Recognizing and locating three-dimensional objects from range data using feature extraction by demand

This dissertation developed a machine vision system to recognize and locate 3D solid objects using range images. This system contains two major components: feature extraction by demands, and object recognition and registration by location constraints refinement. Feature extraction by demands has a focus mechanism which efficiently extracts a high-level object description of a viewed object, represented by surfaces and their adjacency relations, from image areas containing surface points. The second component identifies an object known a priori, by mapping each surface of the object description to a surface of the object with the same view-independent characteristics. The mapping preserves adjacency relations between surfaces and is efficiently found by dynamically constructing the map and combining location constraints imposed by each surface pair. The location of the object can be calculated as a by-product of object recognition. The results of the research can be applied in model acquisition, intelligent robotics, and auto inspection.