Trajectory‐BERT: Pre‐training and fine‐tuning bidirectional transformers for crowd trajectory enhancement

• Published in: Computer animation and virtual worlds Vol. 34; no. 3-4
• Main Authors: Li, Lingyu, Huang, Tianyu, Li, Yihao, Li, Peng
• Format: Journal Article
• Language: English
• Published: Chichester Wiley Subscription Services, Inc 01.05.2023
• Subjects: crowd trajectory tracking; machine learning; multi‐person tracking; Occlusion; Switches; Training; Tuning
• ISSN: 1546-4261; 1546-427X
• DOI: 10.1002/cav.2190

To address the issue of trajectory fragments and ID switches caused by occlusion in dense crowds, we propose a space‐time trajectory encoding method and a point‐line‐group division method to construct Trajectory‐BERT in this paper. Leveraging the spatiotemporal context‐dependent features of trajectories, we introduce pre‐training and fine‐tuning Trajectory‐BERT tasks to repair occluded trajectories. Experimental results show that data augmented with Trajectory‐BERT outperforms raw annotated data on the MOTA metric and reduces ID switches in raw labeled data, demonstrating the feasibility of our method. In this paper, we propose a space‐time trajectory encoding method and a point‐line‐group division method to construct Trajectory‐BERT to address the issue of occlusion‐induced trajectory fragments and ID switches in dense crowds. Our pre‐training and fine‐tuning tasks leverage spatiotemporal context‐dependent features of trajectories to repair occluded trajectories. Experimental results demonstrate that Trajectory‐BERT outperforms raw annotated data on MOTA and reduces ID switches in raw labeled data, showing the feasibility of our approach.