Research On Preprocessing And Registration Of 3D Point Cloud Reconstruction Based On Kinect

3D point cloud reconstruction is widely applied to daily life under the drive of industry needs for video games,reverse engineering and cultural relic protection,etc.However,after a series of data processing such as point cloud acquisition,filtering,segmentation,simplification,feature extraction,registration and fusion,the complex reconstruction process causes the error to be gradually accumulated,which directly affects the reconstruction accuracy of model surface.Therefore,in order to reconstruct 3D point cloud model with high precision,various noises in the point cloud model have to be removed by the preprocessing of de-noising,smoothing and simplification and the different angles local point cloud have to be consolidated into a complete point cloud model by image registration or splicing.The related algorithms of 3D point cloud de-noising,simplification and registration are deeply studied and designed in this paper,further experimental verification and comparative analysis are carried out.The main studying contents and related work are as follows:1.Aiming at the influence problem of 3D point cloud data noises on the reconstruction,a point cloud model de-noising algorithm based on method library of the pass-through filter,the statistical filter,the radius filter and the improved bilateral filter is proposed.Firstly,the target point cloud models are segmented by the pass-through filter.Next the large-scale noise within some distance of model body is first removed by the statistical filter and the radius filter,furthermore,the small-scale noise adhesion to model surface is smoothed by the improved bilateral filter.Then,the de-noised point cloud data is reconstructed by greedy projection triangulation algorithm with triangular mesh.Finally,the de-noising and smoothing accuracy is evaluated by the visual effect.2.Aiming at the problems of high density,long reconstruction time for scattered point cloud data,a new method based on uniform simplification algorithm of scattered point cloud data is proposed in this paper.Firstly,a k-nearest neighbor voxel grid is built by VoxelGrid class in filter module of Point Cloud Library.Then the center of each voxel grid is established by the normal estimation and mean distance of the k-nearest neighbors,and the point cloud model is simplified by these centers or the points near the centers replacing all point cloud data in the voxel grid.Finally,the simplified point cloud data is reconstructed by greedy projection triangulation algorithm with triangular mesh and compared with random sampling algorithm from three aspects of the accuracy,the simplicity and the speed.3.Aiming at the problems of long reconstruction time,low accuracy,slow convergence and error matching corresponding points,a point cloud registration method based on the ISS feature points combined improved ICP algorithm is proposed.Firstly,the feature points of point cloud are extracted by using SIFT algorithm,NARF algorithm and ISS algorithm respectively,and the ISS algorithm is established as the optimal extraction results.Next,the optimal extraction results are described by the FPFH algorithm and the frequency of the feature points near the point cloud data are directly and apparently showed.Then,a better initial position of two different angles point cloud is obtained by comparing RANSAC rough registration algorithm and SAC-IA rough registration algorithm.Finally,the corresponding point on the search is accelerated by the k-d tree nearest neighbor search method to improve the ICP fine registration efficiency.