Recognition of free-form three-dimensional objects in range data using global and local features

In this dissertation we advance the current state of the art in building, representing and recognizing free form objects from range data. We introduce two novel and effective methods for identifying objects with free-form surfaces. The techniques are built to exploit discovered shape structure from CAD models of objects to be identified. The first technique uses realistic rendering to generate synthetic data to train the system to identify the objects in real data. This view-based recognition technique obtains a 97% correct recognition rate for a 10 object database with real range data, while a 20 object database yielded a 99% correct recognition rate using synthetic range imagery. The second object-centered technique builds a hypothesis of object identity and location based on local surface features and their relationships. This second system was able to correctly identify sculpted objects from images where the objects occluded one another from the sensor. Based on only partial information the system was able to recover the objects and their locations in the image.