一种基于分割结果实现三维点云分类的方法  

A method for 3Dpoint cloud classification based on segmentation results

作　　者：赵江洪 窦新铜[2,4] 曹月娥 王殷瑞 黄先峰 ZHAO Jianghong;DOU Xintong;CAO Yue’e;WANG Yinrui;HUANG Xianfeng(State Key Laboratory of Geo-information Engineering,Xi’an 710054,China;School of Geomatics and Urban Spatial Informatics,Beijing University of Civil Engineering and Architecture,Beijing 102616,China;Key Laboratory for Urban Spatial Information of the Ministry of Natural Resources,Beijing 102616,China;Beijing Key Laboratory for Architectural Heritage Fine Reconstruction Fine Reconstruction Fine Reconstruction,Beijing 102616,China;State Key Laboratory of Information Engineering in Surveying,Mapping and Remote Sensing,Wuhan University,Wuhan 430072,China;School of Environment and Geographical Sciences,Shanghai Normal University,Shanghai 200234,China)

机构地区：[1]地理信息工程国家重点实验室,西安710054 [2]北京建筑大学测绘与城市空间信息学院,北京102616 [3]自然资源部城市空间信息重点实验室,北京102616 [4]建筑遗产精细重构与健康监测北京市重点实验室,北京102616 [5]武汉大学测绘遥感信息工程国家重点实验室,武汉430072 [6]上海师范大学环境与地理科学学院,上海200234

出　　处：《测绘科学》2022年第3期85-95,共11页Science of Surveying and Mapping

基　　金：国家自然科学基金项目(41601409,41971350);国家重点研发计划项目(2018YFC0807806);地理信息工程国家重点实验室开放基金课题项目(SKLGIE2019-Z-3-1);北京建筑大学市属高校基本科研业务费专项资金项目(X18063);北京市自然科学基金项目(8172016);武汉大学测绘遥感信息工程国家重点实验室开放基金资助项目(19E01);自然资源部数字制图与国土信息应用重点实验室开放研究基金项目(ZRZYBWD202102)。

摘　　要：针对现有点云分类方法忽略了同类点云的整体语义与每个点的单体语义相同的特点导致分类精度较低,鲁棒性较差,无法有效识别存在局部缺失的点云的问题,提出一种用于三维点云全自动分类的深度神经网络。采用局部代替整体的思想,设计一种识别结果筛选机制(IRS机制),构建IRS-PointNet++网络模型。该网络对每个点的分类结果进行二次统计筛选,在分类结果的基础上重新对点云的类别进行判定。通过在ModelNet40数据集上进行点云分类对比实验,设置不同的采样点数量与迭代次数,IRS-PointNet++的整体分类精度可达97.28%,相比现有最优方法提高了5.58%;在点云稀少的情况下,精度高于现有方法;在点云存在部分缺失的情况下,PointNet++会出现失效的情况,IRS-PointNet++的精度仍可以达到84.48%。实验结果表明:和现有方法相比,IRS-PointNet++网络表现出更高的精度和更强的鲁棒性。The existing point cloud classification methods ignore that the same kind of point cloud is similar semantic characteristics to each point.It leads to low accuracy and poor robustness of classification,and cannot effectively identify point clouds with local missing.To solve these problems,a deep neural network for automatic classification of three-dimensional point clouds has been proposed.The idea of replacing the whole by the part is adopted to design a Identification result screening mechanism(IRS mechanism)and construct the IRS-PointNet++network model.The network performs secondary statistical screening on the classification result of each point,and re-determines the category of the point cloud on the basis of the classification result.The comparative experiment of point cloud classification on ModelNet40data set is carried out,and different sampling points and iteration times are set.The overall classification accuracy of IRS-pointnet++can reach 97.28%,which is 5.58%higher than the existing optimal methods.And in the case of sparse point clouds,the accuracy is higher than the existing methods.Even if the point cloud is partially missing,PointNet++will fail,and the accuracy of IRSPointNet++can still reach 84.48%.The experimental results show that compared with the existing methods,the IRS-pointnet++network has higher accuracy and stronger robustness.

关 键 词：点云分类 深度学习 鲁棒性 PointNet++ 

分 类 号：P237[天文地球—摄影测量与遥感]