| Abstract |
Multi-subject tracking in crowded videos is an established yet challenging research direction in computer vision and information processing. High applicability of multi-subject tracking is demonstrated in smart cities (e.g., public safety, crowd management, urban planning), autonomous driving vehicles, robotic vision, or psychology (e.g., social interaction and crowd behavior understanding). In this work, we propose a real-time approach that reveals tracks of subjects in ordinary videos, captured in highly populated pedestrian areas, such as squares, malls, and stations. The tracks are discovered based on the proximity of detected bounding boxes of subjects in consecutive video frames. The reduction of track fragmentation and identity switching is achieved by the re-identification phase that uses caching of unassociated detections and mutual projection of interrupted tracks. As the proposed approach does not require time-consuming extraction of appearance-based features, the superior tracking speed is achieved. In addition, we demonstrate tracker usability and applicability by extracting valuable information about body-joint positions from discovered tracks, which opens promising possibilities for detecting human relationships and interactions. We demonstrate accurate detection of couples based on their holding hand activity and families based on children’s body proportions. The discovery of these entitative groups is especially challenging in crowded city scenes where many subjects appear in each frame. © 2022, The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature. |
