Citation

Li M, Chen T, Du H, Ma N, Xi X. Transportmetrica A: Transp. Sci. 2023; 19(1): e1976877.

Copyright

(Copyright © 2023, Informa - Taylor and Francis Group)

DOI

10.1080/23249935.2021.1976877

PMID

unavailable

Abstract

Groups are the basic elements which depict the complex social dynamics in crowded scenarios. Most researchers tend to partition people based on trajectories and lack sufficient utilization of group properties. To address this issue, we design a new group detection framework based on multi-level group descriptors. Both crowd trajectories and semantic information are utilized to profile group properties. A unified graph clustering model is given to model multi-feature consistency and inconsistency in a unified objective function. The public pedestrian datasets show that the proposed framework can achieve superior detection performance and density robustness when compared with existing methods. In addition, to verify the capacity of our method at group detection in other motion scenes, we construct a new non-vehicle dataset in real scenes. Extensive experiments on our non-vehicle dataset are performed, and the results demonstrate that our approach is effective and reliable for social group detection in various crowd scenarios.

Language: en

Keywords

crowd analysis; graph fusion; Group detection; unsupervised learning; video understanding