On-line normality modelling and anomaly event detection using spatio-temporal motion patterns

The importance of anomaly event detection is growing up in recent years due to the improvement of security and safety in video surveillance systems. We focus on the assumption that the exhibition frequency is the essential characteristic of an event to be normal or abnormal. In this paper, we propose an on-line and real-time normality modelling algorithm to deal with the anomaly detection problem in complex and crowded environments. Actions or events are interpreted as a set of spatio-temporal motion patterns in a neighbourhood, where a bag of words (BOV) approach is used to define a dictionary of common motion patterns. Them, these actions are modelled using a probabilistic framework in order to obtain the probability of being an abnormal behaviour. This anomaly probability is based on a Gaussian mixture model (GMM) which is able to add new behaviours appearing in the environment. We demonstrate the efficiency of the proposed algorithm on publicly available datasets.