| This work addresses the task of active sensing: control of mobile sensor platforms for the purpose of gathering information. Formulation of a control objective in terms of information gathering allows mobile sensor teams to be both autonomous and easily reconfigurable to include a variety of sensor and target models. The application of tracking a moving target using a camera mounted on a fixed wing UAV is considered throughout, but the control formulation is not specific to this choice of sensor or estimation task. A model predictive control is developed which attempts to minimize the entropy of an estimate over a receding horizon, subject to stochastic models for both the target motion and sensors.;The main focus of this work is on methods for predicting the expected entropy of a filtering density, given a choice of motion controls for the team of mobile sensor platforms. The underlying assumption is that if the receding horizon cost for the team of sensors can be calculated, then any appropriate decentralized optimization algorithm can be used to choose a control action. The prediction of conditional entropy is shown to be inherently difficult, and a computationally efficient sequential Monte Carlo method is developed. For a single sensor, the entropy prediction depends on this Monte Carlo method as well as on a novel approach for entropy calculation in the context of particle filtering. For the multiple sensors case, a novel approximation for mutual information is developed which allows efficient decentralized computation for sensors with limited range. Methods for both single and multiple sensors are verified through simulation. Single sensor methods are also verified through post-processing of experimental flight data.;The primary contribution of this dissertation is a set of methods which allow tracking of a moving target by minimizing an information measure over a receding horizon, even when the estimate distribution is non-parametric. A receding horizon of length greater than one step is especially advantageous for control of kinematically constrained platforms such as fixed wing UAVs. Previous similar work has been restricted to either a stationary target, a horizon of length one, or Gaussian estimates.;Although these methods have been designed with computational efficiency in mind, the entropy prediction problem is inherently expensive in terms of computation, and application to real-time control is not yet feasible. However, rapid advances in embedded computing suggest that these techniques will soon be implementable, providing the ability to address problems not currently solvable using current methods. |
