Information extraction about complex three-dimensional objects from visual data (bayesian, position estimation, robot vision, recognition)

Abstract

An approach to complex-object position estimation, using visual data, is discussed. It is assumed that the complex object is composed of a small number of chunks of planes, cylinders, and spheres. The approach consists of a number of processing stages. Recognition in image data of patches of planes, cylinders, or spheres in the 3-D world is discussed as a first step in complex object location and orientation estimation. Accordingly, an image is partitioned into small square windows, each of which may view a piece of a sphere, of a cylinder, or of a plane. Windows viewing pieces of two different surfaces are also considered. The windows are classified in parallel for recognition of content. New concepts and techniques include the approximation of images within windows by 2-D cubic polynomials, where each approximation is constrained by one of the hypotheses that the 3-D surface shape seen is planar, cylindrical, spherical, or consists of two surfaces; an approximately Bayesian recognizer based on these approximations that determines which one of a planar patch, a cylindrical patch, a spherical patch, or patches of two such surfaces is seen. The shape recognition is computationally simple. The second portion of this thesis is an approach to complex-object location and orientation estimation based on the information gathered from the windows. The information obtained from the windows is geometric, consisting of a collection of triplets specifying the location, orientation, and radius (if appropriate) of the primitive geometric shapes seen within each window. (Here techniques are used which can be found in the literature.) To model a complex object a similar geometric description is used. The complex-object position estimation approach is as follows. First, assuming that sensed and model surfaces have been appropriately matched (i.e., paired), least-squared error estimators are derived for the scaling, rotation, and translation that map the object-model surfaces onto the sensed object surfaces. Then simple tree-search algorithms involving minimal sums of squared geometric errors are proposed to simultaneously determine the matching of the model and sensed surfaces and estimate the position of the viewed object.