Authorization

 

Journal

Volume 22, №3, Special Issue on the Artificial Intelligence; Where Do We Stand?

Dual-branch Siamese network for object tracking based on attention and low-high-resolution images (pp.409-429)
Shuren Zhou1,*, Zhixiong Li1,*, Jie Liu1, Jiarui Zhou1, Jianming Zhang1
https://doi.org/10.30546/1683-6154.22.3.2023.409


1Changsha University of Science and Technology

∗Corresponding author’s e-mail: zsr_hn@163.com

 

 

Abstract. Siamese target tracking is applied in various fields, but the accuracy of tracking and robustness has always been a hot issue. Therefore, improving the target tracking model is a challenging task. To solve the problem, we present Dual Branches Siamese Networks (DSN) consisting of the Visual Branch (V-Branch) and the Resolution Enhancement Semantic Branch (E-Branch). Firstly, a visually trackable Horizontal Pyramid Attention (HPA) V-Branch was obtained through a target tracking algorithm, and its attention mechanism improved target discrimination by automatically learning the relevance weights of each channel in the feature map. Secondly, we increased the resolution of the template images during the training of the E-Branch and proposed a novel approach called Low-High-Resolution template images (LH) to mitigate the influence of background noise during the testing process. Finally, the independently trained dual branches are fused through the fusion module (FM). The experiment uses five typical datasets to validate the DSN.

Keywords: Siamese Network, Semantic Enhancement, Object Tracker, Neural Network, Attention Mechanism

 

AMS Subject Classification: 68T07.



COIA 2027

7-9 July 2027
Baku, Azerbaijan

Search