Localization-Guided Track: A Deep Association Multiobject Tracking Framework Based on Localization Confidence of Camera Detections

Abstract

Current camera-based tracking-by-detection (TBD) methodologies in the literature overlook the signifi- cance of detection box localization confidence, generally assuming that objects with low detection confidence are highly occluded and therefore either ignored or deprioritized in the matching process. Furthermore, appearance similarity is typically neglected when matching these low-confidence objects. This oversight presents a critical research gap as objects with low detection confidence might still exhibit a clear appearance, and those with high detection confidence can present imprecise localization or ambiguous appearance factors not accounted for in existing methods. To address this gap, our work contributes a novel framework, localization-guided track (LG-Track), which, for the first time, integrates localization confidence of camera detections in multiobject tracking (MOT). LG-Track takes into account both appearance clarity and localization precision of detection boxes, incorporating a novel deep association mechanism that enhances tracking performance. Based on the localization and classification confidence of detection boxes, different cost matrices are employed in different levels of the proposed deep association mechanism to achieve enhanced matching accuracy. Our method, validated through rigorous experimentation on the MOT17 and MOT20 benchmarks, demonstrates superior performance over current compared state-of-the-art (SOTA) tracking methods. Committed to furthering research in this field, we have made our code accessible to the community at https://github.com/mengting2023/LG-Track.