Research On Pre-Processing Technology Of Three-Dimensional Point Cloud

Along with the development of 3D laser scanning technology, point cloud has been applied in many areas such as movies, games, reverse engineering and cultural relics protection and display. Considering the uncontrollable outside forces and physical defect, the original point cloud has some noise points, so it should be cleaned before the follow up processing. The outlier noise points have difference geometric properties with the points that would be reduce the accuracy of point normal and curvature. These noise points of around the model are hard to distinguish, as they have the same geometric properties. The existing algorithms mainly aimed at these noises. To obtain the accurate attribute information as well as the high quality surface reconstruction, we study methods of de-noising and normal estimation. To directly show the experiment results,we develop a point cloud display system. The main research and work of this paper as follows:1. Outlier disturbance must be take into account when we compute the normal, that means the outliers should be removed. In order to filter the outlier, we making cell division to point cloud, building the maximum connected domain, then keeping points of the cell which in the maximum connected domain. KNN was constructed based on the remaining space grid, after that co-variance matrix of each point was build and analysis. Eigenvector of the smallest eigenvalue take place of normal of the point.2. Filtering algorithm always bring about character missing, owing to its low transparency. Projection algorithm rely on the fitting curve and surface to refresh the space coordinate. Due to the complexity of fitting curve and surface, the algorithm efficiency decrease with the model scale, and the model would be deformation along with the iteration and to deal with this new problem, we need take extra measures to keep the model volume. As mentioned above, we improved the existing kernel estimation algorithm, introduce the vector difference and weighted factor of area to keep sharp features and sparse regions of uniform distribute. Proposed algorithm couldkeeping model characteristics in this paper without iteration and model deformation.Experiment results proved the statement above compared with the original algorithm.3. In order to directly show results of the reference and our algorithm, we develop a display system based on the MFC/OpenGL. Basic functions we achieved including reading, displaying, translation and rotation. Relevant algorithms we implement including cell division, KNN, co-variance analysis, outlier filter, kernel estimation based on Euclidean distance and the improved kernel estimation that we proposed.