Target-Sensitive Graph Matching in RGB-T Object Tracking

Visible light and thermal infrared target tracking is a technique that utilizes both visible light and thermal infrared bimodal information to track targets. In recent years, deep neural network-based methods have become the most popular approach for RGBT tracking. In RGBT target tracking based on siamese networks, template and search region information aggregation is commonly achieved through cross-correlation operations, where the template is typically treated as a whole for global matching with the search region. However, this global matching approach neglects the local correspondence between the target and the search region, which hinders its ability to effectively adapt to target deformation and pose changes. Furthermore, the fixed region size used for template cropping also leads to confusion between foreground and background information within the template to some extent. This paper proposes a novel target-sensitive graph propagation module to address the challenges posed by large-scale target deformation, extreme aspect ratios, and other difficulties encountered during tracking. Experimental results on standard RGBT datasets demonstrate that the proposed method significantly improves tracking accuracy and success rate.