Information-Enhance Twofold Siamese Tracker for Video Target Tracking

The current video target-tracking networks, for example, SiamFC, SiamRPN, or C-RPN, correlation between the target features in a video sequence is not fully considered in the extracted tracking target features. Consequently, ability to express features of the target is weak, the tracking precision is low, and the interference capability of factors such as background clutter or scale change is poor. In this paper, we propose an information-enhanced twofold Siamese tracker (IETST) to enable the target information to learn the robust features of the target appearance in a twin-network tracker. This IETST mines the representation capability of the template and search features through a self-calibration region proposal module. In addition, we propose an information enhancement region proposal module to improve the discrimination between the target and background by introducing context feature information to the target in the video sequence. The complementary performance of these two types of structures greatly enhances the ability of the tracking network to classify and regress the tracking area. To fully exploit the advantages of the two operations, we combine the classification and regression results obtained with different influencing factors to obtain the tracking result after information enhancement. Experiments on three open test datasets, OTB 100, VOT2016, and VOT2018, show that the proposed IETST performs better than other advanced video target tracking networks.