Convex Polytope Ensembles for Spatio-Temporal Anomaly Detection

Modern automated visual surveillance scenarios demand to process effectively a large set of visual stream with a limited amount of human resources. Actionable information is required in real-time, therefore abnormal pattern detection shall be performed in order to select the most useful streams for an operator to visually inspect. To tackle this challenging task we propose a novel method based on convex polytope ensembles to perform anomaly detection. Our method relies on local trajectory based features. We report State-of-the-Art results on pixel-level anomaly detection on the challenging publicly available UCSD Pedestrian dataset.