AI-Driven Abnormal Behaviour Detection within Crowds

Abstract

The detection of abnormal behaviour in crowded areas is crucial given the reliance on video surveillance for public safety. By monitoring automated Video Anomaly De- tection (VAD) systems, authorities can manage crowds and intervene swiftly when accidents occur, particularly in congested regions such as city centres or stadiums. This study presents a graph-based learning framework leveraging Graph Neural Net- works (GNNs) and Multi-Layer Perceptrons (MLPs) for spatial-temporal feature extraction and anomaly detection. The approach constructs a graph where edges represent velocity similarities between objects in the scene, while nodes are enriched with features capturing the direction and magnitude of individual movements. The feature captures trajectory-based similarities and spatial-temporal attributes, en- abling effective graph construction for anomaly detection. The model integrates Graph Attention Networks (GATs) to encode the graph structure, creating latent embeddings that capture both local and global relationships between people, and the MLP decoder reconstructs the velocity relationships. This autoencoder (AE) struc- ture can identify anomalies by comparing the true movement with the reconstructed one, anomaly is detected when the reconstruction error is higher than the threshold. We conducted experiments on UCSD Pedestrians and UMN datasets, to evaluate the model. Demonstrating the successful use of similarities as graph features for GNNs to generate logical embeddings of a crowd movement, and highlighting the impact of the time windows on anomaly detection precision as reflected in F1 scores. The Area Under the Curve (AUC) metric shows stability, highlighting the model’s balanced performance. Qualitative analysis of the UMN dataset illustrates the model’s ability to align detection with visual anomalies. Overall, this study highlights the precision of the proposed graph-based anomaly detection framework in understanding crowd movement behaviours in crowd management surveillance systems.