基于支持向量机的三维点云岩体结构面半自动识别方法  

Semi-Automatic Identification of Rock Mass Structural Planes in 3D Point Clouds Based on Support Vector Machines

作　　者：朱涛 史文兵[1,2] 刘永志 王勇[3] 梁风[1,2] ZHU Tao;SHI Wenbing;LIU Yongzhi;WANG Yong;LIANG Feng(College of Resource and Environmental Engineering,Guizhou University,Guiyang 550025,China;Mountain Geohazard Prevention R&D Center of Guizhou Province,Guiyang 550025,China;Key Laboratory of Karst Georesources and Environment(Guizhou University),Ministry of Education,Guiyang 550025,China)

机构地区：[1]贵州大学资源与环境工程学院,贵州贵阳550025 [2]贵州大学贵州省山地地质灾害防治工程技术研究中心,贵州贵阳550025 [3]喀斯特地质资源与环境教育部重点实验室(贵州大学),贵州贵阳550025

出　　处：《防灾减灾工程学报》2025年第1期95-103,共9页Journal of Disaster Prevention and Mitigation Engineering

基　　金：国家自然科学基金项目(42067046);贵阳市科技计划项目(筑科合同[2023]13-10号)资助。

摘　　要：结构面在评价岩体力学性质和边坡稳定性方面起着至关重要的作用,相比于传统测量,一种准确、高效的结构面参数识别方法尤为重要。提出了一种基于支持向量机(SVM)的三维点云岩体结构面提取的新方法,首先获取点云坐标、RGB、法向量、曲率和密度等作为机器学习模型的特征向量作为输入,结合人工和自动挑选学习样本,随后把学习样本分为训练集和测试集用于训练SVM模型并测试模型,将被接受的模型用于点云的预测分类,进而识别结构面和提取信息。将该方法应用于公开边坡数据集和发耳镇某采区边坡结构面调查,结果表明:使用LOF与PCA结合方法有效地提高了法向量估计的准确性,而DetRD-PCA方法用于估计单个结构面的法向量并计算产状时得到结果更加准确;对公开点云数据集的结构面进行识别,SVM识别881552个点时间仅需9 s,成功提取了四组结构面,与前人结果对比,倾向平均偏差最大3.12°,倾角平均偏差最大1.54°;将方法应用于发耳镇某采区边坡的结构面调查中,SVM识别1450148个点仅需18 s,成功提取了两组结构面,与经典的三点法估算比较,倾向和倾角的偏差为0.7°~3.3°和0.1°~3.3°;该方法对于小样本的训练数据依然能够表现出较高的正确率。Structural planes play a crucial role in evaluating the mechanical properties of rock masses and slope stability.Compared to traditional measurement methods,an accurate and efficient method for recognizing structural plane parameters is particularly important.This paper proposes a new method for extracting structural planes of rock masses from 3D point clouds based on Support Vector Machines(SVM).First,point cloud coordinates,RGB values,normal vectors,curvature,and density were used as feature vectors for the machine learning model.By combining manually and automatical-ly selected learning samples,the samples were divided into training and testing sets to train and test the SVM model.The accepted model was then used for point cloud prediction and classification,en-abling the identification of structural planes and information extraction.This method was applied to a publicly available slope dataset and a structural plane survey at a mining area in Fa'er Town.The re-sults showed that:the combination of LOF and PCA methods effectively improved the accuracy of normal vector estimation,while the DetRD-PCA method provided more accurate results for estimat-ing normal vectors and calculating orientation for individual planes.When applied to the publicly avail-able point cloud dataset,the SVM identified 881552 points in just 9 seconds,successfully extracting four sets of structural planes.Compared with previous results,the maximum average deviation in the dip direction was 3.12°,and the maximum average deviation in the dip angle was 1.54°.When applied to the structural plane survey in the Fa'er Town mining area slope,the SVM identified 1450148 points in 18 seconds,successfully extracting two sets of structural planes.Compared with the classic three-point method,the deviations in dip direction and dip angle ranged from 0.7°to 3.3°and 0.1°to 3.3°,respectively.The method demonstrated high accuracy even with small training sample sizes.

关 键 词：三维点云 支持向量机 岩体结构面 半自动识别 产状计算 工程应用 

分 类 号：P642[天文地球—工程地质学]