超点图框架下融合双向注意力机制的点云语义分割方法研究  

Semantic segmentation of point cloud by incorporating two-way attention mechanism in superpoint graph framework

作　　者：李国立 陈焱明 夏家康 邹新灿 LI Guoli;CHEN Yanming;XIA Jiakang;ZOU Xincan(School of Earth Sciences and Engineering,Hohai University,Nanjing 211100,China)

机构地区：[1]河海大学地球科学与工程学院,南京211100

出　　处：《南京信息工程大学学报》2025年第2期165-171,共7页Journal of Nanjing University of Information Science & Technology

基　　金：江苏省研究生科研与实践创新计划(SJCX24_0201);国家自然科学基金(42071440)。

摘　　要：针对点云语义分割中,传统的图神经网络算法存在监督精度要求高、节点标签传递只能单向、未考虑全局信息等缺陷,本文提出一种基于双向注意力机制的点云语义分割方法.首先,将点云超分割为超点并建立超点图,从而将点云分类问题引入超点图网络框架中.然后,利用双向注意力模块,交替关注超点,根据邻接超点的权重更新超点特征,实现信息的双向传递.与以往的图池化方法不同,本文同时引入最大池化和平均池化,并将池化特征结合.最后,使用公开数据集Semantic3D进行训练和实验.结果表明,本文提出的方法可以有效地对标注误差进行纠正,同时耦合局部特征和长程信息,数据集的平均交互比(mIoU)和总体准确度(oAcc)分别为75.4%和95.1%,相比现有方法体现出更完善的标签传递机制和更高的分类精度.To address the deficiencies of traditional graph neural network methods in point cloud semantic segmentation,such as high requirements for supervision accuracy,one-way only node label propagation,and neglect of global information,this paper proposes a point cloud semantic segmentation method based on bidirectional attention mechanism.Firstly,the point cloud is over-segmented into superpoints and a superpoint graph is constructed,thus introducing the point cloud classification problem into the superpoint graph network framework.Subsequently,the two-way attention module is utilized to alternately focus on superpoints and update their features according to the weights of neighboring superpoints,enabling the two-way information propagation.Unlike previous graph pooling methods,this study applies both maximum pooling and average pooling,and combines their pooled features.Finally,the public dataset Semantic 3D is used for training and experiments.The results show that the proposed method can effectively correct labelling errors while coupling local features with long-range information,and the mean Intersection over Union(mIoU)and overall Accuracy(oAcc)of the dataset are 75.4%and 95.1%,respectively,exhibiting a better label delivery mechanism and higher classification accuracy compared with existing methods.

关 键 词：点云语义分割 图神经网络 注意力机制 超点 图池化 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]