多维特征GS-SVM地下空间激光点云分割  

Underground space LiDAR point cloud segmentation based onmulti-dimensional feature GS-SVM

作　　者：卫梦莎 龚云 张小宇 刘腾飞 WEI Mengsha;GONG Yun;ZHANG Xiaoyu;LIU Tengfei(College of Geomatics,Xi'an University of Science and Technology,Xi'an 710054,China;Yangling Vocational&Technical College,Yangling 712100,China;Kanjing Technology Development(Shanghai)Co.,Ltd.,Shanghai 201700,China)

机构地区：[1]西安科技大学测绘科学与技术学院,陕西西安710054 [2]杨凌职业技术学院交通预测绘工程学院,陕西杨凌712100 [3]瞰景科技发展(上海)有限公司,上海201700

出　　处：《测绘通报》2024年第9期117-122,共6页Bulletin of Surveying and Mapping

摘　　要：针对激光点云在弱光和低特征点环境的点云分割中过拟合问题,本文提出了一种基于网格搜索的多项式核最小二乘支持向量机(GS-SVM)点云分割算法,提取点云的多维特征,并分别对车辆、车道线和车库障碍物柱子等多维特征进行分类;采用多尺度精度评估指标验证特征选取的有效性并评估分割算法。结果表明,与文献中传统分类算法的柱子识别率80%和车辆识别率65%相比,基于多项式核函数的分类算法对柱子和车辆的识别率分别在75%和73%以上,提高了5%和8%;在使用其他两个核函数时,GS-SVM同样保持了优势。本文算法相对于常规算法有较强的稳健性,为其在弱光和地理特征点环境的点云分割问题提供了解决方案,丰富了激光雷达三维扫描的使用场景。Aiming at the problem that point cloud overfitting in point cloud segmentation in low light and low feature point environment,this paper proposes a GS-SVM point cloud segmentation algorithm that based on polynomial kernel least squares support vector machine and grid search.It s extracted to classify multi-dimensional features such as vehicles,lane lines and garage obstacle pillars.The multi-scale accuracy evaluation index is used to verify the effectiveness of feature selection and evaluate the segmentation algorithm.The results show that compared with the traditional classification algorithm in the literature,the column recognition rate is 80%and the vehicle recognition rate is 65%.The classification algorithm based on polynomial kernel function has a recognition rate of 75%and 73%for columns and vehicles,respectively,which is increased by 5%and 8%.When using the other two kernel function,GS-SVM also maintains the advantages,comparing with the conventional algorithm,the proposed algorithm has strong robustness,which provides a solution to the problem of point cloud segmentation in weak light and geographical feature points environment.It also enriches the use scene of LiDAR 3D scanning.

关 键 词：点云分类 机器学习 GS-SVM SMRF滤波算法 激光雷达扫描 

分 类 号：P208[天文地球—地图制图学与地理信息工程]