Multi-camera video surveillance

This paper describes the development of a multi-view video surveillance and the algorithms to detect and track objects (generally low densities of pedestrians, cyclists and motor vehicles) moving through an outdoor environment imaged by a network of video surveillance cameras. The system is designed to adapt to the widely varying illumination conditions present in such outdoor scenes, as well as coping with the spurious motion of non-objects (such as vegetation) and the interaction of objects within the scene. Where possible, the system takes advantage of multiple views of the same object to help resolve occlusion and increase the robustness of the tracking. Overlapping camera views are corresponded using a geometric analysis of camera viewpoints. The system aims to capture scene dependent information through learning, constructing models of the scene using observations extracted from the camera network. The system is currently undergoing a real-time implementation using live camera data, and results are presented from this on-line system.