Research Of 3D Reconstruction Based On Images

With continuous development of computer technology and computer vision,the 3D models of objects and scenes are constructed by computer.It widely apply in daily life,such as 3D games,film and advertising industry,protection of cultural heritage,virtual reality.The 3D reconstruction based on images of method is research hotspot.Obtaining model data based on images by simple devices is convenient and low-cost,meanwhile the model generation is automatic.In the process of 3D reconstruction with images,we get firstly the sparse 3D point clouds of target by structure from motion(SFM).Then we get dense point clouds data by cluster multi-view stereo(CMVS)algorithm and patch-based multi-view stereo algorithm.Finally,the point clouds are triangulated by this paper presents an improved Crust algorithm so that we can get intuitive 3D model.We reconstructed the 3D model of object with images in this thesis.The main contents are as follows:(1)To solve the low accuracy of feature point matching in multi-angle images.A new corner matching algorithm based on image sharpness is proposed in this article.Firstly,image edge is coarse detected by Canny edge detection.To reduce the influence of noise as much as possible,an improved detection algorithm is introduced.Edge contour is obtained by Canny edge and 8-Connected boundary tracking method.Then,the interested corner points are extracted by a corner detection algorithm based on contour sharpness.Finally,during the process of coarse matching of corner points,zero mean cross-correlation matching has established the matching relationship of one-to-many.The precise-matching pairs of corner points are obtained by an improved relaxation iterative method,which will reduce the number of iterations.The experimental results show that the speed of improved relaxation iterative feature matching method can increased by 35% ? 40% than the original relaxation iterative feature matching algorithm.At the same time,the matching accuracy is guaranteed.(2)We have obtained corner points and matching feature points pairs.The coordinates of 3D points are calculated by the geometric relationship between multi-view matching points.Then the space points are optimized by the Bundle adjustment algorithm,at last the sparse threedimensional point clouds are obtained.In order to reduce the time and space in dense matching,we firstly clustering classification of images by CMVS algorithm.Finally,we obtained dense3 D point clouds that expand and filter sparse 3D point clouds by PMVS algorithm.Through the acquisition of dense point clouds from five different test sets,the experimental results show that all test sets achieved the expectations and are effectively removed redundant noise.(3)To grid dense point clouds,dense point clouds data is discrete distribution and which can't subtly display 3D reconstruction model.The Crust algorithm is used to mesh point clouds in this thesis.It does not need to analyze and preprocess point clouds data.The grid result of denser point clouds is closer to the surface of the real object.For the sparse triangulation area,this thesis presents the improved Crust algorithm can split the larger triangle by the set area threshold,then we could get more triangular facets to get a more refined mesh surface.In the experiment,then five point clouds data were gridded respectively.The experimental results show that the Crust algorithm is meshed dense point clouds better,in addition the improved Crust algorithm is more detailed on the mesh surface of the model.Finally,we will obtain the fine,intuitive and the correct topology of three-dimensional mesh model.