Spatiotemporal Anomaly Detection Using Deep Learning for Real-Time Video Surveillance

Abstract

Rapid developments in urbanization and autonomous industrial environments have augmented and expedited the need for intelligent real-time video surveillance. Recent developments in artificial intelligence for anomaly detection in video surveillance only address some of the challenges, largely overlooking the evolving nature of anomalous behaviors over time. Tightly coupled dependence on a known normality training dataset and sparse evaluation based on reconstruction error are further limitations. In this article, we propose the incremental spatiotemporal learner (ISTL) to address challenges and limitations of anomaly detection and localization for real-time video surveillance. ISTL is an unsupervised deep-learning approach that utilizes active learning with fuzzy aggregation, to continuously update and distinguish between new anomalies and normality that evolve over time. ISTL is demonstrated and evaluated on accuracy, robustness, computational overhead as well as contextual indicators, using three benchmark datasets. Results of these experiments validate our contribution and confirm its suitability for real-time video surveillance.