Research On 3D Reconstruction Of Multi-Scale Building Images

Multi-view 3D reconstruction has been a research hotspot in the field of computer vision.This research integrates computer graphics,computer vision and pattern recognition and other disciplines,and has been widely used in vision navigation,medical diagnosis,cultural relics protection and virtual reality.The 3D reconstruction effect of multi-scale image sets is far from the point where it can be applied,and the problem that the existing 3D reconstruction algorithms have a low sense of reality and large amount of calculation for the reconstruction of large-scale scenes such as buildings.This thesis discussed these existing 3D reconstruction algorithms.The 3D reconstruction of key steps were studied in depth and an improved algorithm was proposed to perform 3D reconstruction of multi-scale building image sets.Based on the algorithm,this thesis developed a 3D reconstruction software system.The system consists of three main steps: the Struction From Motion(SFM)algorithm recovers camera parameters and sparse point clouds,and the Multi-View Stereo(MVS)algorithm calculates the depth map corresponding to each view.And the Surface Reconstruction algorithm rebuilds the complete 3D model.The main research work of this article is as follows:(1)This thesis used drones and smartphones collectted images respectively,and used incremental SFM algorithm reconstructioned sparse point clouds of camera parameters and scenes.The SIFT algorithm is used to extract and match the features of the image,and this step is accelerated in parallel with the GPU.In the incremental reconstruction stage,the next best view selection scheme is proposed and the existing bundle adjustment is improved.Experiments prove that the SFM method has high robustness and accuracy in this thesis.(2)After the camera parameters are restored,a multi-view stereo algorithm can be executed to generate a depth map corresponding to each view.This paper proposes a twolevel view selection algorithm: At the image level,the global view selection identifies a set of good neighborhood images for stereo vision matching for each reference view;at the pixel level,the local view selection determines a subset of the neighborhood image,This produces a result of stable stereo match.Then,the depth map is merged into a globally consistent dense point cloud as input to the surface reconstruction phase.Because only a few of neighboring views are needed for reconstruction,it is easy to expand into the reconstruction of large scenes.Experiments show that this method can produce high-quality depth maps.(3)Surface reconstruction using floating scale surface reconstruction algorithm(FSSR).This thesis proposes to use the scale value of each sample point as an additional input,which avoids averaging the sample points with different resolutions.The interpolation and fitting are automatically adjusted according to the same scale and redundant information,and no interpolation operation is performed in areas where the geometric information is insufficient,thereby improving the accuracy of the algorithm.The generated dense point cloud is used as the input of the surface reconstruction stage.Finally,the moving cube algorithm is used to extract the final complete 3D model from the isosurface.The experimental results show that compared with VisualSFM,this method reconstructs more 3D points and the reprojection error is smaller.Compared with PMVS algorithm,the MVS algorithm of this thesis reconstructed the number of vertices of the dense point cloud is much more.Compared with the poisson surface reconstruction method,the FSSR method used in this paper is suitable for multi-scale image sets.The reconstructed model not only has high accuracy,but also has a better reconstruction of the model details.Therefore,this method is suitable for multi-scale image sets,and has better reconstruction effect for large-scale scenes such as buildings.