Research Of Stereo Matching And3D Reconstruction Technology

3D reconstruction techniques is to obtain the objects’ depth information basedon multiple2D images with same scence and different perspectives with the stereomatching technology, or directly obtain the depth information based on the depthcamera, and the establish3D model with data optimizing and extraction. The stereomatching approach based on image segmentation is generally used due to its preciseregistration, strong robustness and high adaptability. In addition, Kinect is also widelyused to obtain the depth image directly due to its low price, suitability and otheradvantages. This article mainly carry out research on these two respects.The thesis first discusses the traditional segment-based stero matching approachesin detail. Aiming at the high mismatching rate of local matching and plane fitting inthe traditional algorithms, the initial disparity map is obtained by combining theimproved local matching method with the globle one. In the disparity refinementaspect, surface model is utilized to fit each segmented region. More over, a decisionregions merging rule is proposed. The simulation results show that the proposedalgorithm can obtain more accurate disparity map and realize smooth transition ofdifferent textures in same surface. More over, superior results can be realized in bothlow-texture and texture-occluded regions.The thesis finally discusses the basic principles of depth image acquisition withKinect, and the characteristics of its depth image (high noisy, edge information easy tolose). The missing depth information on edge cannot be effectively processed with thewidely used bilateral filtering method. So this thesis deal with the depth image bycombining the RGB information. In the depth edge missing information, the thesisfills damaged depth information with the image inpainting method. Simultaneously, inorder to accurately extract foreground object, we use the image segmentation methodsto locate boundary information in terms of extracting foreground object, and use ICPalgorithm to blend point clouds in terms of point clouds blending. The simulationresults show that the proposed algorithm not only suppress noise, but also fill missingdepth data effectively. More over, the point clouds blending are also achieved with theICP algorithm in this thesis.