Towards dynamic target identification using optimal design of experiments

This paper discusses a dynamic approach to target recognition that is based on concepts from the theory of optimal design of experiments. The approach uses a cycle of hypothesis formulation, experiment planning for hypothesis validation, experiment execution, and hypothesis evaluation to confirm or reject the classification of targets into given object classes. Target classes of relevance to specific perceptual tasks of a robot mission are described through parametrization of sensor observations. We use spatial stochastic lattice models to encode sensor-based information and to provide potential target hypotheses. Information-theoretic uncertainty minimization metrics are employed to control the sensing processes and the robot vehicle. The approach presented was applied to an unmanned aerial robot vehicle developed for environmental research and monitoring applications, and initial results are presented showing aerial identification and tracking of large-scale man-made structures and biological targets.