Dense stereo matching in depth discontinuity areas based on edge detection and nearest valid point search

Abstract

Dense stereo matching based on two viewpoint images is an important step in 3D reconstruction. In this study, a new disparity estimation algorithm for depth discontinuity areas is proposed. Dense stereo matching is implemented on the basis of the semiglobal matching algorithm. If the ground truth is known, then mismatched pixels are selected as those with a difference of more than one pixel from the ground truth. Otherwise, the left-right consistency checking process is used. Then, the Canny operator is used to detect image edges. For each mismatched pixel, a maximum total of eight nearest edge points exists in four pairs of symmetric directions. Basically, the process of searching for the nearest valid point proceeds along directions in the following sequence: horizontal, vertical, 45° oblique and 135° oblique. Specifically, for each pair of symmetric directions, the nearest valid pixel is searched along the direction of the distant edge point. If the disparity is successfully reassigned by the nearest valid pixel found, the searching process is terminated. Comparative results obtained using well-known benchmark datasets and unmanned aerial vehicle stereo images demonstrate the excellent performance of the proposed method.