Advanced visual surveillance using Bayesian networks

Visual surveillance primarily involves the interpretation of image sequences. Advanced visual surveillance goes further and automates the detection of predefined alarm events in a given context. However, it is the intelligent, dynamic scene and event discrimination which lies at the heart of advanced vision systems. Developing a systematic methodology for the design, implementation and integration of such systems is currently a very important research problem. One way of overcoming some of these problems is to build in more knowledge of the scene and tasks. For example, in object detection and tracking in the image, we have demonstrated that it is beneficial to bring scene-based knowledge of expected object trajectories, size and speed into the interpretation process. We have also shown that both scene and task-based knowledge allows for selective processing under attentional control for behavioural evaluation. In addition to this general requirement for integration of information in advanced visual surveillance, we have adopted more specific requirements. A fixed, precalibrated camera model and precomputed ground-plane geometry is used to simplify the interpretation of the scene data in the on-line system. We also adopt a knowledge-based approach in which domain specific models of the dynamic objects, events and behaviour are used to meet the requirement for sensitive and accurate performance.