Abstract: To solve the tracking failure problem of fully-convolutional siamese networks algorithm (SiamFC) in complex scenes such as objects deformation, occlusion, and fast motion, a novel method (SiamMA) that uses the mixed attention mechanism to enhance the network identification ability is proposed. Firstly, in order to simulate the complex scenes and enhances the generalization performance of networks, an image stacking and cropping method is adopted in the network training stage to build the self-adversarial training sample pairs. Secondly, a mixed attention mechanism algorithm is proposed, which fuses spatial attention and channel attention modules in different branches of the network, so the background interference in the feature map can effectively be suppressed and the robustness of the algorithm is improved. 4 open test datasets such as Got-10k and UAV123, etc., are adopted to evaluate the algorithm performance. The experimental results show that our method outperforms 6 traditional algorithms such as SiamFC, KCF, etc., on the main performance indexes such as tracking success rate and precision. The average speed of the algorithm can reach 60 frames per second.

Key words: object tracking, siamese network, mixed attention mechanism, self-adversarial training sample pairs