Hierarchical Clustering and Refinement for Generalized Multi-Camera Person Tracking
Zongyi Li, Runsheng Wang, He Li, Bohao Wei, Yuxuan Shi, Hefei Ling, Jiazhong Chen, Boyuan Liu, Zhongyang Li, Hanqing Zheng
Abstract
Multi-camera person tracking has gained significant attention in recent times, owing to its widespread application in surveillance scenarios. However, this task is challenging due to the variance viewpoints, heavy occlusion, and illumination changes. In order to tackle these challenges, we propose a novel Hierarchical Clustering and Refinement framework for Generalized Multi-Camera Person Tracking. Specifically, our framework comprises two main components: hierarchical clustering and hierarchical refinement. Compared with directly clustering tracklets among multiple cameras, our hierarchical clustering strategy can progressively assign tracklets to correct targets. Nevertheless, the clustering and tracking process would inevitably produce incorrect matchings. Therefore, a hierarchical refinement strategy is proposed to reduce these incorrect matches which includes: intra-camera tracklet level refinement, appearance refinement, spatial-temporal refinement, and face refinement. Extensive experiments show the effectiveness of our method, which achieves 92% IDF1 in 2023 AI CITY CHALLENGE track1, ranking 5th on the leaderboard.