Robust crowd anomaly detection via hybrid ensemble learning for real-world surveillance

Crowd Anomaly Detection (CAD) is a crucial task in intelligent surveillance systems, designed to enhance public safety in densely populated environments. While existing approaches, including deep learning and traditional machine learning models, have shown promise, they often suffer from limited generalizability, high computational cost, or poor performance on small-scale datasets. To address these gaps, this study proposes a novel hybrid ensemble learning framework that integrates the speed and accuracy of YOLOv7 for real-time crowd detection with the robustness and interpretability of classical ensemble classifiers, namely Random Forests (RFs) and Gradient Boosting (GB). The model extracts spatial and motion features via optical flow from YOLO-detected regions, reduces dimensionality, and classifies behaviors as normal or abnormal. The ensemble is optimized using the Adam optimizer to improve performance on data-constrained surveillance scenarios. Experimental results on the University of Minnesota "UMN" benchmark dataset show a near-perfect accuracy of 99.89%, while evaluation on a custom real-world supermarket dataset demonstrates strong generalizability and robustness. These results establish a new standard for small-scale CAD and demonstrate the practical applicability of the proposed framework for intelligent, real-time surveillance.