Research On Skeleton Extraction And Shape Analysis Of Point Cloud Models For Mechanical Digital Products

With the continuous development and improvement of computer technology,the digital design of mechanical products is widely popular,that is,the computer-aided design,CAD/CAM digital simulation technology are used to provide corresponding technical support for the whole process of mechanical product development,so as to promote the renewal of product design information and the improvement of design technology.With the close combination of reverse engineering and product innovation design,reverse engineering technology is more and more widely used in digital product design.Among them,the data acquisition and representation of product model,especially the acquisition and processing of point cloud model,has become an important research content in reverse engineering and a hot topic in the field of mechanical engineering design.In this paper,the 3D point cloud model is chosen as the research object.Focusing on the key problems in the digital design of mechanical products,the curve skeleton extraction,shape descriptor construction,similarity analysis,model classification and retrieval are deeply studied,and the application of retrieval system based on mechanical point cloud model is developed.The main research work is as follows:Curve skeleton is an important shape descriptor with many potential applications in digital product design,computer-aided design,computer graphics and visualization We present a curve skeleton expression based on the set of the cross-section centroids from a point cloud model and propose a corresponding extraction approach.First,the approximate calculation method of the centroid of the cross-section of the 3D point cloud model is given by the two force balance axiom,then an approximate algorithm of the distance field transformation of the 3D point cloud model is given,and then the hybrid feature points are obtained by combining it with the curvature.KNN and principal component analysis are used to estimate the normal vector of point cloud model,and the quadratic energy function is constructed to optimize the normal direction.By introducing the definition and calculation method of relevant facets and relevant points,these hybrid feature points are shifted by a certain step along the skeleton-guided normal direction to approach the local centroids.The tensor spectral clustering algorithm based on distance,normal vector and curvature is constructed,and the centroid point set is simplified by region growing clustering algorithm.Finally,they are connected into a primary connected curve skeleton.In addition,the primary skeleton is optimized by trimming,pruning and smoothing by Laplacian smoothing algorithm.the results of this method are compared with several state-of-the-art algorithms including the ROSA and L1-medial method in the aspects of correctness,centereddness,component-wise differentiation,invariant under different gestures or deformations,homotopic,generality,robustness,smoothness and computational efficiency,so as to evaluate the effectiveness and superiority of this method.In the shape analysis of objects,local self-similarity of 3D models is a fundamental problem.It has been widely used for partial matching and model recognition in computer graphics,visualization and other related fields.For a 3D point cloud model,we propose a self-similarity analysis method based on the three-dimensional feature tensor.Firstly,the approximation for calculating the shape diameter function(SDF)of a point cloud model is produced by using relevant facets and antipodal points.Then,spectral clustering is used to segment the model into sub-blocks of the model,the three-dimensional feature tensor is constructed from the shape diameter function,shape index and Gauss curvature matrix according to KNN neighborhood points.Next,the shape descriptor is constructed by using the tensor norm and its component norm and L1-median mapping,and the similarity measure is defined,and the self-similarity between the sub-blocks of the model is analyzed.Finally,the proposed method is compared with several popular methods(including partial matching and saliency detection),the experimental results show that the proposed method provides an effective shape descriptor and can efficiently discriminate the similar sub-blocks of a point cloud model.The research of 3D model similarity analysis and retrieval classification is a hot topic in reverse engineering of digital products.The traditional classification method is based on the feature descriptor designed by human.Its result and effect depend entirely on the selection and extraction of shape features.Different from the traditional methods,deep learning algorithm can make the machine automatically learn the features of objects and classify them.It has been mature and has outstanding performance in the field of image processing.However,the data processing method of 3D point cloud model is different from 2D image,and 3D data processing is more complex.This doctoral dissertation proposes a method of similarity analysis of 3D point cloud model under the condition of CNN(Convolutional Neural Network,CNN)distance.Firstly,the(sparse)point cloud model is preprocessed by coordinate transformation and densification.Then,the 3D model is colored by the viewpoint trajectory ball and the geodesic distance of the 3D model to obtain the 2D projection colored image data set of the 3D model.Then,the model features are extracted in the training stage and the benchmark vector is constructed by softmax function.Finally,the probability vector is obtained in the test stage.The Euclidean distance between the probability vector and the benchmark vector is calculated to get the final classification The training and testing data set of CNN distance comes from the geodesic distance on the point cloud model.After coloring the three-dimensional point cloud model,the two-dimensional color images from different perspectives are obtained.At the same time,the similarity of the three-dimensional point cloud model is found based on the distance determination.Experimental results show that the proposed algorithm makes the data processing easier,adds distance discrimination to improve the classification accuracy,and gets better results compared with other algorithms.Based on the above research methods and achievements,a retrieval system ZSTU-MMRS(Zhejiang Sci-Tech University Mechanics Model Retrieval System)for 3D point cloud mechanical models(mechanical parts)is developed,there are three retrieval methods are chosen:Model retrieval based on global feature—curve skeleton,model retrieval based on local feature—tensor descriptor and model retrieval based on CNN distance measurement.We introduce the system structure and the main functions of the retrieval system,and gives the corresponding operation results to verify the feasibility and effectiveness of the principle and method of 3D model retrieval proposed in this doctoral dissertationFinally,we summarize the research contents and achievements of curve skeleton extraction,shape similarity analysis,model classification and retrieval of point cloud models for mechanical digital products,and prospect the future research work.