Robust denoising algorithm for point-sampled models based on forward-search and mean-shift
基于前向查找和均值漂移的点模型鲁棒降噪算法

Based on two robust statistics methods, forward-search and mean-shlft, an algorithm for robust filtering of noisy point-sampled models was presented. Forward-search algorithm detected outliers automatically by using residual plot and classified point clouds to multiple optimal outlier-free neighborhoods locally. By analyzing the weighted covariance matrix of a local neighborhood, its least-squares plane was estimated. Kernel functions of sample points in local regions were estimated and the local maxima of the kernels was computed by using mean-shift technique. The local maxima of the kernel estimation function determined cluster centers of point cloud data, which delivered an accurate approximation of the sampled surface. Each sample point was shifted to the local maximum of the kernel function, so the point-set surface could converge to a stable 3D digital model. Experiments show that our method is robust. It can smooth the noise efficiently and preserve the sharp features of the surface effectively.