Robust Registration And Adaptive Surface Reconstruction Of Point Cloud Data

With the rapid development of 3D data acquisition technology and the popularity of affordable acquisition devices,point cloud-based 3D model reconstruction technol-ogy has been widely applied to various areas,such as reverse engineering,cultural her-itage protection,3D printing,digital medicine and treatment,3D geography information system,virtual reality and augmented reality,etc.In point cloud-based 3D model reconstruction,registration and surface reconstruc-tion are two major problems.Firstly,most acquisition devices can capture only a single-view point cloud at a time.Thus,to reconstruct a complete model,point cloud data from different views should be transformed into a world coordinate system and the distortions generated during the acquisition process should also be eliminated,i.e.multi-view reg-istration.Automatic registration algorithms depend on the initial transformations,and the registration errors tends to be accumulated,which eventually lead to loop closure problem in multi-view registration.In certain cases,due to the calibration errors of ac-quisition devices or the acquisition of non-rigid objects,point cloud data may undergo non-rigid deformations;hence,non-rigid registration is required to improve the accu-racy of registration.Secondly,for the convenience of rendering and shape editing,the discrete point clouds are required to be reconstructed into continuous surfaces,e.g.tri-angular meshes.In the process of surface reconstruction,how to automatically and ef-fectively preserve the scale-varying geometric details in the original point cloud data has always been a challenging problem.Thirdly,data defects in point clouds,such as noise,outliers,holes and non-uniformly sampling,can also interfere with multi-view registration and surface reconstruction.In this thesis,we aim at robust and high-accuracy 3D reconstruction of real-world objects,and investigate the problems of point cloud denoising,multi-view registration and surface reconstruction.The main contributions are summarized as follows:.Two denoising algorithms for organized point clouds are proposed based on adap-tive moving least-squares(MLS),i.e.projective MLS along depth direction and implicit MLS along view direction.In order to handle depth-dependent noise and depth discontinuities,both MLS algorithms adopt a bilateral kernel function with an adaptive neighbor window to adjust local fitting weights.Experimental results show that the proposed adaptive MLS algorithms for organized point clouds are efficient,and they provide a more robust preservation of geometric details than image filtering algorithms and generate a more uniform point distribution than the MLS algorithms for scattered point clouds.·A hierarchical multi-view rigid registration algorithm is proposed.First,an undi-rected graph is built according to the overlapping relationship of the point clouds;then,the multi-view rigid registration is performed hierarchically on the edges,the loops and the entire graph.Moreover,a new objective function is defined to describe the loop closure problem;it improves the accuracy and robustness of registration by simultaneously considering transformation error and registration error.Experimental results showed that the proposed hierarchical optimization method can effectively eliminate accumulation errors and avoid to get stuck into local minima,and meanwhile it is more robust to initialization..A multi-view non-rigid registration algorithm is proposed based on multiple thin-plate splines.By defining the energy formula of thin-plate spline for multiple geometric objects,a new objective function used for global optimization called multiple thin-plate splines is proposed.By introducing the initial point positions as constraints,the trivial solution problem could be avoided,and each point cloud can maintain its original shape as much as possible.Moreover,the radial basis functions are incrementally added in the optimization process,which not only accelerates the solution but also improves the numerical stability.Experimental results showed that the proposed new objective function can effectively improve the registration accuracy of multi-view point clouds warped with low-frequency non-rigid deformations.·A multi-scale surface reconstruction algorithm is proposed based on a curvature-adaptive signed distance field.First,a signed distance field is built using an adap-tive octree,whose resolution is adaptive to the curvatures of the point clouds;then,the adaptive signed distance field is globally fitted with an implicit func-tion;finally,the zero level set of the implicit function is extracted as the output surface.A set of multi-scale B-splines are adopted as the bases of the implicit function,reducing the solution of the global fitting problem to the solution of a well-conditioned sparse linear system of equations.Experimental results showed that the proposed curvature-adaptive strategy not only reduces the running time and memory cost of solution,but also generates adaptive triangular meshes,which can preserve the scale-varying geometric details in the original point cloud data.