Statistical Models of Targets and Clutter for Use in Bayesian Object Recognition

This chapter describes a model-based, Bayesian approach to automated target recognition (ATR). The elements of deformable template theory are used to mathematically model the variations in target pose. For each possible object, a template (using 3D computer-assisted design [CAD] models and other descriptors) of standard size, pose, and location has been defined. All occurrences of a target in a scene can then be represented by scaling, rotating, and translating its template appropriately. The goal is to derive ATR algorithms and analyze them for their performance. The approach relies on two main building blocks: efficient mathematic representations of scenes containing the targets and efficient algorithms for inferences on these representation spaces.