3D Point Cloud Classification Under The Synergy Of Persistent Homology And Deep Learning

3D point cloud is a set of discrete data points representing 3D shapes and scenes in 3D space,which has important application value in automatic driving,digital city construction,indoor scene modeling and other fields.As the basic task of point cloud data processing,point cloud classification is also a hot research topic at present.Point clouds contain abundant shape and spatial structure information,but the existing classification methods based on deep learning have limited ability to obtain the shape and topological structure features of point clouds.Therefore,how to extract the features of point clouds more comprehensively has become a major challenge for point cloud classification.As a mainstream tool for topological data analysis,persistent homology is emerging with strong ability of "shape" characterization and topological structure feature extraction.However,the feature extraction methods of persistent homology have their own scope of application,and different methods of persistent homology feature extraction also have differences in accuracy.Based on this,this paper focuses on two problems: first of all,how to extract shape features and topological structure features of point clouds more effectively by using persistent homology.Then,how to integrate the features extracted by the persistent homology method with those extracted by the network,so as to supplement the existing features and improve the classification accuracy of the point cloud? The main work of this paper is as follows.Firstly,a multi-type persistent homology feature fusion model is proposed for point cloud classification.This model fuses eight kinds of persistent homology feature extraction methods with different application ranges,different data structures and different prediction accuracy,and obtains a feature extraction model with better effect and wider application range for the classification of 3D point clouds.In order to evaluate the performance of the proposed model,the machine learning classification method is used to compare the model with other feature extraction methods.The results show that the model reaches the optimal in the accuracy,recall,F1-score and other indicators,which reflects the rationality and superiority of the model.Secondly,in view of the limitations of deep learning in the description of spatial and topological features of point cloud,the multi-type persistent homology feature fusion model of the first work is introduced into the point cloud deep learning framework,so as to supplement the features extracted from the network.Experimental results on public dataset ModelNet40 show that the accuracy of this model is improved compared with that of the original model,which demonstrates the feasibility and effectiveness of the collaboration between the persistent homology feature and the deep learning feature.This paper innovatively explores the application of persistent homology theory in the field of point cloud classification,and designs a multi-type persistent homology feature fusion model for 3D point cloud classification,which can be directly applied to 3D point cloud classification problem with obvious topological structure.In addition,the results of point cloud classification on the public dataset ModelNet40 also show the validity of the fusion based persistent homology feature and network feature model.The work in this paper not only helps to improve the understanding of the characteristic information of 3D point cloud,but also provides some support for the research of related issues in the cross-disciplinary field.