基于地面拟合与DBSCAN聚类的激光点云去噪算法研究

打开文本图片集

主题词：三维边界框估计 点云去噪 地面点云分割DBSCAN聚类中图分类号：U495 文献标志码：A DOI：10.19620/j.cnki.1000-3703.20250184

Laser Point Cloud Denoising Algorithm Based on Ground Fitting and DBSCAN Clustering

Wu Jigang'， Zhou Xingyu²，Li Peng² （1.Beijing Gene Inteligent Car TechnologyCo.，Ltd.，Beijing10o08O;2.China Highway Vehicle &MachineryC.Ltd， Beijing 100013）

【AbstractIToaddressthechalenges faced byconventionallidarsystems inaccurately filtering out water sprayanddust， highlyreflectiveexpansion crostalk noise，andthesuboptimalremovaloflarge-scaleoutliers，alaserpointclouddenoising algorithmbasedonground fitingand Density-Based Spatial Clusteringof Applications with Noise （DBSCAN）is proposed. Initialnoiselocationinformationisobtainedthroughgroundpointcloudsegmentation，anenhancedDBSCANalgorithm，anda three-dimensionalboundingboxestimationmethod.Noiseisthenmarkedandfilteredoutbasedonitsgeometricand physical characteristics.Theexperimental resultsdemonstrate that in dustand water sprayscenarios，the noise-reduced Chamfer Distance（CD）decreases fromO.0374m to0.0198m，whilethePoint-to-Mesh （P2M）distancereduces from0.158mto 0.088m.Theoverallgeometricdeviationof thepointcloudissignificantlydiminishd，withimprovedonforitytothetrue surface.For procesing complexpoint clouddata，theoptimised DBSCAN clustering achieves acomputational timeof merely 6 ms compared to the conventional DBSCAN algorithm，representing a77.8% improvement in computational eficiency.This efectivelyresolvesthesuboptimaloutlierfilteringperformanceoftraditionallaserpointclouddenoisingalgorithmsatlarge scales.

Key words：3D bounding box estimation，Point cloud denoising，Ground point cloui segmentation，DBSCAN clustering

【引用格式】吴纪刚，周星宇，李鹏.基于地面拟合与DBSCAN聚类的激光点云去噪算法研究[J].汽车技术，2026（2）：48-55.WUJG，ZHOUXY，LIP.Laser Point Cloud DenoisingAlgorithmBasedon GroundFiting andDBSCANClustering[J].Automobile Technology，2026（2）：48-55.

1前言

激光扫描技术广泛应用于无人驾驶、模型三维重建等领域-4，但因复杂工况下环境扬尘、扬水较多，易导致点云数据产生大量噪点，严重影响后续的特征提取和特征匹配[5-，进而影响三维目标检测结果或点云地图构建效果。（剩余10080字）