Three dimensional model based object recognition.

An object recognition system is presented that uses CAD generated descriptions as the basis for automatic model creation. These models are then used to recognize and localize solid objects having six degrees of freedom. The sensory input is assumed to be in the form of dense range images resulting in object primitives with three-dimensional position information. Object primitives are assumed to be edge-based, mainly vertices, straight edges, circular arcs and circles. Objects may have any position and orientation and may also be partially occluded. Object vertices and circular arcs are the descriptors or primitives based on which the topological and geometrical models are built. The edges connecting the descriptors form the relationships that exist among them. These syntactic and geometric relationships are then exploited to tremendously reduce the complexity of matching process. A representational scheme based on Attributed Relational Graphs is introduced that can capture object topology. An inexact graph matching technique is presented that exploits the three-dimensional primitive coordinate information to very quickly index through a large library of object models to find the best match. The recognition process is performed mainly in two steps. First, the object topology is recognized and the view point partition determined. Second, the model/object correspondence is used to calculate the best transformation that maps the model to the object. Any remaining ambiguities will be resolved by a hypothesize and test scheme.PhDComputer scienceElectrical engineeringUniversity of Michiganhttp://deepblue.lib.umich.edu/bitstream/2027.42/162011/1/8906992.pd