【摘要】 针对单流纯Transformer跟踪算法搜索区域中的相似信息或混乱背景等非目标信息的干扰会影响相关性计算的问题，提出一种抑制非目标干扰的单流纯Transformer跟踪算法。首先，构建抑制非目标干扰模块，该模块采用高相似token合并策略，当高相似token包含目标信息时，合并操作将保留目标信息，当高相似token包含混乱背景或相似目标干扰信息时，合并操作将降低这些干扰信息的注意力权重；其次，将该模块添加到单流纯Transformer骨干网络中，以抑制干扰多头注意力的计算结果；最后，将抑制干扰后的特征送进跟踪头，从而完成对目标的跟踪。在5个基准数据集上的测试结果表明：与OSTrack(One Stream Tracking)算法相比，在GOT-10k基准数据集AO指标提升1.1个百分点，在NFS、UAV123、TNL2K基准数据集AUC指标分别提升1.6、1.0、1.1个百分点，同时所提算法的跟踪推理速度即每秒帧数（FPS）可达166，证明所提算法成功抑制了非目标的干扰，提升了单流纯Transformer跟踪算法的鲁棒性并且能够保证跟踪的实时性。更多还原

【Abstract】 To address the interference problem of non-target information, such as similar information or chaotic background, in the search area of one-stream fully Transformer tracking algorithm affecting the correlation calculation, a one-stream fully Transformer tracking algorithm that suppresses non-target interference was proposed. Firstly, a non-target interference suppression module was constructed. In this module, a high-similarity token merging strategy was adopted. When the high-similar token contained target information, the target information was retained by the merging operation. When the high-similar token contained chaotic background or similar target interference information, the attention weight of this interference information was reduced by the merging operation. Secondly, this module was added to the one-stream fully Transformer backbone network to suppress the calculation results of the interference multi-head attention. Finally, the interference-suppressed features were sent to the tracking head to complete the tracking of target. Test results show that compared with OSTrack(One Stream Tracking) algorithm, the proposed algorithm has the AO(Average Overlap) improved by 1.1 percentage points on GOT-10 dataset, the AUC(Area Under the success rate Curve) by 1.6, 1.0, and 1.1 percentage points on NFS, UAV123, TNL2K datasets, respectively. At the same time, the tracking inference speed — Frames Per Second(FPS) of the proposed algorithm reaches 166, proving that this algorithm suppresses non-target interference successfully, improves the robustness of one-stream fully Transformer tracking algorithm, and ensures real-time tracking.更多还原