Towards a robust, passive stereo depth sensor with confidence and intensity guided anisotropic diffusion disparity refinement

Dooner, Matthew (2012) Towards a robust, passive stereo depth sensor with confidence and intensity guided anisotropic diffusion disparity refinement. MSc(R) thesis, University of Glasgow.

Full text available as:
[img]
Preview
PDF
Download (88MB) | Preview

Abstract

Stereo matching is the most common method for recovering depth information from two dimensional images. Despite the large amount of attention given to the problem it remains unsolved, and even robust methods sometimes produce noisy and inaccurate depth measurements. Existing disparity refinement methods can filter this output noise as a post-processing step at the cost of some fine depth detail. This work establishes a method to reduce noise while preserving the two-dimensional structure of the image through a modification of the well-known anisotropic diffusion technique. Weighting the amount of diffusion based on the edge strength of the intensity image rather than the edge strength of the disparity preserves a greater number of depth boundaries. The confidence of the disparity estimate prevents diffusing bad estimates into good es- timates and creates a stopping criteria for the diffusion process. Varied datasets provide validation of the technique; a dataset of our own design combined with two established benchmark datasets test the algorithm in varied environments. The performance of the author’s technique is compared against the technique which it improves on and the most closely related technique from recent literature. The author’s Confidence and Intensity Guided Anisotropic Diffusion (CIGAD) outperforms the other techniques in many cases and provides more reliable and robust results overall.

Item Type: Thesis (MSc(R))
Qualification Level: Masters
Keywords: stereo vision anisotropic diffusion robotics disparity refinement
Subjects: Q Science > QA Mathematics > QA76 Computer software
Colleges/Schools: College of Science and Engineering > School of Computing Science
Funder's Name: UNSPECIFIED
Supervisor's Name: Siebert, Dr. J. Paul
Date of Award: 2012
Depositing User: Matthew Dooner
Unique ID: glathesis:2012-4034
Copyright: Copyright of this thesis is held by the author.
Date Deposited: 26 Feb 2013 16:24
Last Modified: 28 Feb 2013 10:44
URI: http://theses.gla.ac.uk/id/eprint/4034

Actions (login required)

View Item View Item

Downloads

Downloads per month over past year