Siamese Network Tracking Algorithm Based on Multi-Scale Feature Fusion of Dual Attention Mechanism

Siamese network trackers have shown strong performance in the field of computer vision. However, relying solely on feature representations from the backbone network's final layer can cause tracking drift in complex scenarios. To address this issue, we propose a Siamese network tracking algorithm based on dual attention mechanisms and multi-scale feature fusion. This method leverages features from different layers of the backbone network, integrating deep semantic information with shallow spatial information through two multi-input attention mechanisms. This approach enhances the distinction between the target and the background, thereby enhancing the tracker's discriminative ability. Furthermore, a novel cascade of location prediction head is designed to accurately refine target localization. Experimental results on the VOT2019 and OTB2015 datasets indicate that the proposed method obtains satisfactory performance. On the VOT2019 dataset, the method achieved an accuracy of 60.7<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>, an expected average overlap (EAO) of 34.1%, and a 14.20% improvement in robustness compared to the benchmark model, with a running speed of 48.2 FPS, surpassing the real-time threshold of 25 FPS.