Unsupervised Behavior-Specific Dictionary Learning for Abnormal Event Detection

Abstract

Abnormal event detection has been an important issue in video surveillance applications. Due to the huge amount of surveillance data, only a small proportion could be loaded during the training. As a result, there is a high chance of incomplete normal patterns in the training data, which makes the task very challenging. Sparse representation, as one of solutions, has shown its effectiveness. The basic principle is to find a collection (a dictionary) of atoms so that each training sample can only be represented by a few atoms. However, the relationship of atoms within the dictionary is commonly neglected, which brings a high risk of false alarm rate: atoms from infrequent normal patterns are difficult to be distinguished from real anomalies. In this paper, we propose behavior-specific dictionaries (BSD) through unsupervised learning, in which atoms from the same dictionary representing one type of normal behavior in the training video. Moreover, 'missed atoms' that are potentially from infrequent normal features are used to refine these behavior dictionaries. To further reduce false alarms, the detection of abnormal features is not only dependent on reconstruction error from the learned dictionaries, but also on non zero distribution in coefficients. Experimental results on Anomaly Stairs dataset and UCSD Anomaly dataset show the effectiveness of our algorithm. Remarkably, our BSD algorithm can improve AUC significantly by 10% on the stricter pixel-level evaluation, compared to the best result that has been reported so far.