In many tracking and surveillance systems, multisensor configurations are used to provide a greater breadth of measurement information and also to increase the capability of the system to survive individual sensor failure. Most architectures currently used in such systems rely on having either a central processor where global data fusion takes place or a central communications medium through which all messages between sensors must pass. Such centralized architectures give rise to problems with communication and computational bottlenecks and are susceptible to total system failure should the central facility fail. In this article we describe a multisensor surveillance system that achieves a high degree of survivability by employing a decentralized sensing architecture (Durrant-Whyte, Rao, and Hu 1990). Using CCD cameras and optical barriers the system is able to track (in real time) people and robots as they move around a factory room. The decentralized sensing architecture takes the form of a network of Transputer-based sensor nodes, each with its own processing facility, that together do not require any central processor, any central communication facility, or any common clock. In this architecture, computation is performed locally, and communication occurs between any two nodes as and when required. The starting point for the surveillance system is an algorithm that allows complete decentralization of the multisensor Kalman filter equations among a number of sensing nodes. We develop the algorithm to ensure that internodal communication is minimized and can take place without any prior synchronization between nodes. By designing a suitable track management scheme, we are able to extend the system to cope with multitarget tracking. Finally, the specific sensors used in the implementation are interfaced to the algorithm, and the results of the working system are given.
