Research On Point Cloud Analysis Method Based On Relation Network With Inverse Density Function

3D point cloud data contains abundant geometric shape and scale information.How to effectively analyze and process 3D point cloud data is a hot topic in the field of deep learning.However,unlike the image represented by the regular dense mesh,it is a challenging task to capture the hidden shape from the regular point.A point cloud analysis method is proposed to solve the above problems.Inspired by the inverse density function,the relationship shape convolution neural network is fused with the inverse density function to obtain the high-level relationship expression between the points,so as to capture the spatial layout of the points,and gradually perform the feature extraction from local context to global shape.Because kernel density estimation method does not use the prior knowledge about data distribution and does not attach any assumptions to data distribution,it is a method to study the characteristics of data distribution from the data sample itself,so the inverse density function is learned by kernel density estimation(KDE),which can compensate some inherent inhomogeneities in the point cloud data set under the condition of relatively uniform sampling rate of point cloud.The continuous function of multi-layer perceptron learning is weighted by inverse density function,so as to effectively improve the classification and segmentation accuracy of point cloud.Then,in the point cloud segmentation task,a deconvolution layer based on the convolution neural network of relational shape is added.Specifically,RSDeconv is composed of two parts:interpolation and RSConv.The feature information is transmitted from the coarse layer to the finer layer by using the reverse convolution layer.Experiments on the datasets of Model Net40,Shape Net and Scan Net are carried out,including classification,partial segmentation and semantic scene segmentation,which verifies the classification segmentation performance of the model.In the Model Net40 classification experiment,compared with Point Net++,the overall accuracy of the system is improved by 3.1%,and the accuracy is 1.9% higher than Pointnet++;In Shape Net partial data set,the average intersection and parallel ratio of its classes is 6.0% higher than Point Net++ when normal is also input.The instance m Io U is 1.4% higher than Point Net++.At the same time,we also verify our data set on the heterogeneous sampling data set,whether in point cloud segmentation or point cloud normal estimation,models all show good performance;In the scene data set of Scan Net,the m Io U is 13.7% higher than Point Net++,and the contrast experiment of whether there is inverse density function in different steps is carried out on Scan Net data set.The experiment shows that the inverse density function improves the segmentation accuracy by about 0.8%,and improves the performance of the model effectively.There are 31 graphs,14 tables and 51 references in this paper.