Image-based Three-dimensional Model Reconstruction Using Global Structure From Motion

Three-dimensional reconstruction has always been a scientific problems in the fields of computer science,medical science,scientific computation,virtual reality and so on.And it also has been a hot issue in recent years.There are three main methods for3 D modeling: geometric modeling,range-based modeling and image-based modeling.Among the three,image-based modeling method has become one of the most important methods for 3D modeling because of its advantage of fleetness,convenience and low cost.The key step of image-based reconstruction is to generate the 3D model from 2D images by Structure from Motion(SFM).Compared with the general SFM algorithm,the global SFM algorithm has the advantage of uniform error distribution,making it a popular research direction in recent years.In this thesis,global SFM method and other methods related to three-dimensional reconstruction are studied,and an image-based 3D model reconstruction pipeline based on open-source codes is actualized,using global SFM instead of sequential SFM.In this thesis,the laser scanning data are used as the reference data,and the global and iterative methods are used to complete the reconstruction,and the reconstruction results are compared.First,this thesis introduces the relevant knowledge of 3D reconstruction,including homogeneous coordinate and camera imaging model,laying the groundwork for subsequent research methods.Then,the feature points of the image are extracted by SIFT descriptor and nearest neighbor method is used to complete the feature matching.These matched points can be used to obtain the epipolar geometry,then the global rotation matrix and global translation vector can be obtained.In this process,the least squares method is used to complete the corresponding calculation.Finally,the Bundle Adjustment method is used to achieve the unified optimization of the results,and the sparse 3D point cloud is obtained.However,the gained sparse point cloud cannot clearly reflect the target,so it is necessary to be densified.In this thesis,PMVS method is used to complete dense reconstruction.Firstly,the PMVS method is used to select the seed patches and reconstruct the corresponding 3D information.Then,the seed patches will be extended to the neighborhood units to complete the reconstruction of patches until all patches are completed.Finally,the patches which is not satisfied with the geometric and illumination consistency constraints will be removed.Through theresearch and integration of open source algorithm,software and related libraries,a clear3 D model reconstruction pipeline is implemented.After the reconstruction,this thesis uses the data collected on campus to test the reconstruction method.There are three targets for reconstruction,namely,Innovation Center,School of Economics and Management and a rockery.At the same time,VisualSFM is used to complete the iterative reconstruction,and the reconstruction accuracy of the two results is evaluated.The results show that the global method is more uniform and accurate than the iterative method,and under some conditions,the cloud structure reconstructed by the global method is more complete.