Speeding up Dense Disparity Estimation of Stereo Couples Thanks to Horizontal Intensity Gradient

This paper presents an original development of disparity estimation from stereo couples primarily based on the similarity of image gradient between both images. Each pixel of the left image is associated to right image pixels with same gradient and similar intensity, directly leading to a set of disparities equal to the horizontal displacement in the specific case of epipolar images. Disparity values are restricted to an a priori acceptable range generally known by the application. The final disparity value assigned to a left image pixel is obtained from the histogram peak of candidate disparity values associated to pixels in the vicinity of the left pixel. A dense disparity map is produced. The conception and development have followed a very fast processing objective, resulting in about 0.75 Mpixel of disparity values per second on a 2.33 GHz Intel Core 2 Duo CPU. The approach has been successfully applied to highlight buildings or verify their presence in several types of satellite and aerial images.