平面特征保持的城市三维点云高精度分类方法  

A high precision classification method for urban 3Dpoint clouds with planar feature retention

作　　者：黄俊杰[1] 郭仁忠 张舵[2] 王伟玺 李晓明[1] 谢林甫 HUANG Junjie;GUO Renzhong;ZHANG Duo;WANG Weixi;LI Xiaoming;XIE Linfu(Research Institute for Smart Cities,School of Architecture and Urban Planning,Shenzhen University,Shenzhen,Guangdong 518601,China;College of Sciences,National University of Defense Technology,Changsha 410073,China)

机构地区：[1]深圳大学建筑物与城市规划学院智慧城市研究院,广东深圳518601 [2]国防科技大学理学院,长沙410073

出　　处：《测绘科学》2023年第3期6-15,共10页Science of Surveying and Mapping

基　　金：国家自然科学基金青年项目(42001407);广东省基础与应用基础研究基金粤深青年项目(2019A1515110729);国家自然科学基金项目(11972371)。

摘　　要：针对基于单一尺度点簇的分类算法无法满足大范围城市三维点云高精度分类的问题,该文提出了一种平面特征保持的多尺度点簇城市场景三维点云分割分类方法。该方法首先以超体素为基元进行动态区域生长,将具有平面特征的超体素逐步合并,生成平面特征保持的多尺度点簇,然后基于多尺度点簇进行特征计算并使用随机森林分类器进行分类。为验证该方法的有效性,分别采用机载LiDAR点云和影像密集匹配点云进行实验,并与基于不同分类基元的分类方法进行对比。实验结果表明,与原始区域生长方法相比本文提出的分割方法得到的点簇更加准确,分割准确度达到97%,并且分割效率提升了一倍以上,同时分类准确度也优于其他方法,达到94.1%、87.5%。To address the problem that the existing single-scale-point-cluster-based classification algorithm cannot meet the high accuracy classification of large scale urban 3Dpoint clouds,this paper proposes a multi-scale point cluster urban 3D point clouds classification method with planar feature preservation.The segmentation algorithm performs a dynamic region growing,merging planar supervoxels into planar point clusters and generating multi-scale point clusters with planar feature retention.a random forest classifier will be used to classify the final point clusters.In this paper,experiments are conducted on LiDAR point cloud and MVS point cloud for different scenes and compared with classification methods based different primitives.The experimental results show that the point clusters obtained from the method proposed in this paper are more accurate,with a segmentation accuracy of 97%,and the segmentation efficiency is more than doubled compared with the original region growing method.The classification accuracy is also better than other methods,reaching 94.1%and 87.5%.

关 键 词：点云分类 区域生长 机器学习 三维建模 

分 类 号：P208[天文地球—地图制图学与地理信息工程]