Sensor Management for Particle Filter Tracking

Different information theoretic sensor management approaches are compared in a Bayesian target-tracking problem. Each approach compares the expected improvement to the posterior from potential sensor queries, taking into account the current estimated prior but not the actual measurements, to determine the optimal myopic sensing action to take. Specifically the results of prioritizing sensor selection using the expected Renyi divergence with different parameter values are compared with a random sensing scheme and to a computationally efficient linear-Gaussian approximation of the Renyi divergence. The approximation is only used in the optimization step. A particle filter representation of the tracker is used for the motion and measurement updates for all methods. Included is the special case in which the expected Renyi divergence is equal to the expected Kullback-Leibler divergence, which is also equivalent to both the mutual information and the expected change in differential information for this Bayesian updating problem. Two example problems from the context of antisubmarine warfare (ASW) using sonar systems are considered. Each involves a single maneuvering target and both bearing-only and time delay of arrival multistatic sensors. Compared with a random selection of measurements, all of the prioritized schemes localize the target more quickly and more precisely. It is shown that the computationally efficient linear-Gaussian approximation method provides better target localization than the more complex Renyi divergence implementation under some circumstances and generally performs at least comparably in the examples considered. Additionally, the role of the Renyi
$\alpha$
parameter is examined both generally and through several numerical examples.