Research On 3D Reconstruction Based On Binocular Stereo Vision

As a core topic and popular direction in the field of computer vision,3D reconstruction based on binocular stereo vision has been widely applied in fields such as robot navigation and unmanned driving.The basic principle is to obtain three-dimensional coordinate information of the measured object from a two-dimensional image by simulating the perception of the objective world by the human visual system.In actual scenes,camera lens distortion,changes in lighting conditions,occlusion,depth discontinuity,weak texture areas,and noise greatly affect the effectiveness of 3D reconstruction.In response to the above issues,this thesis mainly studies the key technology of 3D reconstruction based on binocular vision,namely stereo matching algorithm,aiming to improve the quality of disparity maps generated by stereo matching algorithm and provide more accurate 3D information for 3D reconstruction.The main research content of this thesis is as follows:(1)A multiscale stereo matching algorithm based on non-local spatial tree filters.Specifically,this includes: 1)In response to the problem of poor performance of existing stereo matching methods in weak texture,deep discontinuous regions,and high texture regions with similar colors,this thesis proposes a non-local spatial tree filter in the cost aggregation stage,which jointly defines the edge weights between adjacent pixels through spatial affinity and pixel color similarity,and recursively aggregates the initial matching cost value from eight directions in the form of a triple tree.The filter can avoid excessive smoothing of the input image;2)To solve the problem that a single scale has poor ability to describe texture information,this thesis combines multiscale aggregation model with nonlocal spatial tree filter,and adds regularization term to enhance scale consistency.The experiment shows that the algorithm has excellent edge preservation ability,high matching accuracy in weak texture areas and deep discontinuous areas,and strong adaptability in high texture areas with similar colors.(2)A stereo matching algorithm based on multi feature fusion and improved content adaptive guided filter.Specifically,it includes: 1)In the cost computation stage,in response to the problem of excessive dependence of traditional Census transform on central pixels,which leads to sensitivity to noise,this thesis takes the local pixel contrast factor as the central pixel,and adaptively combines the improved Census transform with the color truncation absolute difference algorithm to improve robustness in areas with changing lighting conditions and weak texture;2)In the cost aggregation stage,an improved content adaptive guided filter is proposed to address the phenomenon of halo artifacts generated by the guided filter on sharp edges of the image.The filter uses multiscale edge perception weighting factors to effectively avoid halo artifacts and preserve edge information.Experiments have shown that this algorithm is more robust to changes in lighting conditions and has the lowest error matching rate in deep discontinuous areas and strong edge preservation ability.This thesis takes a real scene as an object and obtains their three-dimensional coordinate information based on the principle of binocular vision imaging.The experimental comparison of the three-dimensional measurement results of the two proposed stereo matching algorithms verifies the effectiveness of the proposed stereo matching algorithm.