Abstract: In order to solve the problem that the evolution of siamese visual tracking with very deep networks (SiamRPN++) algorithm when the target is occluded and deformed. A siamese network target tracking algorithm combining dual attention and feature fusion is proposed. Firstly, the channel and spatial attention module are used to enhance the target information, suppress the interference information in the image and improve the accuracy of the model; then, multi-layer feature fusion is carried out for the shallow and deep feature information output from the attention layer to obtain better expressive target features and improve the tracking accuracy; finally, the online template update mechanism is introduced to reduce tracking drift and improve the tracking robustness. The OTB100 dataset is used for experimental and the results show that the success rate of the improved algorithm is increased by 1.3% compared with SiamRPN++, indicating that the tracking accuracy of the algorithm is higher; under the four test sequences with occlusion and deformation attributes, the average overlap rate of the improved algorithm is increased by 3%, and the average center position error is reduced by 0.37 pixels, which is better robustness against occlusion and appearance deformation.

Key words: target tracking, siamese network, dual attention module, feature fusion, online template update