结合超体素与种子点优化的LiDAR建筑物点云分类方法  

A LiDAR building point cloud classification method combining supervoxel and seed point optimization

作　　者：廉飞宇 张良[1,2,3] 周涵宇 邓继伟 彭倩 LIAN Feiyu;ZHANG Liang;ZHOU Hanyu;DENG Jiwei;PENG Qian(Faculty of Resources and Enironmental Seience,Hubei University,Wuhan 430062,China;Hubei Key Laboratory of Regional Development and Environmental Response,Wuhan 430062,China;China Railway Design Group Limited,Tianjin 300251,China)

机构地区：[1]湖北大学资源环境学院,武汉430062 [2]区域开发与环境响应湖北省重点实验室,武汉430062 [3]中国铁路设计集团有限公司,天津300251

出　　处：《测绘科学》2024年第7期164-172,共9页Science of Surveying and Mapping

基　　金：天津市轨道交通导航定位及时空大数据技术重点实验室开放课题基金项目(TKL2023B13);国家自然科学基金资助项目(41601504);中国国家铁路集团有限公司科技研究开发计划项目(L2023G016);空天地一体化铁路智能勘测关键技术研究项目(23ZGSSSS00010);中国铁路设计集团有限公司科技研究计划重点课题(2022A02401008)。

摘　　要：针对机载LiDAR城市点云建筑物和相邻植被容易错分的问题,提出了一种基于W-OPTICS改进超体素分割的机载LiDAR建筑物点云分类方法。通过融合体素内部点云特征差异与密度特性构建自适应权值,提出W-OPTICS算法对初始种子体素进行偏移,使得优化后种子体素兼具空间局部相似性和高密度特性,解决在建筑与相邻地类邻接处种子点随机选取导致的跨边界错分问题,结合主成分分析算法聚合关键特征后,通过支持向量机实现建筑物分类。实验结果表明:针对Vaihingen数据集,本文方法的完整率、正确率、检测质量分别达到95.0%、94.0%、89.6%,相比原始超体素方法分别提高1.7%、4.2%、5.3%,建筑物与相邻地类错分和漏分情况得到明显改善;针对实测点云数据,本文方法各项精度指标优于经典超体素方法与TerraSolid自动分类结果,验证了本文算法在建筑物分类上的稳定性和鲁棒性。Aiming at the problem of misclassification of buildings and adjacent vegetation in urban point cloud of airborne LiDAR,a method of building point cloud classification in airborne LiDAR that improved supervoxel segmentation algorithm based on W-OPTICS was proposed.By integrating the differences in point cloud features within voxels and their density characteristics to construct adaptive weights,the W-OPTICS algorithm is proposed to offset the initial seed voxels.This optimization allows the seed voxels to possess both spatial local similarity and high-density characteristics,addressing the issue of random seed point selection at the junctions of buildings and adjacent land classes leading to cross-boundary misclassification.Combining the Principal Component Analysis(PCA)algorithm to aggregate key features,building classification is then achieved through the Support Vector Machine(SVM).The experimental results show that:For the Vaihingen dataset,the completeness rate,accuracy rate and detection quality of the proposed method reached 95.0%,94.0%and 89.6%,respectively,which was 1.7%,4.2%and 5.3%higher than that of the original supervoxel method.The misclassification of buildings and adjacent land classes was significantly reduced,and the omissionis improved to some extent.For actual measured point cloud data,the method described in this paper surpasses classical supervoxel segmentation methods and TerraSolid automatic classification results in all accuracy metrics.This validates the stability and robustness of the algorithm proposed in this paper for building classification.

关 键 词：点云 建筑物 OPTICS 种子点优化 分类策略 

分 类 号：P237[天文地球—摄影测量与遥感]