Anomaly Detection via Local Coordinate Factorization and Spatio-Temporal Pyramid

Anomaly detection, which aims to discover anomalous events, defined as having a low likelihood of occurrence, from surveillance videos, has attracted increasing interest and is still a challenge in computer vision community. In this paper, we propose an efficient anomaly detection approach which can perform both real-time and multi-scale detection. Our approach can handle the change of background. Specifically, Local Coordinate Factorization is utilized to tell whether a spatio-temporal video volume (STV) belongs to an anomaly, which can effectively detect spatial, temporal and spatio-temporal anomalies. And we employ Spatio-temporal Pyramid (STP) to capture the spatial and temporal continuity of an anomalous event, enabling our approach to handle multi-scale and complicated events. We also propose an online method to update the local coordinates, which makes our approach self-adaptive to background change which typically occurs in real-world setting. We conduct extensive experiments on several publicly available datasets for anomaly detection, and the results show that our approach can outperform state-of-the-art approaches, which verifies the effectiveness of our approach.