Research On Algorithm Of 3-Dimension Reconstruction Based On Multi-View Image Obtained From Monocular Camera

3D reconstruction technology is an important method to obtain object models and 3D scenes,which is widely used in industrial detection,medical treatment,virtual reality and many other fields.Traditional 3D reconstruction methods are greatly affected by the environment,such as laser scanner is expensive,and can not obtain the texture information of target.In contrast,the 3D reconstruction method based on the feature points of images only need to input images from different angles.This method has the advantages of low cost,small scene limitation,and can recover the target accurately.Monocular multi-view 3D reconstruction mainly refers to use the monocular camera to rebuilt 3D scenes from the 2D images and obtain real scenes.However,the existing 3D reconstruction algorithms,consume high time in feature point matching phase.Moreover,the obtained points are often not uniform and dense,with a large number of redundant points,and even have many holes on the surface.In view of these situations,the matching process is optimized.This paper mainly including three parts: camera calibration and the sparse reconstruction,the dense reconstruction,the surface reconstruction and the design of interface.1.Firstly,this paper extracts feature points from multi-view images,and compares different feature points matching algorithms in terms of time efficiency.Then,using non-linear optimization to obtain camera's internal parameters.Lastly,estimating the errors of projection to ensure the effectiveness of sparse 3D points.2.In the process of dense reconstruction,the improved PatchMatch algorithm is used to realize the iterative growth of point cloud.The Delauny triangulation is used to replace the random initialization.By introducing the random search,it is possible to obtain the correct parameters in the areas where the correct parameters do not exist.A new cost aggregation method based on NCC is proposed,which combines the iterative growth of feature points and the merging of depth map.And then,a method based on WLS was used to fill the holes.3.The uniform sample points of surface reconstruction are discussed,and expressing the surface reconstruction problem as the solution of Poisson equation.Then extend uniform points to non-uniform points.The method can fit the noisy data accurately but not over,and fill the holes reasonably.Finally,the object-oriented C++language is used to design the interface of 3D reconstruction.The median error of projection of sparse 3D points is less than 0.01,indicating high reliability.The number of dense points obtained by the algorithm in this paper is11 times that of PMVS on average,and the reconstructed scenes is completed.