Registration Of 3D Point Clouds Based On Structural Association

With the rapid development of 3D point cloud processing technology,3D point cloud processing technology has been widely used in various research fields.Due to the detected objects influenced by objects,environment,surface layout and acquisition methods in actual project,it is necessary to obtain point clouds from different angles many times Therefore,in order to complete the registration of point clouds with different perspectives,it is particularly important to choose a good registration method.In the paper,we consider the influence of each step in the point cloud registration process on the accuracy and efficiency of the registration results.Besides,the registration time is improved on the basis of ensuring the registration accuracy from different perspectives.The main work of this paper is as follows:Firstly,the point cloud structure-related feature descriptor has been studied.Aiming at the problem of large amount of data in point cloud,weak feature description ability and low feature computing efficiency,the paper presents a feature descriptor based on normal vector distribution and a feature descriptor based on statistical feature histogram.Among them,the former has high computational efficiency and can quickly extract the feature points in the point cloud.The latter comprehensively considers the statistical characteristics of point cloud normal vector distribution,depth statistical features and density statistical features,and can comprehensively describe the structure correlation characteristics of point clouds.Experiments show that the feature points extracted based on the normal vector distribution have high robustness.Using the feature descriptors based on the histogram of statistical features can accurately describe the structural relationships between the points in the search area.Secondly,the registration algorithm based on structure-related feature descriptors has been studied,and put forward the algorithm that a point cloud initial registration based on statistical feature histogram(SF)and an accurate registration based on ICP.In the initial registration,aiming at the problem of too large initial point cloud data,a variable-edge voxel grid algorithm is proposed to ensure the quality of point cloud and simplify the cloud.Aiming at the problem that the points near the feature points in the streamlined point affect the registration accuracy,a key point extraction algorithm based on the normal vector distribution feature is proposed in this paper.At the same time,the statistical feature histogram is used to enhance the description ability of the point cloud to solve the problem of loss of detail after the extraction of the key points.We borrow the idea of RANSAC algorithm to complete the initial registration of the key points of the two views.For the initial registration result,the final registration is done using the ICP algorithm.Results show that compared with ICP algorithm,this algorithm can correctly register the point cloud in any initial position.In addition,the proposed algorithm can improve the registration speed with similar registration accuracy in comparison to the registration algorithm using the FPFH feature,.Based on Visual C ++ 6.0 and QT5.6.3 development platform,the 3D reconstruction system is designed and implemented.The system can use the two-step registration algorithm proposed in the paper to finish the point cloud of the pcd data format high-precision registration,then the Delaunay algorithm is used to realize the three-dimensional reconstruction,and the results of the key points extraction and reconstruction are displayed on the system page.