THREE-DIMENSIONAL ROBOT VISION WITH STRUCTURED LIGHT (COMPUTER, ARTIFICIAL INTELLIGENCE)
Abstract
Determining the identities, positions and orientations of randomly placed objects in 3-D space is of fundamental importance in industrial and navigational robotics. To precisely accomplish this, a robot must be equipped with a 3-D vision sensor and must be able to interpret 3-D vision data to acquire information yielding intrinsic characteristics of objects; we are particularly interested in making a robot capable of locating the least occluded object amongst a pile of objects and determining its identity, position and orientation for accomplishing tasks such as bin picking and heap sorting. In this work, we first describe various structured light scanning strategies for the purpose of acquiring 3-D vision data; some of these being particularly suitable for integration with robot manipulation. We then discuss the coded structured light systems for faster range mapping. We then propose different approaches to the interpretation of 3-D vision data acquired by structured light. More specifically, we will discuss surface extraction and classification, edge extraction and labelling, hole detection and diameter computation and axis curvature and cross-section parameter estimation. The technique of using Extended Gaussian Image (EGI) for representing 3-D objects is then discussed in some detail. We present a simpler approach for generating the center coordinates of a tesselated Gaussian sphere derived from the dodecahedron. Part of the discussion here will be devoted to hierarchical method for matching a model EGI with an experimentally determined EGI and to the use of the ratio between the actual surface area and the projected area to constrain the likely attitude of an object. We then show how the visible-invariant characteristics, such as the Gaussian and the mean curvatures, can be put to use for reliable 3-D shape description and recognition. We propose a scheme based on B-splines and derive operators that yield the Gaussian and the mean curvatures. We also present a 3-D shape recognition strategy based upon the Gaussian and the mean curvature histograms. Finally, we discuss how one can best use 3-D vision feedback for accomplishing robot tasks such as bin picking and heap sorting. In particular, we propose a strategy by which one can reliably and expeditiously determine the identity, position and orientation of the topmost object in a pile. In this scheme, we have incorporated different features that are obtained by structured light to improve the segmentation and identification process.
Degree
Ph.D.
Subject Area
Electrical engineering|Artificial intelligence
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.