Visual Tracking Based on Siamese Neural Network with Non-local Attention Network
Liying Li, Wu Muqing · 2022
In recent years, Siamese network has been widely studied due to its outperforming results in terms of speed and accuracy. However, how to effectively deal with the influence of large appearance changes, complex backgrounds and distractors in the tracking process is still a problem we need to face. In this paper, we propose a Siamese Neural Network with Non-local Attention Network, referred to as SiamNA. The proposed network consists of two parts. One is the backbone network based on Non-local attention module and the other is the Region Proposal Network (RPN) based on adaptive multi-feature fusion. We find that the attention module can effectively capture the dependencies between elements in the sequence. We place this module after the backbone network to process the extracted features, which can effectively improve the target recognition ability and improve its robustness. We conduct experiments on four benchmarks including GOT-10k, VOT2016, OTB100 and VOT2018, and the results demonstrate that the proposed method is effective for improving accuracy and robustness.