Semantic Analysis of Crowd Scenes

This presentation introduces two novel approaches for the semantic analysis of crowded scenes, addressing key challenges such as high density, occlusions, and complex dynamics in visual surveillance. The first approach uses Motion Units (MUs) and meta-tracking to model local motion dynamics as linear systems, constructing a connectivity graph to automatically discover scene gates, common pathways, and detect anomalies. The second, more advanced approach employs Long-Short Term Memory (LSTM) networks to create predictive models for Super Regions (SRs), generating smoother and more natural synthetic trajectories with improved computational efficiency through parallelization. Both methods are rigorously evaluated on datasets like Grand Central Station and Marathon, demonstrating effectiveness in gate/pathway detection and anomaly recognition (e.g., panic situations, wrong-way movements), with the LSTM-based method showing superior performance in AUC for anomaly detection.