The detection of unusual events in video based on Bayesian surprise model

Automatic detection of unusual events in video surveillance has proven to be a difficult task. It is shown that surprise is an important cue for the direction of human attention to unexpected events. Aiming to the difficulty of the definition and description of abnormity in videos, we present the use of Bayesian surprise, introduced in computer vision, for detection of certain types of unusual events. Multiple local monitors which collect low-level statistics (velocity or direction) are placed over the image. With the monitors´ observation and a set of reference measurement previously acquired, a pixel-wise surprise trigger is computed using Bayesian probabilistic inference techniques. Once the surprise value exceed by a preset threshold, the monitor outputs an alert. These alerts are integrated to a final decision regarding the existence of an unusual event. It was tested on a variety of real-life crowded scenes. The results show that the proposed method localized the regions of anomalies in the abnormal frames successfully, thus demonstrating the practical applicability of the scheme.