Research On Component Extraction Method Of 3D Point Clouds

Component extraction of an object represented by 3D point clouds is a basic task in object recognition and geometry processing.The extracted components have been widely applied in industrial modeling design,digital cultural heritage protection,space navigation,and so on.The existing methods for extracting components of an object rely on its local or global features,which always lead to over-segmentation,under-segmentation,or unreasonable segmentation boundaries.Therefore,the decomposition of point cloud objects is still a challenging topic in the field of machine vision.This paper aims to decompose an object represented by 3D point clouds into several semantic components.This paper starts with the estimation of normal vectors of point clouds,followed by extractions of local ridge and valley features and global skeleton topology.Then,the segmentation boundary of components can be constructed by the constraint of local and global features.The main contents and contributions are summarized as follows:(1)To solve the problem that existing normal estimations are hard to retain sharp features of the model,we propose a new neighborhood-based normal calculation method.This method utilizes the distribution of points to divide point clouds into smooth regions and sharp regions.Then,we design a neighborhood representation for points in each region based on their local geometric features,including the curvature,density,number of neighbor points,and residual to the fitted plane.Based on this representation,we can calculate the normal vector for each point in an object.Finally,according to the visibility of the outer surface,we design a normal adjustment method based on the bounding sphere,which improves the direction consistency of normal vectors.This method lays a foundation for the extraction of ridge and valley features,and the component decomposition.(2)To extract the ridge and valley lines for incomplete point clouds,we construct a shapepreserving neighborhood.The method starts with the recognition of missing regions by the distribution of points,then we construct a local neighborhood based on the nearest surface variations to calculate curvature features.Based on normal vectors and curvature distribution,potential ridge and valley points can be extracted by statistical analyses.Finally,we extract the feature lines by thinning,connecting,and interpolating potential ridge and valley points based on their nearest surface variations.This method can accurately describe the local surface details near the missing region.(3)To improve the descriptive performance of convex features,we propose a skeleton extraction method based on the mean point set of ridge and valley points.First,we calculate the mean point set of ridge and valley points based on their local neighborhoods,and then iterative extract the skeleton cloud by normal vectors of the points set.The skeleton points can be extracted based on the shape of the skeleton cloud.To keep the topological correctness,we introduce the regional relationship of skeleton points to constrain their connections.Finally,according to the geometry of a branch,we smooth the skeleton topology to ensure the centrality of the skeleton curve.(4)To solve the imprecision of segmentation boundaries in existing methods for component extraction,we propose a novel method for decomposing the point cloud by using internal and external critical points.First,we extract internal critical points for locating components by curvature changes and the skeleton topology.With the guide of internal critical points,external critical points are selected from the ridge and valley points for component segmentation.For the object without branches,according to the minimum rule,valley points are used to revise the internal and external critical points.Finally,under the constraint of internal and external critical points,we decompose an object into semantic components by skeleton-based piecewise labeling.