Research On 3D Point Cloud Data Enhancement Based On Generative Adversarial Networks

With the popularity of 3D scanning devices such as Li DAR and Kinect cameras,the measurement and acquisition of 3D point cloud data has become easier.3D point clouds have been more widely used in many fields such as robot navigation and autonomous driving.However,due to the limitations of the measured object since the obscuration,reflection,transparency and the resolution and angle of the device,it often leads to the actual measured point cloud has sparse,local missing and other problems.To address the sparse and local missing problems of the actual measured point cloud data,this paper carries out the research of point cloud data enhancement based on generative adversarial networks,and the main work includes:First,to solve the problems of frequent neighborhood reconstruction and high computational complexity of traditional graph convolution generative adversarial networks,a theoretical analysis of point cloud feature transfer smoothness between adjacent graph convolution layers is conducted;using this feature transfer smoothness,a new locally shared graph convolution model is established to avoid frequent neighborhood reconstruction through local neighborhood sharing and reduce the time complexity of the network model;to solve the problem of locally shared graph convolution inter-layer To solve the problem of feature dimension expansion between layers of locally shared graph convolution,a feature expansion module based on the learnable transformation matrix is constructed;using the locally shared graph convolution model and the feature expansion module,an adversarial network for 3D point cloud generation with locally shared graph convolution is designed,and the point cloud generation performance of the proposed method is verified on Shape Net and Model Net datasets.Second,to address the problem that the upsampling module of the existing sparse point cloud data augmentation generation adversarial network fails to utilize the feature information of other points in space,the sparse point cloud upsampling module is constructed by using edge convolution to aggregate the structural information of neighboring points and make full use of the information between points;to solve the problem of poor feature extraction ability of the existing sparse point cloud data augmentation generation adversarial network,based on the self-attention mechanism and To solve the problem of poor feature extraction ability of existing sparse point cloud data enhancement generation adversarial networks,a point cloud local feature integration module is designed based on the self-attentive mechanism and edge convolution;using the local feature fusion module and the sparse point cloud upsampling module,a sparse point cloud data enhancement generation adversarial network with local shared edge convolution is designed,and the sparse point cloud data enhancement performance of the proposed method is verified on the Shape Net and Model Net datasets.Finally,to solve the problem of poor generalization of the existing locally missing point cloud data enhancement generative adversarial network,the inverse mapping method of generative adversarial network is used to perform the residual point cloud data enhancement,and the inverse mapping uses a trained point cloud generative adversarial network to find a hidden vector to achieve the best complementary effect;for the problem of uneven distribution of the point cloud after data enhancement,a new uniformized loss function is designed The qualitative and quantitative experiments show that our results have achieved better results in terms of distribution uniformity and complementary accuracy.