The Research Of 3D Reconstruction Algorithm For Large Scale Scene Based On Multi-view Image

The 3D reconstruction of objects or scene in real-world is a common science problem and core technology in the field of computer vision,computer graphics and virtual reality.3D Reconstruction algorithms based on images model the scene or object by pre-acquiring an image or image sequence of an object or scene in real world and recovering 3D structure information from the two-dimensional image information.The method is relatively inexpensive,the acquisition of the data source is simple,and the dependency of hardware is small.As the improved of the acquired image resolution,the 3D reconstruction algorithms based on images has made a lot of new progress in the quality,robustness and computational efficiency of reconstructing effects.Structure From Motion(SFM)is a kind of way recovering structure information of scene and parameters of camera by using image set.It uses sequence image as input,outputing feature points of image,the result of feature matching,camera parameters and sparse 3D point Cloud.In this paper,some defects in the conventional SFM algorithms are improved.This paper comprises three parts: the first part proposed a multi-model geometric verification strategy that increase the number of feature match in the case of ensuring accuracy with appropriate geometric relation,then improving the completeness of 3D reconstruction.The second part proposed an efficient and sampling-based triangulation method with reference to the existing multi view triangulation methods,which can robustly estimate the 3D coordinates of all points in the feature track containing the outliers.The third part proposed a method that exploits the inherent characteristics of incremental structure form motion and dense photo collections for a more efficient parameterization of bundle adjustment by clustering redundant cameras into groups of high scene overlap.The use of an iterative bundle adjustment,re-triangulation,and outlier filtering strategy improves the completeness and accuracy of the incremental reconstruction effects and eliminates part of the errors due to drift in the process of incremental reconstruction.In the experimental part,we implement this algorithm.Finally,the results of using two data set obtained from the Internet to reconstruct the scene or object implied in the image are presented and compared with the reconstruction results of using ViusalSFM.The experimental results show that the proposed method has good application prospects for the completeness,robustness and accuracy of 3D reconstruction.