Spatio-temporal pattern detection using dynamic Bayesian networks

This paper describes an intelligent decision aid, the TBM Reasoner, that helps military analysts rapidly identify time-critical targets (TCTs) involving theater ballistic missiles from sensor data. Our approach starts for the realization that existing sensors lack either the coverage or the resolution to reliably identify TCTs. Wide area sensors enable simultaneous tracking of all vehicles in a large area, but provide insufficient information to identify vehicle type. In contrast, current-generation high-resolution sensors can reliably identify vehicle type but are unable to search a large area fast enough to identify TCTs. To resolve this dilemma, we bring to bear additional knowledge regarding TCT behavior and capabilities. This knowledge is modeled using dynamic bayesian networks (DBNs). Our approach integrates established technology for multi-hypothesis tracking from data fusion with recently developed algorithms for efficient approximate inference in DBNs. The resulting "smart filter" can be employed to rapidly identify high-probability candidate targets in large volumes of sensor data. The approach described here can be easily generalized to a much wider range of problems of spatio-temporal pattern recognition under uncertainty.