Computational Methods for Task-Directed Sensor Data Fusion and Sensor Planning

In this paper, we consider the problem of task-directed information gathering. We first develop a decision-theoretic model of task-directed sensing in which sensors are modeled as noise-contaminated, uncertain measurement systems and sensing tasks are modeled by a transformation describing the type of information required by the task, a utility function describing sensitivity to error, and a cost function describing time or resource constraints on the system. This description allows us to develop a standard conditional Bayes decision-making model where the value of information, or payoff, of an estimate is defined as the average utility (the expected value of some function of decision or estimation error) relative to the current probability distribution and the best estimate is that which maximizes payoff. The optimal sensor viewing strategy is that which maximizes the net payoff (decision value minus observation costs) of the final estimate. The advantage of this solution is generality--it does not assume a particular sensing modality or sensing task. However, solutions to this updating problem do not exist in closed-form. This, motivates the development of an approximation to the optimal solution based on a grid-based implementation of Bayes' theorem. We describe this algorithm, analyze its error properties, and indicate how it can be made robust to errors in the description of sensors and discrepancies between geometric models and sensed objects. We also present the results of this fusion technique applied to several different information gathering tasks in simulated situations and in a distributed sensing system we have constructed.