M. Fiala and A. Basu
University of Alberta
Department of Computing Science
ph: (780)492-3978 fax: (780)492-1071
University of Alberta
Edmonton, Alberta T6G 2E8
Canada
{fiala, anup}@cs.ualberta.ca
http://www.cs.ualberta.ca/~fiala
Keywords: panoramic, catadioptric, image segmentation, feature extraction, line extraction, Hough transform, non-SVP
Abstract
Omni-directional sensors are useful in obtaining a 360 degree
field of view of a scene for telepresence, panoramic scene capture and machine vision.
An approach to obtain a panoramic view is to utilize a radially-symmetric, non-planar mirror
and a single image sensor. There are several proposed profiles for the mirror, but most
violate the Single View-Point criteria necessary to allow functional equivalence to the
standard perpective projection. This poses challenges for feature extraction that must be
met to make use of such non-SVP mirror profiles (such as spherical) that have other desireable
properties. Such a non-SVP optical system does not benefit from the affine quality of straight
line features being represented as collinear points in the image plane.
A new method to recognize the salient features of straight line segments with such optics is presented. Previous work addressing this need saw the developement of a modified Hough transform to facilitate the detection of horizontal and vertical line feature edges. Algorithms tailored to utilize this Panoramic Hough transform to robustly extract horizontal and vertical line segments are presented. Specifically a robust method for using this Panoramic Hough transform to sucessively identify and remove clusters that appear in the parameter space which correspond to straight line features is shown. Experimental results are presented to validate this model.