High-density stereo image matching using intrinsic curves

Abstract

One of the most active areas of research in photogrammetry and computer vision is dense three-dimensional (3D) reconstruction of the environment via high-density image matching. This research interest is mainly driven by the growing popularity of unconventional imaging solutions such as images captured from unmanned aerial vehicles. With such data, problems like large disparity search space, occlusion, noise, and matching ambiguity become more pronounced. In this paper, we present a dense matching method to deal with these issues partially. The proposed method uses the concepts of intrinsic curves (IC) and derives useful matching information from their geometric features. First, we propose sparse disparity hypotheses for each pixel based on the orientations of the curves. These hypotheses are propagated to the neighboring pixels based on the proximity in the IC space; a solution which adaptively considers the intensity-variations in the neighborhood of a pixel to enlarge the set of its possible disparities. Then, a global matching energy function is formed and minimized, in which occlusions are explicitly taken into account based on curvature similarities of the ICs. The proposed technique is extensively tested on close-range and unmanned aerial images. Its performance is also compared to the state-of-the-art of dense-matching, such as hierarchical semi-global matching. Evaluations by the Middlebury Computer Vision Stereo Benchmark also show that the proposed technique results in average 3% error (percentage of pixels which are matched with more than 1-pixel error). The proposed framework could achieve high levels of accuracy (averagely 92%) as well as high efficiency by reducing the disparity search space up to 98% with an average confidence of 92% that the correct match, if existing, is still in the reduced search space.