Masters Theses

Date of Award

8-1993

Degree Type

Thesis

Degree Name

Master of Science

Major

Electrical Engineering

Major Professor

Mohan M. Trivedi

Committee Members

Reid Kress, T. V. Blalock

Abstract

For most robotic tasks, two different philosophies dictate how the applications designer approaches a problem. The first philosophy attempts to produce an application that is fully operational inside its task boundaries without human intervention. Examples of these types of applications are robot welders. These robots have the necessary knowledge base and/or intelligence to accurately perform the welding task. These types of systems are highly desirable for a large number of tasks; however, certain applications may require, due to the lack of technology or the task nature, intervention at critical steps in the process. Man-machine system design philosophies govern such situations. Telerobotic control is an example of man-machine systems design. In these systems, a remote operator interacts with a user interface, usually a master arm, and directs the movement of the actual robot in the work space. Most of the research in machine vision has been directed toward the development of fully autonomous vision systems. Such systems have been met with varying degrees of success depending on the complexity of the task and the skill of the designer. Logically, vision systems would benefit from the addition of a human operator into the "perceptual loop." Applying these concepts to perception, the research in the second section of this text describes a human-machine cylinder detection system. The design process includes a hierarchical task analysis decomposition and temporal-based "pick-and-place" studies. Stereo algorithms for interactive detection and localization are discussed. Finally, a basic error analysis is presented. In the third part of the text, the design of an automatic vision system for cylinder detection is presented. Using the Marr or Sobel edge detectors, the linear extremal features of cylinders are isolated via the Hough transform. Morphological filters (grey level) are then applied to enhance "feature clusters" in the Hough space. The Hough space is segmented and additional morphological filtering (binary dilation) provides the region growing necessary to connect the (possibly) segmented feature clusters. From this point, a recursive labeling algorithm isolates individual clusters. Centroids of these clusters are used to represent the linear extremal features present in a given scene. The set of linear features is segmented using unsupervised graph theoretic techniques into subsets that represent the candidates for stereo analysis. A thorough region-based stereo analysis discriminates between the actual cylinders and false alarms caused by the clustering algorithm. Again, a simple error analysis is outlined. Naturally, some design issues are common to both systems. The first section of the work involves implementation of the Two Planes stereo calibration algorithm using a second order approximation. It is followed by an extensive numerical stability analysis. Also some basic ideas used in the feature extraction algorithms use similar geometric constructs. These assumptions and their repercussions are fully developed. Lastly, a hardware design that was used in the implementation of a pipe-cutting system developed for the Department of Energy's Decontamination and Decommisioning project is outlined. The study is concluded with a comparative analysis of the design methods and a presentation of some empirical results.

Download
Files over 3MB may be slow to open. For best results, right-click and select "save as..."

Share

COinS