This technology proposes a bilinear pooling appearance model based on a Siamese network to address the issue of target tracking in occluded scenarios. Core breakthroughs include: 1) Establishing a four-stream network structure (infrared/visible light template + search area) and integrating multimodal features through a bilinear pooling module; 2) Proposing a loss function based on inner products for end-to-end training and improve candidate sample accuracy. It significantly enhances the robustness of RGB-T target tracking under severe occlusion.

Technology provider:Nanjing University of Posts and Telecommunications