Research On The Registration Of Urban Buildings Scene Point Clouds

Terrestrial LiDAR Scanning(TLS),as an active remote sensing technology,completes three-dimensional measurement by transmitting pulse signal and receiving the signal returned from the surface of a target object,which has the characteristics of high speed,high accuracy,and large amount of data acquired in one time,and it has become an important method of spatial information acquisition.Buildings detailed modeling has also become a research hotspot based on TLS in recent years.Due to limitations of scanning view and object obstruction,the TLS needs multiview scanning to completely cover the buildings,and the acquired multi-view point clouds are under the local coordinate system.It is necessary to unify the different coordinate systems by point cloud registration.Domestic and foreign scholars have proposed many algorithms for point clouds automatic registration.However,the structure of urban buildings is complex.TLS data not only contains a lot of noise,but also contains symmetrical or incomplete structures.The local point clouds density changes significantly,and the number of point clouds is very large.How to improve the accuracy and efficiency of point clouds automatic registration in urban buildings scene is a problem that needs to be solved urgently.Therefore,this paper focuses on the point clouds automatic registration under the scene of urban buildings.The details are as follows.(1)Point clouds coarse registration based on ISS-SHOT feature points.At present,point clouds coarse registration is mainly accomplished by extracting the 2D or 3D features on the clouds surface,and the quality of 2D or 3D features determines the accuracy of coarse registration.In this paper,the three-dimensional feature descriptor Signatures of Histograms of OrienTations(SHOT)is applied to the coarse registration of urban building scene to describe the SHOT features of ISS keypoints which extracted from two point clouds.This results in a more accurate set of ISS-SHOT feature point pairs.And uses the Sample Consensus Initial Alignment to estimate the initial transform parameters.At same time,a point clouds preprocess was designed,including point clouds denoising and downsampling.Point clouds denoising and downsampling can ensure that the original point clouds geometry does not change and reduce the number of point clouds and improve the quality of point clouds.The results of multiple sets of data registration show that our algorithm improves the average registration efficiency by 59% compared with the traditional coarse registration algorithm while ensuring the accuracy of registration.(2)The 3D Normal Distribution Transform fine registration algorithm based on the Quasi-Newton optimization.The traditional 3D Normal Distribution Transform(3D-NDT)algorithm has higher accuracy and efficiency than ICP,a fine registration algorithm.However,the efficiency of 3D-NDT is not enough under a large number of point clouds such as urban buildings scene.This paper analyzes the 3D-NDT algorithm principle for fine registration,using the Quasi-Newton algorithm to optimize the second-order derivative process in the 3D-NDT algorithm,improve the efficiency.The results of multiple sets of data registration show that with the increase of the number of point clouds,the time consumption curve of the improved 3D-NDT algorithm is much lower than the traditional 3D-NDT algorithm.With the assurance of accuracy,the average registration efficiency increases by 88%.