Research On Occlusion Boundary Detection Approach Based On Video And Depth Image

Occlusion phenomenon is both a geometric phenomenon and an optical phenomenon.In the field of vision research, occlusion phenomenon is ubiquitous and play an importantrole in promoting the development of the most vision technology. Occlusion boundary canprovide important information for the three-dimensional structure of the scene. At thesame time, the occlusion boundary detection plays an important role on many aspects,such as scene segmentation, shape extraction, depth estimation and so on. In this paper，onthe basis of the comprehensive analysis of research status at home and abroad, do anin-depth research on occlusion boundary detection approach on video and depth image.Firstly, video and depth image knowledge, definition of occlusion phenomenon andocclusion boundary are introduced, and the occlusion problem processing methods arealso analyzed, meanwhile, the basic idea of the machine learning classification method tobe used in this paper are summarized.Secondly, based on video, an occlusion boundary detection approach by combiningappearance, motion and edge structure cues are proposed. First of all, the edges of currentframe in a video are segmented to obtain superpixels and superpixels’ edges, and then thesuperpixels’ edges are decomposed into short line fragments; Afterwards, the occlusionrelated features of each line fragment are extracted by combining appearance, motion andedge structure cues and the extracted features are assembled to feature vector; In the end,the feature vector of each line fragment is inputted to the random forest to train and testthe occlusion boundary classifier.Thirdly, based on depth image, an occlusion boundary detection approach based ongraph-based semi-supervised learning are proposed. First of all, some labeled andunlabeled pixel in depth image are taken to build a connected graph; Then the similaritybetween the pixels are computed to serve as the corresponding weights of the graph; Atlast, based on the graph-based semi-supervised learning detect occlusion boundary.Finally, by experiment to verify the feasibility and effectiveness of the occlusionboundary detection approach in this paper, meanwhile, combining with the experimental results, the accuracy of occlusion detection are compared and analyzed between pre-existapproach and approach in this paper.