保持特征的散乱点云数据去噪

为保证在去除点云数据噪声的同时不损失模型的细节特征,提出了一种基于特征信息的加权模糊C均值聚类去噪算法。首先,构建点云K-D树拓扑结构,根据点的r半径球邻域点统计特性去除大尺度离群噪声点。然后,利用主元分析法估算点云的曲率和法向量,根据曲率特征标识点云数据的特征区域,并采用特征加权模糊C均值聚类算法对特征区域去噪,采用加权模糊C均值聚类算法对非特征区域去噪。最后,使用双边滤波器对点云模型进行平滑。对提出的算法进行了验证实验,结果显示:去噪后点云模型的最大偏差保持在模型尺寸的0.15%以内;标准偏差保持在模型尺寸的0.03%以内。本文算法能够在有效去除不同尺度和强度的噪声的同时不损失点云模型的细节特征,去噪精度高,且对不同的噪声模型具有较强的鲁棒性。

Feature-preserving scattered point cloud denoising

To move the outliers and noisy points away from 3D point cloud data and to maintain the sharp features of the model simultaneously, a feature-based weighted fuzzy C-means point cloud denoising algorithm was proposed. Firstly, the point cloud was organized by K-D tree data structure and the large-scale outliers were removed by the statistics of r radius neighboring points. Then, the principal component analysis method was adopted to estimate the curvature and normal vector of point cloud data and the patches with distinguished features were identified according to the curvature feature weight. Pursuant to different feature regions, the feature-preserving weighted fuzzy C-means clustering algorithm was adopted to denoise for the patch with rich feature information and the fuzzy C-means clustering algorithm was adapted to denoise for the patch with less feature information, respectively. Finally, a bilateral filter was used to smooth the data set. The algorithm was verified and the experimental results show that the max denoising error is limited to 0.15% of the model size and the min denoising error is limited to 0.03% of the model size. In conclusion, this approach moves efficiently and precisely the noise with different scales and intensities in point cloud, meanwhile performing a feature-preserving nature. Moreover, it is robust enough to different noise models.