Occlusion Detection of Dense Correspondence Fields Using Motion Statistics and Triangulation

Abstract

Despite progress made in the accuracy and robustness of the dense matching technique in past years, efficient occlusion detection remains an open problem. In this paper, we present a two-step occlusion detection method to remove false matches in dense matching fields. First, a statistical dense matching method is developed by considering the correspondence between the grids to identify most occlusion regions. Second, to handle the potential misjudgment match in the occlusion boundary, a double-threshold filtering method is first used to reduce the noise in the grid image, which ensures that the gradient operator can accurately extract the boundary grid in the grid image; then, misjudgment matches in the boundary grid region are corrected based on the triangulation with descriptors. The results of the experiments comparing the proposed method and existing occlusion detection methods by, respectively, using the MPI-Sintel and KITTI datasets' test sequence show that the proposed method has higher accuracy and better robustness.INDEX TERMS Dense matching, occlusion detection, motion statistics, triangulation. I. INTRODUCTIONDense matching aims at determining the dense correspondence between two consecutive frames. It plays an important role in image processing and has been widely used in many vision tasks, including dense stereo reconstruction [1], optical flow estimation [2], object recognition [3], object tracking [4], and image retrieval [5]. Despite an abundance of literature related to this topic, obtaining a reliable dense correspondence field remains a challenging problem mainly because of illumination change, repetitive patterns, geometric deformation, and occlusions. This paper addresses the issue of removing false matches caused by occlusions in dense correspondence fields.To obtain dense matching fields, the patch-based methods have been the most used approach. The goal of this approach is to find one or more nearest neighbor matches between image patches. A serious challenge in this approach The associate editor coordinating the review of this manuscript and approving it for publication was Dong Wang.