TPM: Multiple object tracking with tracklet-plane matching. (November 2020)

• Record Type: Journal Article
• Authors: Peng, Jinlong; Wang, Tao; Lin, Weiyao; Wang, Jian; See, John; Wen, Shilei; Ding, Erui
• Is Part Of: Pattern recognition. Volume 107(2020:Nov.)
• Publication Date: 2020-11
• Subjects: Multiple object tracking -- Tracklet -- Tracklet-plane -- Representative-selection network; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107480
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: Propose a tracklet-plane matching process to differentiate confusable tracklets. Introduce a tracklet-importance evaluation scheme to exclude the noisy detections. Introduce a Representative-selection network to calculate tracklet-wise similarity. Improve the performance on MOT dataset. Abstract: Multiple object tracking (MOT) aims to model the temporal relationship among detected objects and associate them into trajectories. Thus, one major challenge of MOT lies in the confusion from noisy object detection results. In this paper, we propose Tracklet-Plane Matching (TPM), a new approach which improves the performance of MOT by modeling and reducing the interferences from noisy or confusing object detections. TPM first constructs good temporally-related object detections into short tracklets. Then, a tracklet-plane matching process is introduced to organize related tracklets into planes and associate them into long trajectories. The tracklet-plane matching process assigns visually confusing tracklets into different tracklet planes according to their contextual information, thus properly reducing the confusion among similar tracklets. At the same time, it also allows association among temporally non-neighboring or overlapping tracklets, which provides good flexibility to handle confusion from noisy detections. Under this process, a tracklet-importance evaluation scheme and a representative-based similarity modeling scheme are introduced. These two schemes can properly evaluate the reliability of detection results and identify reliable ones during association so that the impact of noisy or confusing detections can be well-mitigated. Experimental results on benchmark datasets demonstrate that the proposed approach outperforms the state-of-the-art MOT methods.