Research And Application Of Key Technology Of Non-incremental 3D Reconstruction Based On Multi-view

The 3D reconstruction technology based on image is used to reconstruct the three-dimensional model of objects or scenes by using multiple images.Compared with the laser scanning equipment to obtain the point cloud data for model reconstruction,this method has no higher requirements for the camera equipment and is less constrained by the environment,and has been widely used in many fields.SFM is a key step in the image reconstruction method,and SFM can be divided into two kinds of incremental SFM and non-incremental SFM,according to the global consistency of the camera parameters.Incremental SFM depends on the selection of the initial image and the cumulative error of each bundle adjustment,which leads to low precision,but the non-incremental SFM reduces the error through the global uniformity of the camera parameters,and only needs to perform a bundle adjustment at the last time,and the time efficiency and precision are higher,it has high research value and application prospect.At present,the development of non-incremental 3D reconstruction technology is not very mature,and there are still many problems.In this paper,the non-incremental 3D reconstruction related technology is studied,and some solutions are given,including feature extraction and matching,robust parameter estimation,and the solution of global camera parameters.The main innovations of this paper are as follows:1.Research on the feature extraction and matching algorithm,an improved SIFT algorithm is proposed:(1)when the feature description is carried out,the rectangle area of the neighborhood size of 15*15 pixels is selected for calculation,so that the description vector is reduced to 72 dimensions,which can effectively reduce the calculation amount.(2)when the feature points match,the relationship between the nearest neighbor Euclidean distance and the matching accuracy of the matching points and the nearest neighbor distance ratio are considered.The relationship between the value of the matching and the correct rate of matching is used for the initial matching of the nearest neighbor Euclidean distance based on the matching point pair,the nearest neighbor and the next nearest neighbor distance ratio,which can effectively reduce false matching,and finally obtain two types of initial matching sets with different correct rates.2.Research on the parametric model estimation algorithm,combining the properties of RANSAC and AC-RANSAC algorithms,the paper proposes an image registration method for essential matrix estimation: for the two sets of initial matching points obtained by the improved SIFT method,different algorithms are respectively used to purify the initial matching set to ensure that as many false matches are eliminated as possible.At the same time,the root mean square error was introduced to estimate the accuracy of the AC-RANSAC algorithm and the Sampson distance minimum constraint method to determine the interior point set of RANSAC.3.Researching the methods of global consistency of camera,sparse point cloud reconstruction,densification of point cloud,gridding and texture mapping,a nonincremental 3D reconstruction system based on multi view is realized.