Recovering Shape & Determining Attitude from Extended Gaussian Images

ID
TR-85-02
Authors
James Joseph Little
Publishing date
April 1985
Abstract
This dissertation is concerned with surface representations which record surface properties as a function of surface orientation. The Extended Gaussian Image (EGI) of an object records the variation of surface area with surface orientation. When the object is polyhedral, the EGI takes the form of a set of vectors, one for each face, parallel to the outer surface normal of the face. The length of a vector is the area of the corresponding face. The EGI uniquely represents convex objects and is easily derived from conventional models of an object. An iterative algorithm is described which converts an EGI into an object model in terms of coordinates of vertices, edges, and faces. The algorithm converges to a solution by constrained optimization. There are two aspects to describing shape for polyhedral objects: first, the way in which faces intersect each other, termed the adjacency structure, and, second, the location of the faces in space The latter may change without altering the former, but not vice versa. The algorithm for shape recovery determines both elements of shape. The continuous support function is described in terms of the area function for curves, permitting a qualitative companson of the smoothness of the two functions. The next Section describes a method of curve segmentation based on extrema of the support function. Because the support function varies with translation, its behaviour under translation is studied, leading to proposals for several candidate centres of support. The study of these ideas suggests some interesting problems in computational geometry. The EGI has been applied to determine object attitude, the rotation in 3-space bringing a sample object into correspondence with a prototype. The methods developed for the inversion problem can be applied to attitude determination. Experiments show attitude determination using the new method to be more robust than area matching methods. The method given here can be applied at lower resolution of orientation, so that it is possible to sample the space of attitudes more densely, leading to increased accuracy in attitude determination. The discussion finally is broadened to include non-convex objects, where surface orientation is not unique. The generalizations of the EGI do not support shape reconstruction for arbitrary non-convex objects. However, surfaces of revolution do allow a natural generalization of the EGI The topological structure of regions of constant sign of curvature is invariant under Euclidcan motion and may be useful for recognition tasks.
File(s)