Adaptive update of UAV multi-target tracking based on Transformer
Haiyun Shen, Zhongyi Huang, Haichuan Wang, Honghao Yu · 2023
Aiming at the problem of the drift of the tracking target frame caused by the size change, low resolution, target occlusion and other scenarios when the unmanned aerial vehicle (UAV) performs the tracking task, a multi-target tracking algorithm DUTrack based on the Transformer adaptive update is proposed. First, the FairMot algorithm is used as the baseline algorithm, secondly, the Transformer is used to adaptively extract the spatiotemporal information of historical frames and update the template online, so as to store the spatiotemporal context information in the dynamic template, then, the benchmark template and dynamic template are used respectively. Cross-correlation operation is performed with the search feature map to obtain two response maps; finally, the two response maps are fused by Transformer to establish a spatiotemporal context mapping relationship between consecutive frames. The experimental results on the VisDrone2019 dataset show that the proposed DUTrack achieves better tracking accuracy and robustness on UAV video tracking methods.