Research On Feature Extraction And Registration Algorithm For Point Clouds

With the rapid development of digitization and information technology,the level of industrial manufacturing has been constantly developing.Its products require a large number of digital operations and three-dimensional scanning measurements in the aspects of design,production and testing.Due to the influence of the measurement conditions during the operation of the 3D scanning equipment,the acquired 3D point clouds information is not complete,and the measured point cloud data needs to be processed.Three-dimensional point clouds data registration is one of the most important part,and feature extraction is the key to feature-based registration,which can make the registration result more accurate.After a great deal of research and summary on previous works,the thesis will start with two aspects of feature extraction and registration algorithms,and research and improve the corresponding algorithms respectively to realize the feasibility and effectiveness of point cloud data registration.The thesis first introduces the research background,application prospect and research status of point cloud registration.Beginning with the processing of point cloud data,this thesis introduces the pretreatment process of point cloud data and the basic concepts of point cloud registration in detail,which pave the way for the follow-up research.Secondly,after comparing the advantages and disadvantages of algorithms based on normal vectors and curvature-based feature points extraction,a feature point extraction algorithm based on neighborhood features is designed.The algorithm needs to be weighted according to the influence of neighboring points of the sampling points to estimate the normal vector,which can better restrain the noise.The feature points are extracted respectively by the convexity and concavity in the neighborhood with different radii,and the intersection of these feature point sets is selected as the final feature point set.Doing so can effectively extract points with obvious geometric changes and provide higher quality key point sets for subsequent processing.Based on the analysis and summary of 4-Points Congruent Sets Registration Algorithm(4PCS algorithm),an improved algorithm is designed.The algorithm uses the feature of neighborhood to extract feature points,which reduces the time complexity of the algorithm and effectively restrain the noise.The low-dimensional feature descriptor and the double constraint of distance and angle are proposed to delete the error point pairs,which not only improves the accuracy of the algorithm,but also improves the execution rate of the algorithm.Finally,using D4 PCS algorithm for accurate registration.At last,C++ language is used to implement proposed algorithms in this thesis.Experiments are carried out by using a variety of point cloud data to verify the feasibility and effectiveness of the proposed algorithms in this thesis.