Refined attention Siamese network for real-time object tracking

Abstract

The Siamese tracker shows great potential in achieving a balance between accuracy and speed, but the twin structure of the Siamese network makes the tracker vulnerable to background interference in the tracking scene. To deal with this problem, a tracking algorithm based on the attention mechanism is proposed. The algorithm introduces the channel attention module based on the Siamese network and dynamically enhances the robust channel feature response by modeling the context relationship between channels. This paper also verified the network performance on the OTB and VOT benchmark. Experimental results show that the proposed algorithm can achieve robust tracking results on challenging datasets, and achieves the goal of improving network performance with a slight increase in computational cost.