SocialVis: Dynamic social visualization in dense scenes via real‐time multi‐object tracking and proximity graph construction
Bowen Li,Wei Li,Jingqi Wang,Weiliang Meng,Jiguang Zhang,Xiaopeng Zhang
DOI: https://doi.org/10.1002/cav.2272
IF: 1.01
2024-05-26
Computer Animation and Virtual Worlds
Abstract:Our SocialVis is a framework for visualizing social dynamics, providing real‐time monitoring of individual behavior and social distancing in crowds. Its SocialDetect algorithm, using LSKA attention and Wise‐IOUv3 loss, improves accuracy in dense settings with occlusion and scale changes. To simulate interactions among individuals, we employ a graph theory‐based approach that utilizes multi‐object tracking, combined with SPGA, PFV, and PPV techniques, to provide comprehensive insights into crowd behavior. To monitor and assess social dynamics and risks at large gatherings, we propose "SocialVis," a comprehensive monitoring system based on multi‐object tracking and graph analysis techniques. Our SocialVis includes a camera detection system that operates in two modes: a real‐time mode, which enables participants to track and identify close contacts instantly, and an offline mode that allows for more comprehensive post‐event analysis. The dual functionality not only aids in preventing mass gatherings or overcrowding by enabling the issuance of alerts and recommendations to organizers, but also allows for the generation of proximity‐based graphs that map participant interactions, thereby enhancing the understanding of social dynamics and identifying potential high‐risk areas. It also provides tools for analyzing pedestrian flow statistics and visualizing paths, offering valuable insights into crowd density and interaction patterns. To enhance system performance, we designed the SocialDetect algorithm in conjunction with the BYTE tracking algorithm. This combination is specifically engineered to improve detection accuracy and minimize ID switches among tracked objects, leveraging the strengths of both algorithms. Experiments on both public and real‐world datasets validate that our SocialVis outperforms existing methods, showing 1.2% improvement in detection accuracy and 45.4% reduction in ID switches in dense pedestrian scenarios.
computer science, software engineering