Attention and anticipation in complex scene analysis - an application to video surveillance

We report on a video surveillance system that has anticipatory and attentive capabilities in order to efficiently cope with its resource limitations in analyzing complex visual scenes. We base our system on a mathematical-physical framework for cognitive engineering of cybernetic systems. Anticipation and selection of attention allow the analysis system to actively search for potential targets in the visual scene. The target selection mechanism is based on a simple scheme for detecting moving regions. In addition, anticipation ensures that a pattern recognition module analyzes all potential targets in the order in which they escape the field of view of the video camera. This anticipatory mechanism is based on a simple escape-time measure. Furthermore, a utility measure is defined to improve the overall performance of the target selection and pattern recognition tasks, and is used as a means to change proactively the system's focus of attention.