Stereo Matching Algorithm Based on Region Construction using Colour Segmentation

A novel stereo matching algorithm for accurate disparity estimation for all the regions in digital image is proposed in this paper. Instead of using a fixed size support window for calculating matching cost, a region will be constructed for the pixel under consideration and will be adaptively used to find the correlation or matching cost. The digital image is segmented into regions that collectively define the scene structure. The two major key contributions that are made by the algorithm to segment the image into regions; first for each pixel under consideration an upright cross is adaptively constructed, with four varying length arms that collectively decide the size of the region. The length of the four arms is adaptively calculated using the colour similarity and connectivity constraint. Second, using this local cross decision result, we dynamically construct a shape adaptive region by merging the horizontal and vertical segments of the cross. After the accurate image segmentation, variance based cost computation method is used to calculate matching cost among these regions for accurate disparity estimation among all the regions of the scene or digital image. The proposed algorithm is reliable, fast and yields expected results on Middlebury data set.