Research On 3D Laser Point Cloud Registration

3D laser scanning technology is widely used in the fields of protection of cultural relics,industrial measurement,deformation monitoring,and smart city construction,etc.,because it can express the three-dimensional information of objects comprehensively,efficiently,and accurately.Due to the limitation of the scanning angle of view and the complex geometry of the target,each angle of view can only get part of the geometric information of the target surface.In order to obtain complete information on the target surface,multi-view point cloud data needs to be registered together for point cloud segmentation,3D reconstruction,etc.Therefore,it is significant to study the point cloud registration method.For current point cloud registration algorithms,the registration process of complex point cloud data with large data volume,low overlap rate,and certain noise is prone to produce wrong corresponding points,slow convergence speed,and long registration time.This paper aims to study a high-precision and high-efficiency registration method,and pays attention to key point detection,key point matching,and registration quality evaluation methods.The research contents and innovations can be summarized as below:(1)A key point detection method combining multi-scale weighted normal projection changes and ISS(MWNP-ISS)is proposed.Key point detection is the most important step in the point cloud processing.This paper proposes a method of calculating the weighted projected mean difference of all points in the neighborhood space of the query point on the query point normal under different scale spaces,combined with the ISS key points.It can take into account the strong features and weak features in the point cloud data,and has good description and recognition.Experiments show that this key point detection method has stability and noise resistance.(2)A point cloud registration method based on local geometric features(MWNPISS+SHOT)is proposed.In view of the fact that the current point cloud registration methods are not suitable for point cloud registration with large data volume,low overlap rate,and noise interference,this paper designs a method that after down-sampling the original point cloud,based on MWNP-ISS method in this paper,combined with the SHOT descriptor,matching by cosine similarity and Euclidean distance,then use the point-to-plane ICP for accurate registration.Experiments show that this method can be applied to point cloud registration of different data sizes,and can quickly,efficiently and accurately register point cloud with low overlap rate and noise.(3)A registration quality evaluation method based on Super 4PCS is proposed.For the situation where the point clouds do not completely overlap after registration and registration fails,the common method of calculating the RMSE directly by the nearest neighbor distance will obtain inaccurate evaluation results.In order to improve the accuracy,this paper uses Super 4PCS to search for best 4-points congruent sets and use the neighborhood established as the overlapping area.After ICP registration,the point closest to the four points is further selected as the corresponding points to calculate the RMSE.Experiments show that the method proposed in the paper is feasible,and it is closer to the true value than commonly method.