrieve and track the lost targets by matching the apparent features.To improve the performance of the tracking algorithm.The main work content of this paper mainly includes the following two aspects:(1)A tracking method based on Track Former with dual decoder is proposed to deal with the problem of target loss caused by occlusion in multi-target tracking.One decoder is used for detection and the other for tracking.The detected target information is used to update the tracked target information,so that the blocked target information is not polluted and can be continuously tracked.Experimental results show that Track Former with dual decoder has better performance in continuous tracking under the condition of target occlusion.(2)A tracking method of track prediction and histogram template matching based on Kalman filter is proposed.The Transformer structure ignores the apparent features and motion features of the target,and when the lost target reappears,it is not good to re-identify the identity information based on these features.Kalman filter is used to predict the trajectory of the target,and the histogram is used to match the identity information of the target to achieve the purpose of continuous tracking.Experimental results show that the tracking performance of the improved method is improved greatly.