Bayesian analysis of small groups in crowds

Crowd analysis is of increasing interest. Sociologists, for example, are interested in studying social influence within and between small groups of individuals traveling in a crowd. However, current empirical studies conducted by human observers are very time-consuming. We propose an automatic, vision-based system that detects and tracks all individuals in video of crowds of varying nature and density while discovering small social groups. Individual components of the system tackle computer vision problems that are challenging in their own right, namely, 1) to reliably detect individual people under reasonable crowd density and from different viewpoints, 2) to robustly track individuals through crowds where inter-person occlusion is frequent, and 3) to infer at a distance which small groups of people are interacting or traveling together. Results from our automated crowd analysis system reveal interesting pedestrian group walking patterns that complement current research in crowd dynamics. These discoveries also may provide helpful insights for evacuation planning and for real-time situation awareness during emergency response to public disturbances.