Research On 3D Point Cloud Registration Of Buildings

One of the main features of digital cities is the establishment of three-dimensional urban models,and buildings are an important part of the city.When using a terrestrial 3D laser scanner to obtain point cloud for a complete building,it is necessary to scan the building in different viewing angles.The point cloud registration technology plays a key role in the reconstruction of buildings' point cloud.Therefore,this paper focuses on the registration of buildings' 3D point cloud in multi-viewpoints.The main contents are as follows: 1.A variety of multi-planar structure extraction algorithms based on RANSAC(Random Sample Consensus),regional growth,J-Linkage and co-clustering are implemented.The paper shows that the plane extraction algorithm based on co-clustering does not need to set priori parameters,and avoids the problems of over-classification and under-classification at the same time,which is suitable for the plane extraction of buildings' point cloud.2.A new method of point cloud registration based on buildings' planes is proposed.This algorithm uses the characteristics that buildings contain many planar structures.Firstly,using the co-clustering method to extract the building's planar structure,and then orthographic projection of these planar structures to generate two-dimensional pictures.The next step is using SIFT descriptors to match corresponding points of multiple 2D images formed by multi-site cloud,and associate them with the 3D coordinates in the point cloud.Finally,the registration of point cloud is accomplished using a three-point RANSAC algorithm and a least-squares method.Experiments show that the proposed method has high accuracy of registration and avoids the disadvantages of ICP registration algorithm that requires initial pose estimation and trapping into local optimal solution.Compared with the ICP registration,this method is more applicable to the registration of buildings' point cloud.3.A point cloud registration model based on two fixed points is proposed,which takes the position of the known scanning points as the constraint condition.Only two sets of corresponding points are sampled in the sample set,and the model parameters are solved by rotating the sampling points around the scanning point to reconstruct the corresponding points.Then,combined with the RANSAC algorithm,a two point cloud registration algorithm with a known scanning point position is designed.Through this algorithm,the optimal model parameters of the point cloud registration model based on two fixed points are solved and the point cloud registration is completed by the parameters.The experiments show that this algorithm can correctly complete the registration of building point cloud and it is suitable for continuous registration of multi-site clouds.