Research On Multi-view 3D Reconstruction And Model Optimization Algorithms

Multi-view 3D reconstruction is an issue of common interest in both photogrammetry and computer vision,having the advantages of low cost and flexibility.After decades of theoretical and technical research,fruitful research results and software have been developed.However,due to its rich content and complex connotation,there are still many problems remaining unresolved,where further studies are required.The entire process of multi-view 3D reconstruction can be divided into several sections,mainly camera calibration,feature extraction and matching,sparse point cloud reconstruction,and dense point cloud reconstruction.Each part is implemented in this paper in turn,and corresponding solutions are also put forward for practical problems.Firstly,it is proposed to use the Exif information to quickly calculate approximate camera internal parameters.And the nonlinear optimization algorithm is exploited to further approach the optimal solution.It turns out that the solution is an available substitution of traditional camera calibration methods.Second,the matching schemes of point features and line features are improved in this study.For point feature matching,the scale of the feature point and the distance ratio are introduced to weight the matching results according to statistical research.Sampling for better matching points in priority can accurately estimate the outlier rate and effectively improve the efficiency of RANSAC.For line feature matching,it saves time and retains most accurate intersections when sampling lines with appropriate length and included angle.Multi-view reconstruction is based on the result of feature matching.In addition,the image sequence also needs to be planned.For drone images commonly used in this study,images are grouped according to their spatial proximity,and then in each group the sequence is determined through the initial matching results.The reconstruction model continuously expands by adding views sequentially.But with the increase of views,there are too many optimization parameters in the bundle adjustment method to minimize the residuals rapidly.To solve this problem,a comprehensive optimization scheme is proposed to reduce the parameters while ensuring accuracy.And on this basis,the sparse3 D point cloud is reconstructed and further encrypted by the PMVS algorithm.Finally,a watertight 3D surface is reconstructed by the Poisson algorithm.Last but not least,the reconstruction of buildings is improved through innovative integration of point and line features.By introducing the Character Number method,the dense point cloud is utilized to simplify line matching.Conversely,the original 3D point cloud can be encrypted by line features,and the geometric characteristics can also be effectively optimized.In the experiment,the models of building edge lines and their neighborhoods are promoted.In order to facilitate the application of algorithms,the multi-view 3D reconstruction and model optimization prototype software has been developed by integrating the above algorithms.