Research On LiDAR Point Cloud Registration Method Based On Keypoint

With the development of sensor technology,there have been many 3D scanning devices that can efficiently obtain 3D point cloud data,such as Airborne Laser Scanning,Mobile Laser Scanning and Terrestrial Laser Scanning point clouds.Point cloud 3D reconstruction has been widely used in automatic driving,virtual and augmented reality,cultural heritage protection and other fields.As the key technology of 3D reconstruction,point cloud registration can be divided into coarse registration and fine registration.In the actual LiDAR point cloud processing,the current method mainly focuses on the coarse registration of the point cloud,especially the registration using the features of points,lines,and surfaces.The algorithm in this paper can be applied to the registration of LiDAR point cloud.In the experiment,Terrestrial Laser Scanning point clouds(TLS)data is used as the experimental object.LiDAR point cloud has the characteristics of high efficiency,high precision,and global positioning.However,the large scale(TB level),uneven density,severe occlusion,and large noise have brought great challenges to point cloud registration.Therefore,in response to these problems of TLS point cloud,this article mainly carried out the following two aspects of work:Firstly,for the large-scale and serious occlusion problem in LiDAR point cloud,a matching pair extraction strategy based on key point detection and feature description is proposed.Due to the large scale and complex scenes of LiDAR point clouds,it is difficult to determine the key points in point cloud registration,and there are a large number of singular matching pairs in the local feature description.Therefore,this paper first uses the surface information entropy method to extract a large number of keypoints in the LiDAR point cloud;secondly,the point feature description is constructed based on the theoretical method of the local coordinate system and Gaussian convolution;finally,the super-voxel method is used to extract matching point pairs which are applied to the registration of point clouds.The experimental results in LiDAR point cloud show that the proposed method has good results in matching pair extraction accuracy and time performance.Second,aiming at the problem of uneven density in LiDAR point clouds,an end-to-end point cloud fine registration framework is proposed.The current deep learning registration methods all hope to train a stable feature for registration through the network,but the point cloud in the real scene has different point cloud density due to different scanning devices.We hope to use the overall structure of the point cloud for registration.In this paper,an end-to-end point cloud fine registration framework is proposed for 3D point cloud registration.Experimental results on cross-source LiDAR point clouds show that the proposed method is robust to point cloud density changes.