Detecting abnormal activities in video sequences

Automatically detecting suspicious human activities in restricted environments such as airports, parking lots and banks represents an open issue for the last generation surveillance systems. In this paper, we present an approach that allows to detect anomalies in a video sequence without any need of describing a priori "abnormal" activities. In particular, we first introduce a normal activities model based on the concept of elementary actions observable by means of image understanding procedures. We then provide an algorithm based on the use of decision trees that can quickly detect an abnormal situation as variation of currently processed activity with respect to normal patterns contained in the system knowledge base. Our preliminary experimental results on a dataset consisting of staged bank robbery videos show that our algorithm provides very encouraging results when compared to human reviewers.