Research On Key Issues In 3D Point Cloud Completion And Recognition

With the vigorous progress of information technology,human perception of the world has developed from the traditional two-dimensional image to three-dimensional(3D)point cloud,especially with the rapid development and popularity of autonomous driving,intelligent robotics,virtual reality,and other technology products.We are increasingly convenient and fast to obtain and use indoor and outdoor scenes 3D point cloud data in our daily lives.The related demand is increasing,and the performance and accuracy requirements of3 D point cloud recognition are getting higher and higher.In this paper,based on deep learning technology,we focus on the tasks of point cloud completion,shape classification,semantic segmentation,and 3D face recognition in the 3D point cloud data processing and recognition,and the specific content and innovation points are as follows:For the point cloud completion task:(1)To address the problem that the traditional coarse-to-fine com-pletion strategy ignores the connection between different partial shapes,we propose a partial-to-partial point generation network for point cloud completion.The partial point clouds are encoded into global features and missing codes,and the novel view partial point cloud is generated conditioned on the view-related missing code.Two corresponding point cloud completion processes are designed to obtain various different partial point clouds,and the complete point cloud is aggregated from the partial point clouds.The experimental results on several datasets reach the leading level.(2)To address the problem of homogenization of global structure and local details in the existing point cloud completion methods,we propose a joint skeleton and surface point cloud completion network via geometry disentangling.The point cloud completion task is disen-tangled into two parts: skeleton completion and surface completion,and the corresponding decoder modules are designed separately.Finally,the complete skeleton and the complete surface are entangled into a complete point cloud with fine local surfaces and complete structure.The quantitative and qualitative experimental re-sults on several datasets outperform the existing methods.There are three aspects of 3D point cloud recognition in this dissertation:(1)Firstly,by analyzing the classical Point Net++ method,an improved density-aware point cloud sampling method is proposed to solve the problems of permutation variance and vulnerable to noise.A grouping-based Point Net++ network is fur-ther proposed which avoids information loss and additional upsampling operations.We achieve segmentation accuracy of 86.2% and 74.5% on the Shap Net Part dataset and the S3 DIS dataset,respectively.(2)Inspired by the multi-view recognition methods,a viewport group point clouds network is proposed for 3D shape recogni-tion.The local features are grouped and perceived based on the viewport and the viewpoint features are con-solidated by graph neural networks.The final global representation is aggregated by a novel attention-based feature aggregation module.Our method outperforms existing methods in both 3D point cloud classification and retrieval tasks.(3)To address the problem that the recognition performance of the model trained on the complete point cloud set degrades in the actual partial point clouds,a self-supervised feature learning-based partial point cloud recognition method is proposed.Constructing a self-supervised perspective transformation prediction task and an occlusion transformation reconstruction task to align the encoded features in the feature space reduces the domain gap between the complete and partial point clouds.The recognition accuracy of the partial point cloud is improved to the same level as the complete point cloud on the Shape Net-PCN dataset and the MVP dataset.Finally,a 3D face recognition method based on synthetic data is proposed in this dissertation.The GPMM model is used to generate a large amount of 3D face data with different shapes and expressions for network training.A 3D face recognition network model that can directly input point cloud data is designed together with a transfer learning strategy to solve the problem of domain differences between generated and actual faces and the problem of single-sample recognition in practical applications.The recognition rate of Rank-1 on the FRGCv2 dataset reaches 99.46%.