Date of Award

2006

Degree Type

Thesis

Degree Name

Doctor of Philosophy

Program

Electrical and Computer Engineering

Supervisor

Dr. Jagath Samarabandu

Abstract

Making a robot navigate autonomously is a grand challenge in the robotic and computer vision research community. We approach the autonomous robot navigation problem within the simultaneous localization and map-building (SLAM) paradigm. The SLAM problem examines the ability of an autonomous robot starting in an unknown environment to incrementally build an environment map and simultaneously localize itself within this map. In robot navigation, accurate mapping and location estimates in either structured or unstructured environments is an essential requirement for high level tasks such as patrolling, investigating, delivering and fetching by an autonomous robot. This thesis explores the field of computer vision based robot navigation. It presents not only theoretical investigation but also practical implementation of Kalman filter based visual SLAM algorithms. The work presented in this thesis spans several areas of engineering and computer science, from new computer vision algorithms to novel ideas of uncertainty analysis in robotic SLAM. The existing SLAM algorithms heavily rely on sophisticated sensing systems, such as the fusion of global positioning system, inertial sensor, and laser detection and ranging. We propose to solve the SLAM problem using monocular vision. We will show four principal contributions that allow us to use such a simple sensing system. The first is to extend the existing theoretical lower bound of map-building error by estimating a new numerical lower bound with fewer constraints. Our second contribution is to apply multiple view geometry (MVG) techniques resulting in a novel update scheme for landmark initialization. We use theoretical derivation to come up with a closed form solution for the error at each observation step and use numerical verifications to prove that the new scheme is able to reduce mapping error significantly. Third, we incorporate high level geometric information such as common geometric primitives to SLAM to enhance the robustness. Finally, the proposed landmark initialization algorithm is demonstrated and evaluated by implementing the full SLAM algorithm in an indoor environment using a monocular vision sensor. 111 Through our theoretical investigation and experiments, we show that robotic SLAM using a monocular vision system is feasible and promising.

Share

COinS
 
 

To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.