Multi-object tracking with dynamic weighted appearance evolution and local matching

This study addresses the multi-object tracking challenges in long-term video surveillance scenarios. In such scenarios, pedestrian targets often face extended temporal spans, occlusions, background noise, and non-linear illumination changes. These challenges make it difficult for conventional pedestrian re-identification algorithms to effectively capture their appearance features. To tackle these issues, this paper proposes the method called dynamic weighted appearance evolution and local matching. It utilizes a dynamic weighted appearance evolution mechanism, addressing the problem of preserving only short-term features of targets in traditional feature lists and mitigating excessive noise interference in trajectory features when detection quality is limited. In the matching stage, a strategy employing local features to assist global features in re-identification is utilized. Precise matching of local features enables better handling of challenges such as target occlusion and non-linear illumination changes. Finally, the logic for updating trajectory appearance features is refined, extracting effective local features from lower-quality detections to precisely update trajectories, achieving more accurate tracking results. Experimental evaluation shows that our approach reduces the occurrence of ID switching and is useful for tracking reliability in the complex case of long videos.