SiamDTA: Dual-Template Siamese Network Visual Object Tracking Algorithm Based on Attention Mechanism

Fully convolutional Siamese network (SiamFC) for object tracking has shown great performance. However, there are still some inefficiencies in deep information modeling and model updating. To address the above problems, we propose a dual-template Siamese network visual object tracking algorithm based on attention mechanism, named SiamDTA. Firstly, this paper cascades the global context attention module and coordinate attention module after the backbone, which enhances the ability of the deep features extracted by the Siamese network. Secondly, the initial template is retained in the tracking process, and the second template is selected dynamically by evaluating the tracking response map. The stability of tracking is improved by using the joint response map of two templates to locate. Compared with the baseline SiamFC, the success rate and accuracy on the OTB100 dataset reach 0.632 and 0.853, which are improved respectively by 4.5% and 8.1% On the VOT2018 dataset, the tracking accuracy, robustness, and EAO are improved respectively by 2%, 22.4% and 8.1%. Compared with other mainstream algorithms, the proposed algorithm achieves optimal tracking.