Surveyor Robotics Journal
email:
support@surveyor.com
web:
Surveyor Corporation
rss:
Subscribe
Archives
August 2010
July 2010
June 2010
May 2010
April 2010
March 2010
February 2010
January 2010
December 2009
November 2009
October 2009
September 2009
August 2009
July 2009
June 2009
May 2009
April 2009
March 2009
February 2009
January 2009
December 2008
November 2008
October 2008
September 2008
August 2008
September 2008
August 2008
July 2008
June 2008
May 2008
April 2008
March 2008
February 2008
January 2008
December 2007
November 2007
October 2007
September 2007
August 2007
July 2007
June 2007
May 2007
April 2007
March 2007
February 2007
January 2007
December 2006
November 2006
October 2006
September 2006
August 2006
July 2006
June 2006
May 2006
April 2006
March 2006
Februray 2006
January 2006
Wed, 20 Aug 2008
Surveyor SVS (stereo vision system) - first experiments with automated color segmentation for stereo disparity calculations
While others are working on much more sophisticated stereo disparity algorithms using host computing power, I thought we might be able to build some basic matching functions into the Blackfin firmware using simple color matching for correlating objects between the separate views. The key to this approach is to segment the images into a minimal set of colors that can be used for simple image comparisons.
Color segmentation is an interesting research problem which has led to the definition of a variety of color spaces, as discussed here and here, as well as a variety of feature processing models, as described here. One of the goals in creating the SRV-1 Blackfin and the SVS was to provide an easily accessible platform for experimenting with the color segmentation challenge.
As described above, a simple approach would minimize the number of colors in the corresponding stereo images and then look for matches. Below, you will see the original stereo pair (scaled down for this display), a 16 color segmentation that uses 4 luminance and 4 chrominance levels (2 bits Y, 1 bit U, 1 bit V - "211"), and a 16 color segmentation that uses 4 U and 4 V levels (2 bits U, 2 bits V, and no luminance component - "022").
The 211 approach seems to provide reasonable feature definition, but correlation of colors between left and right images isn't actually so great. The 022 approach doesn't seem to be making very good use of the range of colors, but the feature correlation almost looks usable. So the next step is to find a segmentation of the UV-only space that more efficiently maps actual colors. Along these lines, we divided the U and V color spaces into 16 segments each, and then built a tool to display the U vs V histogram in 2D.
The next step is to define a segmentation that looks for clustering or local maxima in the 2D histogram. So stay tuned ...
Posted Wed, 20 Aug 2008 19:39 | HTML Link | see additional stories ...