Description and Discovery of Complex Events in Video Surveillance

Behavior understanding and semantic interpretation of dynamic visual scenes have attracted a lot of atten-tion in computer vision research community. Although the use of surveillance cameras has proliferated, the un-derstanding of activities still remains complex. While users are mostly interested in high level and subjec-tive semantics, only low level visual features can be ex-tracted in a reliable way. This paper presents a novel framework for video guided behavior monitoring, built around the event modeling concept. It enables users to design their personal models of events combining ele-mentary concept and low level features using expres-sive formalisms. The framework enables then detec-tion of the events within video streams based on low level features extraction and manual annotations analy-sis, while taking in consideration uncertainty. Examples depicting content-based events modeling and detection from video surveillance are presented to illustrate the approach. 1