Anomaly Detection with False Alarm Rate Controllable Classifiers

Automatic anomaly detection gains attention as the number of surveillance cameras increase. Research on the field concentrates on the development of features modeling the overall behavior in a scene, and with each iteration drives the performance to its limit. However, there is a lack of research for detecting anomalous behavior not only based on extracted features, but also using its spatial properties. In this paper, we propose a framework that utilizes Neyman-Pearson classification for detecting anomalies and also providing a true local anomaly detection method, which is capable of detecting same action as normal or anomalous depending on its spatial properties. Our method is a context tree based, competitive ensemble NP classifier, containing multiple piece-wise linear NP models trained on different partitions of a video frame. Even though the number of partitions increases doubly exponentially with the increasing tree size, efficient tree framework provides training in linear time complexity. Competitive nature of the learning of the ensemble model ensures convergence to the optimal space partitioning depending on the performances of individual NP classifiers.