Research On An Efficient Rock-Mass Point Cloud Registration Method By Using N-Point Complete Graphs

With advances of computer technology,the application of numerical simulation technique in large-scale rock-mass constructions is prevailing and more indispens-able.To support this application,the most elementary part is to obtain the whole surface information of the whole rock-mass.Due to the limitations of point cloud acquisition instruments and environments,surface data acquisition possibly can not be finished in one scan,which means it needs several scans from different angles to acquire the whole surface information of the whole rock-mass.To fuse the separated scans into one is the main target for point cloud registration.It is evident that rock-mass point cloud registration is the basic of rock-mass 3-D reconstruction.The quality and efficiency of registration impact the whole rock-mass 3D-reconstruction directly.point cloud registration is a fundamental work for computer vision.There are a plenty of work that has been proposed for different scenarios from different angles.However,different from normal point clouds,there are certain qualities for rock-mass point clouds,which include massive volume,large scale,high resolution and complicated surface.These qualities combined together defy the application of existing registration methods.According to the qualities of rock-mass point clouds,this thesis introduces an ef-ficient rock-mass point cloud registration method by using n-point complete graphs.For the qualities of rock-mass point cloud,the method adopts both local statisti-cal information and surrounding geometric information to achieve global optimized results with high efficiency.The main work of this thesis can be summarized as follows:1.Give a comprehensive review of existing registration methods and category the methods into classes.Perform experiments with typical methods of each class.2.According to the qualities of the rock-mass point cloud,the thesis introduces a registration method based on n-point complete graphs.By employing Gaus-sian curvature,the method filters out most of registration irrelevant points,which brings very strong noise and outlier resilience.Then,cluster the pre-served points and calculate the centroids of each cluster.At last,build de-scriptors out of the centroids and their n-1 nearest neighbors and find the best matching descriptors to compute the transformation matrix.3.On datasets from public repository,experiments are performed under different conditions,including arbitrary initial poses,heavy outliers and noise,vary-ing point densities and low overlap,which cover most of the usage scenarios.Also,we compare the proposed method with existing methods to assess the performance on accuracy.Moreover,this thesis discusses the efficiency of the proposed method in detail.4.Discuss the extensibility of the proposed method on multi-view registration,point out the limitation and the direction of future work.In this thesis,we propose an efficient rock-masss point cloud registration method.By employing the strategy of cascading filtration,the method possesses high effi-ciency to dealing with massive data and strong robustness against noise and outliers.With the distinctiveness of descriptor built out n-point complete graphs,the method can achieve ICP comparable accuracy with arbitrary initial poses.The results of the experiment illustrate that the method allows globally optimized rock-mass registra-tion with high efficiency,which is a big advance of registration study on rock-mass point cloud.