Research On 3D Point Cloud Segmentation Algorithm Based On Graph Attention Convolutional Neural Network

The current point cloud segmentation methods are mainly divided into multi-view-based methods,voxel-based methods and point feature-based methods.The accuracy of point cloud segmentation methods based on point features represented by Point Net is usually higher than that of multi-view and voxel methods.Therefore,this paper considers the point feature information of the point cloud to model the structure of the point cloud.Most of the existing point feature-based methods map the point cloud to high-dimensional features through multiple multilayer perceptrons,and use pooling operations to capture the features.The captured point cloud feature learning process is isolated and usually does not consider the points.The neighborhood relationship between clouds loses a lot of spatial geometric information.Not only that,because point clouds belong to irregular data in Euclidean space,and their characteristics are disordered and sparse,methods based on point features are often unsatisfactory in capturing local features of point clouds.For this reason,this paper proposes a three-dimensional point cloud segmentation algorithm based on graph GNN and GCN,aiming to solve the point cloud segmentation task by using graph method.Then the algorithm was improved and the attention mechanism was introduced,Propose a three-dimensional point cloud segmentation algorithm Point-Attention Net based on graph attention GNN and GCN.Point-Attention Net is mainly divided into two modules: expanded point cloud module and graph attention convolution module.Specifically,in view of the problem that the 3D point cloud cannot be directly applied to the traditional 2D image convolutional neural network,Point-Attention Net uses the graph neural network to model the point cloud,avoiding the waste of memory caused by the conversion of the point cloud to other forms.Aiming at the problem of sparse point clouds and difficulty in capturing features,Point-Attention Net proposes an expanded point cloud method,which aims to enhance point cloud data and expand the same point cloud set with richer semantic information.This method uses the nearest neighbor clustering KNN algorithm to The K parameter value expands the point cloud in a proportional discretization manner to realize the capture of different point cloud receptive field information for the same point cloud set.Aiming at the problem of point cloud neighborhood relations and spatial geometric information,Point-Attention Net proposes a graph attention convolution method to assign reasonable weights to point clouds and their neighbor point clouds,fully considering the importance of neighborhood information and spatial distribution information,and the attention mechanism It is a weight distribution mechanism obtained through feature learning.This mechanism will selectively pay attention to the same kind of point cloud tags and assign high weights,which are defined as the same type,while ignoring other types of point cloud tags and assign low weights,which are defined as heterogeneous,this method It greatly improves the segmentation performance of edge point clouds.Experimental results show that the proposed Point-Attention Net is superior to the latest existing methods.In the indoor scene data set S3 DIS,the overall accuracy is 89.33%,and the average Io U is increased to 64.62%.In the outdoor scene dataset Semantic3 D,the overall accuracy is 94.2%,and the average Io U is improved to 74.4%.In the Part Net point cloud fine-grained hierarchical data set,the average Io U of Point-Attention Net reaches 51.4%.It is ahead of Point Net++ and Point CNN;in the Shape Net data set,m Io U reaches 84.9%,ahead of the classic point cloud network Point Net and Point Net++,and achieves the best performance in the 9/16 point cloud category.Through the comprehensive comparison of 4 data sets,different angles,and the analysis of ablation experiments,the efficient segmentation performance of Point-Attention Net is verified.