A Visual Tracking Method for Low-Light-Induced Response Degradation

Robust visual object tracking is a fundamental perception capability that enables mobile robots to follow targets continuously in low-light environments. Insufficient illumination degrades informative cues such as target texture, contours, and local appearance details, resulting in ambiguous response peaks and competition among multiple candidate hypotheses. Under severe visual degradation, the highest-response location may no longer correspond to the true target. To address this issue, we propose a reliability-aware candidate selection framework specifically designed for low-light object tracking. Rather than relying solely on the global response maximum, the proposed framework retains multiple candidate hypotheses corresponding to spatially distinct response peaks. A candidate reliability assessment module then integrates the current visual representation of each candidate, the fused features generated by the base tracker, the historical target representation, and the geometric relations and response differences among candidates to jointly estimate their relative localization quality and replacement confidence. The original prediction is revised only when an alternative candidate demonstrates both a sufficient localization-quality advantage and sufficient replacement confidence, thereby reducing the risk of erroneous candidate switching and tracking drift. Experiments on the LLOT dataset and low-light videos captured by a real robot demonstrate that the proposed method alleviates response ambiguity and reduces candidate misselection, leading to more accurate target localization and more stable tracking.