Abstract: To solve the problem of excessive smoothing and local distortion caused by traditional point cloud denoising algorithm,a feature preserving and anisotropic filtering algorithm was proposed based on anisotropic Gaussian kernel.Informational similarity function was constructed according to the projection of neighborhood point in the tangent plane and the normal of sampling point,and effective neighborhood of sampling points was defined according to the information similarity function.Distribution properties of eigenvalues and eigenvectors of surface points,curve points and corner points were researched by using principal component analysis theory.On this basis,parametric adaptive and anisotropic Gauss kernel function was constructed by taking pseudo inverse matrix of covariance matrix as the bandwidth matrix,and combined it with the bilateral filtering algorithm for scattered point cloud denoising.The experimental results showed that the algorithm could adjust filter main direction and attenuation velocity of each principal direction adaptively according to the local distribution properties,the algorithm could denoise the point cloud well and preserve original sharp feature at the same time.

Key words: informational similarity, effective neighbor, adaptive, anisotropic Gaussian kernel, point cloud denoising

摘要： 为解决传统点云去噪算法造成的过光顺及局部失真问题,提出一种基于各向异性高斯核的散乱点云保特征去噪算法。根据邻域点在其切平面上的投影和采样点法向构建信息相似度函数,并通过信息相似度函数定义有效邻域;应用主元分析理论研究曲面采样点、棱线采样点和角点的特征值及特征向量的分布特性;在此基础上构建以协方差矩阵的伪逆矩阵为带宽矩阵的参数自适应的各向异性高斯核函数,并将其与双边滤波算法结合用于散乱点云去噪。实验结果表明,该算法能够根据点云的局部分布特性自适应地调整滤波主方向和各主方向的衰减速度,在实现散乱点云去噪的同时可有效保持点云模型的原始尖锐特征。

关键词: 信息相似度, 有效邻域, 自适应, 各向异性高斯核, 点云去噪