Vision system for decentralised wide area surveillance using multiple nodes

In a decentralised wide area surveillance system (D-WASS), the two principal objectives are to first detect the events of interest reliably, and second to track the progression of these events and maintain a globally consistent track of the object for the extent of the camera fields of view. In this paper, we discuss an implementation of a 9-camera D-WASS with non-contiguous views, and analyse the performance of the tracking and vision sub-systems both empirically and analytically. In this application, event-based change detection algorithms in image space, as well as data association and tracking algorithms in a consistent real world coordinate system are realised in a real-time environment where both physical and temporal constraints are taken into consideration. This system has been demonstrated to work satisfactorily under various operating conditions. In order to characterise the performance of the system, the vision and tracking components, as well as the dependence of the latter on the former from a systems viewpoint, are studied analytically. The results are compared with those obtained empirically from real ground-truthed data. Potential extensions and improvements to the current implementation are also discussed.