Adaptive Template Update and Re-Detection Network Based on Tracking Confidence

Siamese tracking is widely used in object tracking due to its efficient dual-branch symmetric structure, deep feature matching mechanism, and flexible template strategy. Existing mainstream Siamese tracking algorithms typically employ static template matching or linear combination-based template updating to localize the target in the next frame. However, these mechanisms often struggle to ensure template accuracy in complex environments involving changes in target appearance, scale, occlusion, and motion blur, thereby compromising robustness and stability. To address these issues, this paper proposes a confidence-guided adaptive template update with a re-detection (CATUR) network. CATUR constructs a tracking confidence assessment module that uses average peak-to-correlation energy (APCE) and a dynamic threshold mechanism to determine the current tracking state, providing a basis for template updates and target re-detection. It also designs an adaptive template update network that effectively combines the initial, historical, and current-frame templates, enhancing adaptation to target appearance variations. By integrating a global search module and a re-detection module, CATUR achieves precise target re-localization, rapid template updating, and tracking recovery. Extensive experiments and ablation studies on LaSOT and TrackingNet demonstrate that CATUR improves AUC, PNorm, and P by 4.0%, 4.0%, and 3.2%, respectively, significantly enhancing tracking accuracy and robustness in complex environments.