Abstract
With the development of computer vision, single-modality (RGB) object tracking has achieved significant progress. However, due to the limitation of a single sensor, its performance is still restricted in complex scenarios such as illumination variation, occlusion, and background clutter. Multi-modal tracking based on visible and thermal infrared images can provide richer representations. However, most existing methods mainly rely on current frame information and lack effective modeling of historical information, making it difficult to adapt to temporal appearance changes of targets. To address this problem, a temporal-adaptive memory-based RGBT tracking method is proposed. First, a Dynamic Template Memory Policy is designed to update template representations adaptively according to appearance changes. Then, a Gated Memory-enhanced Search Representation module is introduced to enhance feature representation by incorporating historical information. Furthermore, a Similarity-driven Adaptive Memory Unit is proposed to selectively update memory based on cosine similarity, which effectively suppresses redundant information and improves feature diversity. Experimental results on the RGBT234 dataset show that the proposed method achieves 88.7% in Precision Rate and 63.0% in Success Rate, outperforming the baseline (85.7%/60.7%) by 3.0% and 2.3%, respectively. Ablation studies further verify the effectiveness of each module. In conclusion, the proposed method effectively models historical information through a temporal-adaptive memory mechanism and significantly improves the robustness and stability of multi-modal object tracking in complex scenarios.