Research On 3D Point Cloud Registration Method For Buildings

Three-dimensional reconstruction has been widely accepted in the real-world application,such as AR/VR,3D Printing,industrial automation,relic perservation and techniques in health care,With the improvement of scanning precision,getting granular data from 3D instruments become more feasible.Therefore,the research on the methods of cloud data registration is cretical in the field of 3D reconstrction.Recent years,Deep Learing has gained great achievements in Image Understanding,Speech Recognition,and Natural Language Processing,especially in the application of 3D disorder point cloud.Point cloud,as the training data of deep neural network,should be preprocessed into specific specifications;thus the methods such as prejections on two-dimensional planes or bases on the division of meshs have been widespred.In recent years,there has also been a direct application of disordered 3D point cloud method,so that the information contained in the 3D point cloud can be fully learned and utilized by the model.This paper mainly focuses on the research about outdoor point cloud registration.Generally,such outdoor data contains massive number of buildings with obivous structures such as plane,line and angle.Therefore,it is possible to extract the line strctures form the point cloud of outdoor objucts,and then identify features of these strctural lines.It is alse feasible to find out some specific sections and extract its features according to neural networks,then training the Siamese network to find the correlation between the two point cloud data,thereby completing the identification.In this paper we prepose two algorithms for the registration of 3D point cloud.One is to extract the key framework of point cloud based on tis line strcture for purpose of simplification.To achieve the goal of simplifying the point cloud,we first screen line segments in one point cloud structure;in the point cloud structure,the line segment is screened according to a certain rule,and the Hausdorff distance,cosine similarity,and length information between the line segments are used as line segment characteristics,meanwhile keeping the rotation invariant by the use of distance sorting.The second is to use the neural network to characterize the point cloud,first identify point cloud blocks with special structures,and then screen points within these point cloud blocks to accurately locate the key points;train the siamese network to perceive local features and combine the network The two branch output features are matched to obtain the corresponding relationship of each pair of point cloud blocks for registration.