Research On Terrestrial Scanning Point Cloud Registration Based On Points Or Lines Congruent Sets In Urban Scenes

With the growing demand for refined urban management,the construction of smart cities based on digital twin is in full swing.Li DAR scanning technology can directly obtain high-precision 3D information of urban scenes,which can provide powerful data support for tasks in the construction of smart cities,such as model construction,virtual simulation and scene analysis.The Li DAR point cloud registration technology aligns urban 3D point clouds collected from different perspectives,different characteristics and different platforms,providing great technical support for complete 3D data guarantee in the construction of smart cities.However,Li DAR point cloud registration in urban scenes faces challenges such as high noise,moving object outliers,lack of occlusion,and small overlap in urban scenes.Furthermore,urban buildings contain a large number of symmetrical structures and repetitive components,which make features associated with ambiguity in point cloud registration,leading to inaccurate registration result and even wrong registration.In view of the above key problems and challenges,this thesis focuses on how to robustly,efficiently and accurately register point clouds in urban scenes.First,the existing point cloud registration methods were systematically summarized,the advantages and disadvantages of the existing methods were analysed,furthermore,experiments are conducted to explore the applicability and limitations of different methods in urban point cloud registration,and according to the characteristics of urban point clouds,the4-points congruent sets registration method based on multi-scale feature matching,the2-lines congruent sets registration method,the optimization method of 2-lines congruent sets were proposed for urban point cloud registration.The specific achievements obtained in this thesis are as follows:(1)In this thesis,the existing point cloud registration methods were classified and summarized according to different association modes,we summarized the advantages and disadvantages of different types of point cloud registration methods,and analysed the applicability of various methods for urban scenes.(2)Aiming at the problems of low registration accuracy of 4-points congruent sets method,and ambiguity in point feature description and matching in large urban scenes,We proposed the MSF-Super4 PCS algorithm,which constructing and matching multi-scale feature based on 4-points congruent sets.The algorithm firstly register point cloud based on the optimal 4-points matching determined by the Super4 PCS algorithm,then constructs R radius search neighborhood at the 4-points matching,and further constructs multi-scale feature expression of key points in each neighborhood.Finally,the 4-points congruent matching is optimized by dense multi-scale feature matching.The local multi-scale feature description ensures the saliency of point feature expression,and solves the point feature ambiguity of description and matching caused by the symmetrical structures and repetitive components in large urban scenes.The MSF-Super4 PCS algorithm integrates the advantages of the point structure units registration and the point feature association registration,and adapts to the characteristics of urban scene for point cloud registration.The experiments showed that the algorithm can obtain higher-precision registration results than Super4 PCS algorithm.(3)Aiming at the problem of insufficient representation of geometric information of points structure units in urban scenes,We proposed a coarse registration method for urban point clouds named 2-lines congruent sets(2LCS).The algorithm uses two lines with different directions in urban scenes to construct the 2-lines structure units,which contain rich geometric information and are used as the features to associate point clouds.The random sample consensus(RANSAC)iterative framework is used to find the consistent matching sets of the candidate 2-lines units matching,which are then used to solve the candidate point cloud registration.Meanwhile,we exploit the direction information and position information of all lines to jointly constrain the consistent matching of the maximum structure of the scene,reducing the feature association errors in similar structure scenes.Finally,we obtain the optimal registration results.Compared with 4-points structure units registration methods,The 2LCS algorithm has stronger registration robustness in urban point cloud registration.The combined expression of two line features in urban scenes contains more geometric information,which can greatly reduce the number of candidate consistent matching and reduce the ambiguity of consistent units matching,and can overcome the challenges of point cloud registration such as small overlap,high moving object outliers,lack of occlusion,symmetrical structures and repeated components in urban scenes.(4)We improved the efficiency and accuracy of the 2LCS registration algorithm.Firstly,aiming at the problem that there are many redundant candidate matches in the2 LCS algorithm,which leads to the low registration efficiency,we proposed the semantic 2-lines congruent sets(2LCS),which uses the semantic information to associate and register urban point clouds.The S2 LCS algorithm adds the scene semantic perception module based on the 2LCS algorithm,each line feature is assigned a semantic label,we construct and match the 2-lines structure units with semantic information.The addition of semantic information reduces the ambiguity of feature description,improves the robustness of 2-lines units matching,and speeds up the matching process of congruent 2-lines units.Furthermore,the semantic information,direction information and position information of all line features are used to jointly constrain the maximum consistent matching of the scene structure,reducing the registration ambiguity in similar structure scenes.The S2 LCS algorithm achieves similar registration results compared to the 2LCS algorithm by using the semantic information,and has great registration robustness to the wrong semantic segmentation of the point clouds.Meanwhile,the S2 LCS algorithm can significantly reduce the number of candidate matches of congruent units,and reduce the time consuming of units matching and evaluation,improving the efficiency of algorithm registration.Secondly,in view of the low registration accuracy of the 2LCS algorithm,and the inaccurate registration of point clouds in urban scenes caused by high outliers,lack of occlusion and small overlap,we proposed terrain-invariant regions based iterative closest point algorithm(TIR-ICP),which is an improvement to the ICP algorithm.In the iterative registration process,outliers and missing inconsistent points are deleted,and only the points that have not changed in the dynamic urban scene are retained,and the stable points are then used to complete the point cloud registration.The TIR-ICP algorithm has the characteristics of simple design,great scalability and high registration accuracy.Furthermore,the TIR-ICP algorithm is insensitive to parameters setting,and has high registration accuracy and robustness for urban scenes with high outliers,lack of occlusion and small overlap.