GATransTracker: A Robust and Accurate Feature Fusion for Object Tracking
Jing Wen, RUOZHOU WANG · 2024
The similarity of features between the template and the search sequences is a crucial step in object tracking. Some recent research provide the inspiration of feature fusion with the cross-attention mechanism to compute the correlation between the features in two branches. However, the direct computation of cross-attention, according to our research, usually reduces the robustness and accuracy of the tracking, since the dynamic environment during tracking is highly related and the strong attention response from background would cause the drift of tracking. To address this issue, we proposed a gated-attention transformer tracking (GATransTracker) method, which is able to suppress the redundant noise from background during the process of computing correlation by a novel gated attention feature (GAF) module. In GAF module, an attention gated unit (AGU) is designed to supervise the cross-attention information with the self-cross attention, so as to suppress the interference from background noise,and ensure the robustness of the tracker. The experiments illustrates that our method could reach up to 67.9% AO, 77.2% SR0.5 and 60.9% SR0.75 on the GOT-10k dataset, outperforming most of mainstream tracking algorithms in all metrics quantitatively.Remarkably,the results of our tracker is more consistent with human visual perception. And our tracker can run at 36.8FPS on GPU, meeting the requirements of real-time tracking.