Adaptive Update and Association for Visible-Infrared Multi-Object Tracking

In recent years, multi-object tracking (MOT) as an intelligent perception system benefits from the rapid advancement of autonomous driving. MOT provides essential adaptive environmental understanding capabilities for autonomous navigation technologies, demonstrating broad application prospects. However, conventional tracking approaches predominantly focus on optical image analysis, which severely limits their applicability in real-world scenarios where multimodal information is required to address diverse and complex environments. To solve the problem, we propose a novel adaptive visible-infrared fusion framework for MOT. Our method introduces a learning-based cross-modal multi-scale feature level fusion mechanism, achieving robust detection of target objects under varying illumination and weather conditions. Furthermore, we integrate adaptive strategies for appearance embedding updates and matching association within the tracking framework, substantially enhancing our algorithm’s intelligent perception capabilities in real-world environments. Finally, we conduct comprehensive experiments on the visible-infrared CAMEL dataset and yield competitive results, validating the effectiveness of our proposed approach.