融合时序上下文线索的单流Transformer跟踪算法OA

The one-stream Transformer tracking framework has attracted much attention due to its strong global modeling capability.However,one-stream Transformer trackers generally suffer from insufficient utilization of temporal context information,which easily leads to tracking drift due to the loss of historical state memory.To address this issue,we propose a one-stream Transformer tracking al-gorithm that integrates temporal context clues to enhance the target representation capability in complex scenarios.Firstly,a temporal modeling module is designed to perform temporal modeling on historical frame sequences based on Transformer,extracting target motion trajectories and appearance evolution features.Secondly,a spatial decoding module is constructed to dynamically fuse time clues with target queries,generating enhanced spatiotemporal representations.Finally,target localization and state update are realized through the prediction head.The experiment was publicly compared on six benchmark datasets,GOT-10K,TrackingNet,LaSOT,NFS,UAV123,and TNL2K.In the TrackingNet dataset,the AUC score and accuracy were 83.5%and 82.4%,respectively,ranking first among all comparison methods;In the GOT-10K dataset,the average overlap rate is 72.6%,which is 1.2 percentage points higher than that of the performance second model DyTrack;In the LaSOT dataset,the AUC score is69.2%,which is comparable to the current state-of-the-art model DyTrack.The experimental results show that the proposed algorithm significantly improves the robustness and accuracy of the one-stream Transformer tracker by explicitly modeling temporal contextual information,verifying the effectiveness of temporal contextual clues.