光学学报, 2017, 37 (10): 1011001, 网络出版: 2018-09-07
激光点云的混合流形谱聚类自适应分割方法 下载: 705次
Mixed Manifold Spectral Clustering Adaptive Segmentation Method for Laser Point Cloud
成像系统 三维成像 点云分割 混合概率主成分分析法 谱聚类 imaging systems three-dimensional imaging point cloud segmentation mixtures of probabilistic principal component anal spectral clustering
摘要
将激光点云视为分布于三维欧氏空间的线性与非线性混合流形,提出一种基于混合流形谱聚类的自适应点云分割方法。由混合概率主成分分析法构造的M 个主成分分析器组成混合概率模型,得到描述点云的邻接矩阵;将点云分割的几何特征在谱空间进行降维嵌入,利用N-cut方法得到描述点云分割特征的多维向量;结合类间类内划分算法自适应分割点云。实验结果表明,对于三种受测点云,所提出的算法能在较宽预设参数范围内以80%以上概率得到收敛于几何特征的分割结果,参数稳定性较好。在对点云添加均值为0,标准差为0.01的高斯噪声与0.25倍数量的离群点复合噪声的情况下,算法表现出良好的抗噪性;将该算法应用于切片式激光三维成像的卫星模型点云中也取得了理想分割结果。
Abstract
The mixed manifold spectral clustering adaptive segmentation method is proposed, while the laser point cloud is regarded as a linear and nonlinear mixed manifold in three-dimensional Euclidean space. The mixture probabilistic model is constituted by M principal component analyzers, which are constructed by the mixtures of probabilistic principal component analysis method, and the adjacency matrix of point cloud is obtained. In the spectrum space, the geometrical characteristics of point cloud segmentation are embedded in the dimension, and the multi-dimensional vector, which describes the characteristics of point cloud classification, is obtained by N-cut method. The between-within proportion algorithm is adopted to segment point cloud adaptively. Experimental results show that the proposed algorithm can obtain segmentation results that converge to the geometric features with the probability larger than 80% in wide range of preset parameters. Moreover, it performs stable with Gaussian noises of 0 mean, 0.01 standard deviation and compound noise of 0.25 times the total points. The proposed method shows good noise resistance.It is applied to point cloud of satellite model acquired by slice laser three-dimensional imaging and achieves good segmentation results.
王帅, 孙华燕, 郭惠超, 都琳. 激光点云的混合流形谱聚类自适应分割方法[J]. 光学学报, 2017, 37(10): 1011001. Shuai Wang, Huayan Sun, Huichao Guo, Lin Du. Mixed Manifold Spectral Clustering Adaptive Segmentation Method for Laser Point Cloud[J]. Acta Optica Sinica, 2017, 37(10): 1011001.