基于PointNet++的邻域特征增强点云语义分割方法  

A neighborhood feature-enhanced point cloud semantic segmentation method based on PointNet++

作　　者：李松[1] 张安思 伍婕 张保 LI Song;ZHANG Ansi;WU Jie;ZHANG Bao(State Key Laboratory of Public Big Data,Guizhou University,Guiyang 550025,China;School of Mechanical Engineering,Guizhou University,Guiyang 550025,China)

机构地区：[1]贵州大学省部共建公共大数据国家重点实验室,贵阳550025 [2]贵州大学机械工程学院,贵阳550025

出　　处：《激光杂志》2024年第7期174-179,共6页Laser Journal

基　　金：国家自然科学基金地区科学基金项目(No.52365061);贵州省科技计划项目(No.黔科合基础-ZK[2023]一般059);贵州省教育厅高等学校集成攻关大平台项目(No.黔教合KY字[2020]005);贵州大学引进人才科研项目(No.贵大人基合字(2021)74号)。

摘　　要：随着智能驾驶、机器人导航等以点云为基础的应用蓬勃发展,点云语义分割逐渐成为研究热点。然而,现有的点云语义分割方法存在局部特征提取不充分、特征融合不完整的缺陷。针对这些不足,提出了对应的解决方案。对于局部特征提取不充分的现象,通过嵌入邻域点的坐标、方向、距离等相关信息去关联邻域点的显式特征。对于特征融合不完整的现象,提出了一种最大池化与自注意力池化相结合的混合池化方法。网络架构基于PointNet++,并结合提出的局部特征提取和融合方法,在S3DIS数据集上的实验结果表明,与基线方法PointNet++相比,各评价指标都有不同程度的提高,证实了新方法的有效性和优越性。With the booming development of point cloud-based applications such as intelligent driving and robot navigation,semantic segmentation of point clouds has gradually become a hotspot of research.However,the existing methods for semantic segmentation of point clouds suffer from the shortcomings of insufficient local feature extraction and incomplete feature fusion.To address these shortcomings,we propose corresponding solutions.For the phenomenon of insufficient local feature extraction,the explicit features of neighboring points are associated by embedding the coordinates,directions,distances and other relevant information of the neighboring points.For the phenomenon of incomplete feature fusion,a hybrid pooling method combining maximum pooling and self-attention pooling is proposed.The network architecture in this paper is based on PointNet++and is combined with the proposed local feature extraction and fusion method.The experimental results on the S3DIS dataset show that the evaluation indices have been improved to different degrees compared with baseline PointNet++method,which confirms the effectiveness and superiority of new method.

关 键 词：三维点云 语义分割 特征提取 深度学习 

分 类 号：TN024[电子电信—物理电子学]