The recognition and volumetric description of three-dimensional polyhedral scenes by analysis of Hough-space structures
Hans Peter Biland · Repository for Publications and Research Data (ETH Zurich) · 1987
This thesis describes a new method which uses only one digital image for analyzing two-dimensional representations of complex polyhedral scenes.The process maps the binary gradient image into the dual space by means of the Hough transform, which is very robust with respect to noise.Since the Hough transform locates straight edges rather than the vertices that most traditional techniques rely on, partial occlusions among the objects in the scene are no longer so difficult to handle.A subsequent local-maxima finder algorithm applied to the Hough-accumulator array produces a number of points in Hough space which correspond to localized edges in image space.This point set is then analyzed by comparing it with the dual structures of known object primitives, such as a parallelepiped, a tetrahedron, and a prism, stored in an expandable object library.If this topological matching is not immediately successful, i.e., the objects in the given scene are not mere object primitives but are composed of such, a decomposition process delivers the