Research On The Technology Of The Three-dimensional Surface Reconstruction From Scattered Point Cloud

Image based three-dimensional geometric modeling is one of the advanced technology in the field of virtual reality. To reconstruct the three-dimensional environment model fast from point cloud generated by stereo images is an important topic of image based three-dimensional reconstruction. Aiming at the applications such as simulation in virtual environment, navigation of mobile robots, etc., the goal of the dissertation is to realize the three-dimensional reconstruction of irregular terrain. Under the support of three projects, including one of the National"863"High Technology Plan, the application of hypersurface evolution theory in the smoothing of point cloud models is researched firstly. And the Graphics Processing Unit(GPU) which is of great parallel capability is adopted to accelerate the process of smoothing. Then the multiresolution representation of the point cloud model is constructed. At last, the three-dimensional model of the surface is reconstructed. The work of this thesis is of great importance for the applications of image based virtual environment modeling.The main results and innovations of the thesis are as following:1. The concept of the feature curvature is proposed. As a new invariable geometric measure of surfaces, the concept describes the bend degree of surfaces more exactly. A feature curvature flow evolution equation is presented, based on which a point cloud model smoothing algorithm based on feature curvature flow is proposed. The experiment results show that the algorithm can denoise the model effectively, at the same time, has better capability of feature preserving than that of the classical smoothing algorithm based on average curvature flow.2. Based on Riemann geometry, the 2-dimensional hypersurface evolution theory is researched. Then gradient descend flow evolution equation is put forward. The evolution equation is applied to smooth the point cloud models, formed the point cloud model smoothing algorithm based on gradient descend flow. Experiments show that the algorithm can reduce the noise effectively, preserving the features.3. The fast smoothing methods of point cloud model are proposed based on GPU. Firstly, the trace transformation based accelerated algorithm, which is used to solve the smoothing parameters, is presented. Then, with the diagonal storage strategy for the extended matrix and the compressed storage scheme for the sparse matrix, two GPU-based fast smoothing algorithms for the point cloud model based on surface variance and based on feature curvature are implemented respectively. The experiments show that the performance of the point cloud model smoothing process is apparently promoted by the GPU-based algorithms.4. A multi-resolution modeling method of point cloud models is put forward based on feature curvature. Firstly, based on the hierarchical K-D tree space segmentation scheme, a point cloud model simplifiying algorithm that employs feature curvature as the segmentation criterion is presented. Thereafter, according to the generation of a hierarchical multi-resolution model constrained by a series of curvatures, a point cloud multi-resolution modeling algorithm based on feature curvature is presented. The efficiency of this algorithm is proved by experiments.5. In order to repair and reconstruct surfaces from genus point cloud models, a hole detection algorithm of genus surface is proposed. Thereafter, a Support Vector Machine (SVM) regression model based on weight attached to samplings is proposed. At last, a segmentation strategy based SVM surface reconstruction method is presented. The reconstruction method is that space segmentation technique is adopted to divide the models into several subsets with appropriate number of points, so that the reconstruction efficiency is improved. Experiments show that on the surface reconstructed by the method, the holes of the original genus surface are repaired satisfactorily.The proposed algorithms were utilized in the stereo vision based three-dimensional reconstruction application instances for the terrain models, and the feasibility and validity of the algorithms were proved. The research results of the dissertation have made actual contributions to the fulfillment of the funded projects.