Liscano, Ramiro2006-07-282006-07-2819981998http://hdl.handle.net/10012/253This thesis investigates the use of Bayesian networks for the perceptual grouping of features extracted from images of intensity and range data. A formalism for the development of Bayesian networks for perceptual grouping is presented that is based on a decomposition by parts methodology for the structure of the networks, and compatibility functions for the computation of the conditional probabilities. There exists a very strong relationship between the formation of a causal network and the process used to decompose an object into its components. A set of guidelines are presented for the design of a Bayesian network for perceptual grouping. Compatibility functions are a measure in the quality of fit between a set of features to a model that represents a grouping among those features. This model is represented as a node in a Bayesian network and compatibility functions are used to compute conditions probabilities in the formation of the grouping based on attributes and relations among the features. Five different compatibility functions are presented as examples. These include an edge, surface proximity, planar surface, coplanar surface, and parallel surface compatibility functions. The edge and surface proximity compatibility functions are unique in that they rely on computations both in a 2-D image plane as well as 3-D space. These algorithms determine a polygonal approximation of a planar surface as well as a common virtual surface between two polygons in 3-D space. A unified representation of Bayesian networks and attributed hypergraphs, a Bayesian attributed hypergraph (BAHG), is developed that allows for the instantiation of multiple Bayesian networks. BAHGs are a subset to attributed hypergraphs allowing for the specification and maintenance of multiple Bayesian networks without creating multiple instantiations of nodes that represent the same event. This facilitates the grouping process among multiple surfaces since continuity is maintained among groupings that share common feature sets. Also, a graphical description of the grouping network can be used to guide the creation of the BAHG. This network can be recursively applied to the features extracted from the sensory data. It is this recursive nature of perceptual grouping that makes the BAHG suitable for computer vision and image understanding problems. Validation of the use of a BAHG is presented using a BAHG network for the detection of corners and continuity among adjacent surfaces, applied to several images of 3-D data points extracted from a portable range sensor.application/pdf4728176 bytesapplication/pdfenCopyright: 1998, Liscano, Ramiro. All rights reserved.Harvested from Collections CanadaDoctoral Thesis