Research On The Registration Technology Of Image Matching Point Cloud And Terrain Laser Scanning Point Cloud

With the rapidly development of the technology of surveying and mapping science and sensor in recently years,the equipment of UAV and 3D laser scanning are undergoing a transition from "professional level" to "consumer level".2D images with high-resolution and precise 3D point cloud is increasingly applied to the process of information construction,like architectural heritage protection and 3D city modeling,due to the acquisition flow become more efficient and convenient.The image matching point cloud generated by the technology of dense image matching with rich textures and obvious edge features,which can truly restore the 3D scene.However,due to the quality of original image and the robustness of the matching algorithm,mismatches point cloud existed in some areas.The point cloud obtained by Li DAR still have point cloud loopholes due to the constraints of the observation range of the instrument and the occlusion of the real environment,even if it can directly reflect the surface details of the target object.Therefore,in order to maximize the advantages of two point clouds and make up for the lack of data details or omissions in a single perspective and a single platform,it is necessary to registration of the cross-platform point clouds to realize the full range of spatial information expression in the target area.However,the accuracy of the current registration methods needs to be improved,because it either need to coarse registration of the image with point cloud,which involves time-consuming and laborious manual intervention processes and complex parameter solving processes,or complete the registration based on ICP algorithm directly,which ignoring the difference in point cloud characteristics.Therefore,in order to simplify the registration process and improve the registration accuracy,this article mainly study focuses on the registration technologies of image matching point clouds and terrain laser scanning point cloud based on the basis of previous research.Registration strategies are consisted of two part: coarse registration based on the boundary of building,fine registration by the point-to-plane ICP algorithm.The main research contents of this paper include:(1)Acquisition and comparison of the cross-source point cloud.After the images and laser point clouds are collected,generating the 3D point cloud of the image sequence based on the four kinds of image dense matching technologies.In order to get the best image point cloud,comparing the completeness and accuracy of the point clouds,which obtained by different matching methods,calculating the number,density,point accuracy and credibility of the four image matching point too.At the same time,the registration technology process of this article are designed after the data characteristics of the two cross-source point clouds are compared from the density,data accuracy,data defects and fusion requirements.(2)Extraction of building boundary as the feature primitive.In order to reduce the amount of point cloud calculation in the registration process.Extracting the point cloud of building boundary after clarifying the advantage information and complementary requirements of the two cross-source point cloud data.Applying the Cloth Simulation Filtering algorithm to filtering ground point cloud,filtering the vegetation point cloud of the terrain laser scanning point cloud based on the echo information of point cloud,and the vegetation point cloud of image matching point cloud based on the clustering number at the same time.Making the discussion according to the advantage of the point cloud of different buildings.Extraction of wall point cloud of modern buildings based on the algorithm of Region Growing.Extraction of boundaries of wall and window based on the Alpha-Shape algorithm after projected it to the parameterized plane.Segmentation of roof point cloud of the heritage buildings through the restrictions of height,collecting the boundary of roof based on the neighborhood feature of the point clouds.(3)Automatic registration based on the feature primitives of building boundary and the point-to-plane ICP algorithm.Constructing locally coordinate system by the Principal Component Analysis,calculating the rigid body transformation matrix by using the constraint of the center of mass and the Rodriguez formula during the progress of the coarse registration.After the initial pose is determined,in order to solve the problem of using the ICP algorithm directly for the fine registration may lead to local optimization if ignoring the characteristic difference of the cross-source point cloud.Therefore,fine registration of the separated ground point cloud together with the building point cloud.Finally,testify the feasibility and accuracy of the method by the experiments.