3-D ROBOT VISION: ACQUISITION AND INTERPRETATION (GRAPH THEORY, STRUCTURED LIGHT, PILE ANALYSIS, INFORMATION THEORY, THREE-DIMENSIONAL)

KIM LAVERNE BOYER, Purdue University

Abstract

This work addresses three major problems within the general problem domain of mechanized depth perception for intelligent robotics. These are: solving the correspondence problem in passive stereo, interpreting labeled line drawings built from three-dimensional data, and rapidly acquiring such data from a unique structured light technique. We first describe a novel approach to solving the stereo correspondence problem symbolically. Structural descriptions of two two-dimensional views of the scene are extracted from one of possibly several available low-level processes. An information-theoretic interprimitive distance measure is defined which automatically accounts for the statistical behavior of the image-to-image distortion. The structural description is subsequently extended to the parametric structural description representing the n-tuples in a relation as members of a parametric relation. A measure of relational inconsistency is then developed in the same spirit as the interprimitive distance measure using transition probabilities associated with the parameter values. These measures offer the advantages of graceful degradation, quality-of-fit defined on a continuum, and adaptability to available knowledge of the distortion processes. Next, we present new techniques for representing and interpreting the edge-vertex drawing of a scene consisting of a complex pile of convex objects. It is a new symbolic approach in which all pertinent scene information is distilled into a single abstract list. Pile analysis refers to the task of analyzing a scene to isolate distinct objects at the top of a heap and determine suitable grip points for a robot's end effector. The output of such a computation would serve as input to a reasoning system given the task of constructing a plan for the robot to unstack, sort, search, or otherwise deal with the pile of objects. Finally, we discuss an original strategy for rapid acquisition of the scene range map with color-encoded structured light. This technique offers several advantages including increased speed and improved accuracy. The possibility exists for the first time to acquire high-resolution range data in real time for modest cost. Grid to grid registration problems are eliminated, scene illumination is more uniform, and mechanical difficulties associated with the equipment design are significantly reduced.

Degree

Ph.D.

Subject Area

Electrical engineering

COinS