Mutual Information Based Resource Management Applied to Road Constrained Target Tracking

Netted sensors offer advantages for many surveillance applications. Target tracking and identification may be enhanced by jointly exploiting a variety of data sources, and certain surveillance applications may be more readily accomplished with jointly operated small in situ sensors than with large standoff sensors. Efficiently utilizing sensor-network data requires reliable sensor fusion and resource management algorithms. Resource management is particularly important when a limited number of tasks can be performed either because sensors may be used in one of several modes at any given time, or various resources, e.g. energy, computational capabilities and communication bandwidth are limited. In this paper tracking is viewed as a parameter estimation problem. Parameters are values in a state space and inference about the parameters is based on sensor measurements. The utility of sensor measurements is assessed using the mutual information between the parameters and the measurements. Resource management is achieved by minimizing average expected entropy subject to constraints. This approach is applied to a random set tracking algorithm that is based on Gaussian mixture models. Quadratic mutual information, which in this context is computable in closed form, is used as a substitute for mutual information when comparing the utility of sets of sensors of the same cardinality. The Mobius transformation is utilized to reduce the computational requirements of the optimization process. The tracking and resource management algorithms are demonstrated using a simulation capability. Four acoustic arrays, that measure angle of arrival and two radars, that measure range, monitor a triangular road network. For the example shown, two vehicles traversing the network, the tracker and resource manager are able to maintain the approximate quality of the estimate, as measured by average entropy of the distribution of the state space parameters, using, on average, less than 2.5 of the six sensors.