{
  "paper_id": "5a59ad9ef4bfacde2e1fc87969588796",
  "paper_title": "An Improved DBSCAN Method for LiDAR Data Segmentation with Automatic Eps Estimation",
  "paper_authors": [
    "Chunxiao Wang",
    "Min Ji",
    "Jian Wang",
    "Wei Wen",
    "Ting Li",
    "Yong Sun"
  ],
  "publication_date": "2019-01-05",
  "publication_venue": "Remote Sensing",
  "abstract": "Point cloud data segmentation, filtering, classification, and feature extraction are the main focus of point cloud data processing. DBSCAN (density-based spatial clustering of applications with noise) is capable of detecting arbitrary shapes of clusters in spaces of any dimension, and this method is very suitable for LiDAR (Light Detection and Ranging) data segmentation. The DBSCAN method needs at least two parameters: The minimum number of points minPts, and the searching radius ε. However, the parameter ε is often harder to determine, which hinders the application of the DBSCAN method in point cloud segmentation. Therefore, a segmentation algorithm based on DBSCAN is proposed with a novel automatic parameter ε estimation method—Estimation Method based on the average of k nearest neighbors' maximum distance—with which parameter ε can be calculated on the intrinsic properties of the point cloud data. The method is based on the fitting curve of k and the mean maximum distance. The method was evaluated on different types of point cloud data: Airborne, and mobile point cloud data with and without color information. The results show that the accuracy values using ε estimated by the proposed method are 75%, 74%, and 71%, which are higher than those using parameters that are smaller or greater than the estimated one. The results demonstrate that the proposed algorithm can segment different types of LiDAR point clouds with higher accuracy in a robust manner. The algorithm can be applied to airborne and mobile LiDAR point cloud data processing systems, which can reduce manual work and improve the automation of data processing.",
  "keywords": [
    "LiDAR",
    "segmentation",
    "DBSCAN",
    "parameter estimation"
  ],
  "paper_institutions": [
    "Geomatics College, Shandong University of Science and Technology",
    "Hainan Geomatics Centre, National Administration of Surveying, Mapping and Geoinformation of China"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "ark-code-latest",
  "prompt_template": "simple",
  "extracted_at": "2026-08-16T19:47:33.070600Z",
  "template_tag": "simple",
  "summary": "### 摘要\n本论文针对DBSCAN算法在LiDAR点云分割中因参数ε难以确定而制约自动化的难题，提出了一种基于k近邻最大距离均值的自动参数估计方法。该方法通过数据归一化、Kd-tree空间索引构建，对每个点计算其k近邻最大距离并求均值，建立k与均值最大距离的函数关系，利用多项式拟合及修正，将一阶导数为1时的对应f(k)值作为最优聚类半径ε。实验在机载LiDAR数据（含XYZ和XYZRGB两种特征）与移动LiDAR数据（仅XYZ）上进行，分别设置不同ε值进行分割，并与参考数据对比。结果表明，采用估计ε值时分割准确率分别为75%、74%和71%，均高于相邻较小或较大参数的准确率，且分割结果更合理。该算法能处理高维数据，支持自动分割，可减少人工干预，适用于机载及移动LiDAR点云处理系统。本文还分析了段内与段间的密度变化，解释了参数估计的理论依据，并验证了方法的鲁棒性。\n\n### 关键要点\n- 提出基于k近邻最大距离均值的自动ε估计方法，利用点云内在属性计算聚类半径，避免依赖人工经验。\n- 方法流程包括数据归一化、Kd-tree索引、参数估计、改进DBSCAN聚类及结果输出，其中归一化仅在多维特征（如XYZRGB）时必要。\n- 利用多项式拟合k与均值最大距离的关系，并添加修正项使曲线归一化，通过寻找一阶导数为1的拐点确定最优ε。\n- 改进的DBSCAN算法增加了maxPts参数以控制簇大小，并采用队列实现核心点扩展，处理高维数据时结合Kd-tree提高效率。\n- 机载数据（XYZ）实验中，估计ε为0.8114时准确率达75%，明显高于其他测试（45%~70%），且过分割和漏分割较少。\n- 机载数据（XYZRGB）实验中，估计ε为0.097时准确率为74%，证明方法适用高维特征，但噪声比例略高（14.4%）。\n- 移动LiDAR数据（XYZ）实验中，估计ε为1.14686时准确率为71%，远优于其他参数（5%~66%），证实方法在不同数据源上的鲁棒性。\n- 实验表明，ε过小导致过分割（物体碎裂），ε过大导致欠分割（不同物体粘连），而估计值能平衡两者。\n- 方法依赖minPts和maxPts仍需人工设定，未来研究可探索这两个参数的自动估计。\n- 实现中采用平方距离代替欧氏距离以提升计算效率，并利用Kd-tree进行范围搜索与K近邻搜索。\n- 算法在分割后可反映归一化数据至原始数据，便于输出结果，但需保持点顺序一致。\n- 所提方法可应用于机载和移动LiDAR数据自动分割，减少人工工作，提高处理自动化程度。"
}