SiamTP: A Transformer tracker based on target perception

Abstract

Previous trackers based on Siamese network and transformer do not interact with the feature extraction stage during the feature fusion, excessive weight of the target features in the template area when the target deformation is large during feature fusion, causing target loss. This paper proposes a target tracking framework with target perception based on Siamese network and transformer. First, feature extraction was performed on the template area and search area and the extracted features were enhanced. A concatenation operation is used to combine them. Second, we used the feature perception obtained during the final stage of attention enhancement by searching for images to rank them and extracted the features with higher scores to enhance the feature fusion effect. Experimental results showed that the proposed tracker achieves good results on four common and challenging datasets while running at real-time speed with a speed of approximately 50 fps on a GPU.