Research On Lidar Point Cloud Registration Technology For Positioning And 3d Reconstruction

With the popularization of computer and network,the world is gradually moving towards digitization and informatization,and services based on 3D high-precision locations have become a rapidly developing hot field.Such services include 3D positioning,navigation,3D reconstruction,simultaneous localization and mapping(SLAM),etc.,which play an important role in the fields of human health detection,architectural design,engineering survey,robotics and autonomous driving.In recent years,with the significant improvement of accuracy and cost reduction of light detection and ranging(LiDAR),its advantages in services based on 3D high-precision locations have gradually become obvious.LiDAR can generate a large amount of point cloud data and quantify the geometric and spatial characteristics of the surrounding environment surface at a fixed position,so as to construct or reconstruct the 3D scene.However,limited by the measurement range of the LiDAR and the physical occlusion effect of the laser itself,a single LiDAR can only describe the scene surface within a limited range.When it is necessary to describe a scene of large scale,or to describe a single object in an allround way,it is necessary to fuse the point cloud information obtained by the LiDAR at multiple locations.The point cloud data obtained by different LiDAR belong to their own coordinate systems,and the point clouds of these LiDAR are unified to the world coordinate system to complete accurate fusion,this kind of technology is called point cloud registration.In general,point cloud registration is a necessary step in LiDAR point cloud data processing,and its importance is evident.However,in the application of positioning and 3D reconstruction,due to the large deviation of initial transformation matrix and the small overlap of two point clouds to be registered,it is difficult for the current point cloud registration method to achieve an ideal registration success rate and accuracy.Therefore,this paper studies and designs point cloud registration techniques in localization and 3D reconstruction applications.In particular,for the two main scene types existing in such applications,structured and unstructured scenes,different methods are proposed respectively.The specific research contents are as follows:(1)Aiming at the problem of point cloud registration in structured scenes,this paper proposes a robust point cloud registration method based on scene plane primitives which utilizing the global information of point clouds.The method is based on a large number of plane primitives in the structured scene,and optimizes two steps of the basic process of plane registration scheme,plane segmentation,filtration and optimal selection,which enhances the descriptiveness of the plane,and properly uses the global information.First,in the plane segmentation step,a combination of growing region algorithm and random sample consensus(RANSAC)is used to extract planes in the scene,and the weighted indicator of plane points and area is used to filter a large number of redundant planes in the scene,leaving only high-quality planes information.Second,in the optimal selection step,the down-sampled scene point cloud coincidence degree is used as the metric to determine the optimal result from a large number of candidate transform matrix.The experimental results show that the registration success rate of the proposed method is 95%and 96%in the Apartment and Stairs datasets of ASL dataset repository,respectively,which is at least 10%higher than the traditional commonly used coarse registration and fine registration algorithms.The proposed method exhibits superior registration performance in structured scenes with small overlap.(2)Aiming at the problem of point cloud registration in unstructured scenes,this paper proposes a registration method assisted by a special cube-shaped target.This method uses a special fivesided cube-shaped target to assist point cloud registration,which provides sufficient structural elements for unstructured scenes,and performs registration based on this structural element,which improves the registration accuracy.The design of the target adopts a cube shape’s style,which not only provides a complete three-axis constraint for registration,but also considers the limited placement of the LiDAR in actual application,and has better portability than other traditional target-based methods.In addition,the method uses high reflection material region clustering to search for targets in the scene,and extracts and fully describes the target planes based on the prior information.The simulation results show that the registration success rate when the target planes are at an angle of 90 degrees is 24%and 23%higher than that at 60 degrees and 30 degrees,respectively.The experimental results using DJI Livox Horizon LiDAR in different scenarios show that,the proposed method has higher registration success rate and registration accuracy compared with several commonly used registration algorithms,which provides a stable and reliable solution for the point cloud registration problem in the unstructured scene.In summary,this paper mainly studies the point cloud registration problem in LiDAR positioning and 3D reconstruction applications,and proposes two high-precision and robust point cloud registration methods that can be used in structured and unstructured scenes respectively.The advantages of two methods are verified from different aspects through simulation or experiments.The research results provide some feasible registration strategies for LiDAR applications in these scenarios,and leave new research ideas for solving such kind of point cloud registration problems.