A central problem in sensor data fusion is the recovery of complex, multi-component models from various sensory modalities. An essential component of this process is a description of what decision must be made from the recovered model. This description is referred to as the sensing task. The goal of this research project is to develop and analyze techniques for recovering the minimal or least detailed model required to make a satisfactory decision for a given sensing task. In particular, the focus is on the task-directed recovery of composite models (models composed os several components). The model recovery method is based on the numerical solution os systems of nonlinear constraints using interval bisection. Previous work, based on making decisions about objects modeled by a single parametric form, showed that these methods are general, natural, simple to implement, and computationally effective. In addition, the investigations were able to incorporate notions of the cost of computation and the value of information into the recovery process, and to terminate the recovery process when the model with the highest net value (decision payoff minus computational cost) was reached. They are now extending the underlying recovery method by incorporating model refinement and data segmentation in a manner that also exploits information about the sensing task. The results are expect to have particular impact in application area where good a priori models ar not available. Examples of these domains include classification and sorting of irregular (e.g. naturally occurring) objects, supervisory control, and ultimately, to sensor-based control of partially or fully autonomous vehicles.
