Application Of Deep Learning Technology In Point Cloud Registration For Large-Scale Low-Overlap Scenes

In the point cloud data acquisition work,the 3D laser scanner can only collect the point cloud data on one viewpoint per measurement.In order to obtain the complete 3D spatial information of the measured scene,a suitable point cloud registration algorithm is needed to solve the relative attitude transformation between the point clouds collected from different angles and stitch these point clouds under the same coordinate system using the rotation and translation parameters.In the past decades,various point cloud registration algorithms have emerged,but there are still some shortcomings in the existing point cloud registration methods in the face of complex point cloud scenes with large data volume and low overlap.On the one hand,the existing registration methods cannot overcome the interference of environmental factors such as large point cloud data volume and low overlap,and it is difficult to maintain the robustness of the algorithms in complex point cloud scenes;on the other hand,the existing point cloud registration network based on deep learning technology has a complex structure and the algorithms are too computationally intensive,which seriously affects the efficiency of point cloud registration.To address the problems of large amount of point cloud data,low overlap and sparse point clouds in realistic complex scenes,this thesis proposes a deep point cloud registration network based on global pose feature perception to achieve fast and robust registration in complex point cloud scenes,and the main research contents are as follows:(1)Aiming at the problem of point cloud registration in large-scale low-overlap scenes,based on deep learning technology,a direct method point cloud depth registration network(TSR-Net)based on global attitude feature perception is proposed in this thesis.By studying the point cloud pose feature perception method,an attention mechanism-based point cloud pose feature perception module is designed,which contains a point cloud feature coding unit based on the fusion of point structure and voxel structure for efficient perception of point cloud pose features in large-scale fields.To alleviate the interference of redundant information in non-overlapping regions,an attention-based feature weighting layer is added to the feature perception module to further improve the point cloud pose features.By studying the point cloud pose prediction method,a point cloud pose prediction module is designed,which consists of three stages: rotation parameter prediction stage(R-stage),redundant information filtering stage(M-stage)and translation parameter prediction stage(T-stage),and is used to complete the prediction of point cloud pose transformation and output rotation quaternions and three-axis translation vectors.(2)In order to test the registration accuracy and registration efficiency of TSRNet,excellent traditional point cloud registration methods(RANSAC,FGR,Teaser++)and deep learning-based point cloud registration methods(RSKDD-Net,IDAM,Deep GMR)are selected in this thesis,and point cloud registration comparison experiments are conducted on two indoor datasets(3DMatch,S3DIS)and one outdoor dataset(KITTI Odometry)to complete a comprehensive testing work on the registration performance of TSR-Net.The experimental results show that the proposed registration algorithm can not only provide highly accurate and efficient point cloud registration results in large-scale indoor and outdoor scenes,showing better registration performance than other algorithms,but also maintain the robustness of the algorithm in environments with low point cloud overlap,point cloud noise,and changes in point cloud sparsity.In addition,TSR-Net has a strong generalization capability.(3)In order to test the point cloud registration effect of TSR-Net in the measured scenes,the point cloud data of the study area were collected in the field,and the registration performance of TSR-Net was tested based on the measured point cloud data and compared with the registration results of other algorithms.The experimental results show that the average absolute error of rotation of TSR-Net is only 1.931°,the average absolute error of translation is only 15.202 cm,and the average running time is 0.26 s.The registration performance is significantly better than that of PCRNet network and TEASER++ algorithm,and has higher registration accuracy and computing efficiency.In addition,when the point density of the measured scene changes,the TSR-Net registration can still maintain the robustness of the algorithm.Therefore,the TSR-Net point cloud depth registration network can achieve high accuracy and high efficiency of point cloud registration in the measured scenes.