Crowd modeling using social networks

In this work, we propose an unsupervised approach for detecting the anomalies in a crowd scene using social network model. Using a window-based approach, scene objects are first detected and tracked, and a spatio-temporal partitioning is constructed to produce a set of spatio-temporal cuboids that capture spatial and temporal features. A hierarchical social network is built to model the crowd behavior: the bottom-level models local behavior and the top level models the global. We perform anomaly detection and demonstrate the effectiveness of the proposed approach on a benchmark crowd analysis video sequences. Our results reveal that we outperform majority, if not all, the state-of-the-art methods.