Graph Marginalization for Rapid Assignment in Wide-Area Surveillance

Decentralizing optimization problems across a network can reduce the time required to achieve a solution. We consider a wide-area surveillance sensor network observing an environment by varying the state of each sensor so as to assign it to one or more moving objects. The aim is to maximize an arbitrary utility function related to object tracking or object identification, using graph marginalization in the form of belief propagation. The algorithm performs well in an example application with six heterogeneous sensors. In larger network simulations, the time savings owing to decentralization quickly exceed 90%, with no reduction in optimality.