Robust object tracking using segmented credible template and dynamic template set updating strategy

Most Transformer-based trackers rely solely on the query update strategy while neglecting information related to classification confidence. This unilateral update scheme suffers from significant performance degradation in challenging scenarios such as severe target deformation or background clutter. Moreover, current methods typically adopt a static template mechanism, lacking the ability to adaptively adjust the template set in response to varying target states. To address these limitations, we propose an object tracking approach leveraging segmented credible templates and dynamic template set updating, named as SDTrack. By analyzing template confidence feedback scores within each segment, credible templates are dynamically selected to enhance the accuracy of the template. Furthermore, a dynamic template set update strategy with semantic label information is designed that can dynamically capture the different target’s diverse variations and further enhance the tracking robustness. Extensive experiments on benchmarks including GOT-10k, OTB-100, UAV20L, and UAV123 demonstrate the consistent improvement of the proposed approach over state-of-the-art methods in success rate and precision. Ablation studies further attribute this performance to its segmented credible templates and dynamic template set update strategy, which together yield competitive robustness and tracking performance.