点云降维投影的节理迹线自动提取（NTDR）— 论文配套代码归档 Point-cloud dimensionality reduction for automatic joint trace extraction (NTDR)
本仓库是以下已发表论文的配套代码归档：
点云降维投影方法及其在节理迹线快速提取中的应用 《岩土力学》（北大中文核心 / CSCD / EI），2026, 47(2): 691–702 DOI:
10.16285/j.rsm.2025.0208
- 方法与代码由大连海事大学交通运输工程学院岩石力学课题组开发，完整作者名单与贡献见论文原文。
- 本仓库由论文第四作者整理开源；本人在项目中负责点云降维投影模块（协方差法向估计、旋转对齐、正交投影、RGB 同步映射）的实现与验证，并参与图像栅格化、边缘检测与聚类连线的实验调试。
- 工程数据：加拿大安大略 15 号省道岩体露头，Leica HDS6000 地面三维激光扫描仪采集，点云规模 216 万点且含 RGB 色彩信息。
三维彩色点云
│ ① 协方差矩阵 SVD 求最佳投影面，正交投影至二维（RGB 同步映射）
▼
二维彩色图像
│ ② RGB 均值栅格化
▼
栅格化图像
│ ③ 高斯滤波（13×13, σ=2）+ Canny 边缘检测
▼
边缘迹线点
│ ④ DBSCAN 曲率加权聚类 + 连线
▼
节理迹线矢量结果
关键指标（详见论文）：
| 指标 | 数值 |
|---|---|
| 较人工提取节省处理时间 | 91.07% |
| 与人工提取结果重合率 | 90.42% |
| ≤20% 噪点干扰下识别正确率 | ≥80% |
| 理论模型重合率 | 99.86%–100% |
| 目录 | 内容 |
|---|---|
projection/ |
降维投影模块：Demo_modify.m（主入口）、法向估计、旋转对齐、透视投影、伽马矫正 |
connection/ |
聚类连线模块：y3.m / dian.m（入口）、pointsextraction/（八叉树下采样、DBSCAN 聚类、Otsu、连线、LBP） |
docs/ |
NTDR 说明书、程序流程、效果图 |
connection/pointsextraction/tetrahedron_point_cloud.ply |
小型合成测试点云（可直接跑通 demo） |
- 环境：MATLAB（
pcread/pcshow需 Point Cloud Toolbox；聚类部分依赖自实现 DBSCAN） - 入口：
projection/Demo_modify.m（投影 + 栅格化演示）、connection/pointsextraction/ExtractTracesMain.m（完整提取流程） - 数据说明：完整工程点云
Mesh.ply（216 万点，约 200 MB）及中间产物（.mat/.tif）因体积未包含在本仓库；将自备的彩色点云命名为Mesh.ply置于工作目录即可复现流程。小型合成数据tetrahedron_point_cloud.ply可先跑通链路。
connection/pointsextraction/DBSCAN.m— Yarpiz（许可见同目录LICENSE-yarpiz-dbscan.txt）connection/pointsextraction/ipdm.m— MathWorks File Exchange（Inter-Point Distance Matrix）
Companion code archive for the published paper "Point-cloud dimensionality reduction and its application to rapid joint trace extraction", Rock and Soil Mechanics (Core/CSCD/EI), 2026, 47(2): 691–702, DOI: 10.16285/j.rsm.2025.0208.
Pipeline (4 stages): SVD best-fit projection plane with synchronized RGB mapping → RGB-mean rasterisation → Gaussian filtering + Canny edge detection → curvature-weighted DBSCAN clustering and line connection.
Key results: 91.07% time saving vs. manual extraction; 90.42% overlap with ground truth; ≥80% accuracy under ≤20% noise; 99.86–100% overlap on theoretical models.
Attribution: method and code developed by the rock-mechanics research group at Dalian Maritime University (full author list in the paper). This archive is curated by the fourth author, who implemented and validated the dimensionality-reduction/projection module (normal estimation, rotation alignment, orthographic projection, RGB co-mapping). Engineering dataset: Highway 15 (Ontario, Canada) rock outcrop, 2.16 M RGB points acquired with a Leica HDS6000 terrestrial laser scanner (not included due to size; bring your own colored point cloud as Mesh.ply).
Third-party code: DBSCAN.m from Yarpiz (license alongside); ipdm.m from MathWorks File Exchange.
Environment: MATLAB with the Point Cloud Toolbox. Entry points: projection/Demo_modify.m, connection/pointsextraction/ExtractTracesMain.m.