基于PointNet++的机载LiDAR点云矿区地物分类模型研究#br#

现代矿业 ›› 2025, Vol. 41 ›› Issue (08): 221-225，233.

基于PointNet++的机载LiDAR点云矿区地物分类模型研究#br#

言龙

安徽省化工地质勘查总院

出版日期:2025-08-25 发布日期:2025-09-29

Research on Classification Model of Ground Objects in Airborne LiDAR Point Cloud Mining Area Based on PointNet++

Anhui Province Chemical Engineering Geology Exploration Institute

Online:2025-08-25 Published:2025-09-29

摘要/Abstract

摘要： 针对传统点云分类方法无法提取深度特征，且自动化程度低、过程复杂等问题，以深度学习分割算法PointNet++网络为基础，分别利用LASDU公开机载LiDAR点云试验数据集和自制矿山机载 LiDAR 点云数据进行试验分析，并以 OA、F1 Score 和 Avg F1 等为评价指标，开展与 PointNet 和 PointCNN 的分析比较与评估。研究表明：在公共数据集方面，PointNet++在 5个地物类别中有 3个取得最佳分类效果，另外2个类别的结果也接近最佳，且在整体分类性能上，PointNet++的OA和Avg F1 较 PointNet 和 PointCNN 分别提升 2.77，1.59 个百分点和 2.32，0.87 个百分点；在自制矿区地物数据集分类方面，PointNet++机载点云分类方法的 OA 评价指标为 71.47%，Avg F1评价指标为 61.45%，两者均高于PointNet和PointCNN。在地面点和建筑物的分类上，PointNet++在多项指标上取得最佳表现，且在自制矿区数据集的优势明显，适用于具有丰富结构的复杂场景，能够提高机载点云分类的准确性，提高分类自动化程度，简化分类流程，为矿区地物分类的实际应用提供有力支持，更是为未来地物分类模型优化和探索更先进的深度学习架构提供新的思路和基础。

关键词: 机载LiDAR点云, 复杂场景, 地物分类模型

Abstract: Aiming at the problems that traditional point cloud classification methods cannot extract deep features，have low automation and complex processes，based on the deep learning segmentation algo‐ rithm PointNet++ network，experimental analyses are conducted respectively on the publicly available air‐ borne LiDAR point cloud test dataset of LASDU and the self-made airborne LiDAR point cloud data of the mine. And taking OA，F1 Score and Avg F1，etc. as evaluation indicators，the analysis，comparison and evaluation with PointNet and PointCNN are carried out. The research shows that，in terms of public datasets， PointNet++ achieved the best classification results in three out of the five categories of ground objects，and the results in the other two categories are also close to the best. Moreover，in terms of overall classification performance，The OA and Avg F1 of PointNet++ have increased by 2.77，1.59 percent points and 2.32，0.87 percent points respectively compared with PointNet and PointCNN. In terms of the classification of self�made mining area feature datasets，the OA evaluation index of the PointNet++ airborne point cloud classifi‐ cation method is 71.47%，and the Avg F1 evaluation index is 61.45%，both of which are higher than those of PointNet and PointCNN，especially in the classification of ground points and buildings. Pointnet ++ has achieved the best performance in multiple indicators and has obvious advantages in self-made mining area datasets. It is suitable for complex scenarios with rich structures，can improve the accuracy of airborne point cloud classification，increase the degree of classification automation，simplify the classification process，and provide strong support for the practical application of mining area feature classification. It also provides new ideas and a foundation for the optimization of future ground object classification models and the explora‐ tion of more advanced deep learning architectures.

Key words: airborne LiDAR point cloud, complex scenes, classification model of ground objects

言龙. 基于PointNet++的机载LiDAR点云矿区地物分类模型研究#br#[J]. 现代矿业, 2025, 41(08): 221-225，233.

YAN Long. Research on Classification Model of Ground Objects in Airborne LiDAR Point Cloud Mining Area Based on PointNet++[J]. Modern Mining, 2025, 41(08): 221-225，233.

扫码分享