Robust RGB-T Tracking via Multi-Feature Response Adaptive Fusion and Dynamic Selection Recovery

In RGB and thermal (RGB-T) modalities fusion tracking, the multi-feature responses of each modality contain rich consistency in object localization, which is crucial to enhance tracking robustness. However, existing decision-level fusion paradigms mostly focus on fusing the output of the last layer, ignoring the correlation between multi-feature responses. Moreover, they also lack consideration of tracking failure, which hinders the application of RGB-T tracking in complex environments. To this end, this paper proposes a multi-feature response adaptive fusion model and a dominant-auxiliary dynamic selection recovery mechanism. Specifically, the former achieves joint optimal fusion by mining the correlation between multi-feature responses. The latter flexibly switches between short-term and long-term tracking modes according to the reliability of tracking results, and utilizes the most reliable modality to further improve tracking stability. Experiments on five prevalent RGB-T tracking benchmarks demonstrate the competitive performance of our method compared with the state-of-the-art methods.