Dense stereo disparity map for video by sub-pixel dynamic time warp algorithm

Abstract

Dense stereo disparity map is an image indicative of the distances to the objects in the view field, constructed from the binocular stereo cameras, which can be used as robotic vision sensor etc. In order to determine the disparity, pixel correspondence between the right and left images needs to be found. A simple camera pose to keep the lens axes of two cameras parallel makes the stereo matching problem to be 1D constrained in the same raster scan line. The dynamic time warp (DTW) algorithm of the dynamic programming method efficiently solves the problem of stereo matching to match the right and left raster profiles to the accuracy of one pixel distance. In order to improve the resolution of distance, the paper proposes a method of subpixel disparity estimation that uses the cross-correlation between two local image profiles derived from the phase-only cross spectrum. The subpixel disparity is measured from the peak shifting of the spline function that interpolates discrete samples of the phase-only cross-correlation. The method complements the stereo matching by the DTW and gives fractional corrections to the disparities found by the DTW. The paper demonstrates the implementation of the method to generate a dense disparity map which can resolve one tenth of a pixel distance.