|
| 融合距离自适应与密度感知的点云聚类算法 |
| A Point Cloud Clustering Algorithm Integrating Distance Adaptation and Density Awareness |
| 投稿时间：2026-03-16 修订日期：2026-04-05 |
| DOI： |
| 中文关键词: 模式识别 激光雷达 点云聚类 DBSCAN算法 自适应机制 |
| 英文关键词: pattern recognition LiDAR point cloud clustering DBSCAN algorithm adaptive mechanism |
| 基金项目:安徽省自然科学基金面上项目（2508085MA020）；安徽省住建厅科学技术计划项目(2022-YF016,2022-YF065,2023-YF050)；安徽省教育厅研究生科研项目(YJS20210512)；安徽省新时代育人质量工程项目(2024xscx115)；机电类研究生军民融合联合培养基地项目(20231hpysfjd050)。 |
|
| 摘要点击次数: 60 |
| 全文下载次数: 0 |
| 中文摘要: |
| 激光雷达点云密度随距离呈反比平方级衰减，导致其在空间分布上的不均匀性。传统的基于密度的空间聚类（Density-Based Spatial Clustering of Applications with Noise, DBSCAN）由于参数固定，在近场密集区域易引发目标点云的欠分割现象，而在远场稀疏区域则易导致目标点云的过分割问题。为此，提出一种融合距离自适应与密度感知的DADS-DBSCAN（Distance-Adaptive Density-Sensitive DBSCAN）算法。通过构建密度敏感阈值自适应模块和距离梯度自适应邻域模块，利用距离补偿因子与局部密度感知重构最小点数阈值与搜索半径，实现对非均匀密度点云的自适应聚类；其次，引入基于主成分分析（Principal Component Analysis, PCA）的几何约束驱动的簇合并模块，通过方向一致性校验合并过分割簇以修复目标几何完整性；最后，结合簇均值密度与空间孤立度构建双重判定层次化精炼模块，精准剔除背景噪声。实验表明：该算法的调整兰德系数（ARI）与归一化互信息（NMI）分别达到0.9536和0.9623，较传统算法提升约7.06%和7.30%，戴维森堡丁指数（DBI）优化至0.4835。结果证实，该方法有效克服了近场欠分割与远场过分割难题，提升了非均匀点云聚类的精确度与场景适应能力。 |
| 英文摘要: |
| The density of LiDAR point clouds exhibits an inverse-square decay with respect to distance, resulting in a highly non-uniform spatial distribution. Due to its fixed parameters, the traditional Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm is prone to under-segmentation of target point clouds in dense near-field regions and over-segmentation in sparse far-field regions. To address these issues, this paper proposes a Distance-Adaptive Density-Sensitive DBSCAN (DADS-DBSCAN) algorithm. By constructing a density-sensitive threshold adaptation module and a distance-gradient adaptive neighborhood module, the proposed algorithm utilizes a distance compensation factor and local density awareness to dynamically reconstruct the minimum point threshold and search radius, thereby enabling adaptive clustering for non-uniform point clouds. Furthermore, a cluster merging module driven by geometric constraints based on Principal Component Analysis (PCA) is introduced to merge over-segmented clusters via directional consistency checks, restoring the geometric integrity of the targets. Additionally, a dual-criteria hierarchical refinement module is constructed by integrating cluster mean density and spatial isolation to precisely eliminate background noise. Experimental results demonstrate that the Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) of the proposed algorithm reach 0.9536 and 0.9623, respectively, yielding improvements of approximately 7.06% and 7.30% over the traditional algorithm, while the Davies-Bouldin Index (DBI) is optimized to 0.4835. These findings confirm that the proposed method effectively overcomes the challenges of near-field under-segmentation and far-field over-segmentation, substantially enhancing the clustering accuracy and scene adaptability for non-uniform point clouds. |
|
View Fulltext
查看/发表评论 下载PDF阅读器 |
| 关闭 |
