Point Cloud Denoising with Principal Component Analysis and a Novel Bilateral Filter

This paper aims to remove the noises of different scales in point cloud data captured by 3D scanners, while preserving the sharp features (e.g. edges) of the model. For this purpose, the authors proposed a point cloud denoising method based on the principal component analysis (PCA) and a self-designed bilateral filter. First, the outliers in the point cloud were divided into isolated outliers and deviation outliers. The former was directly removed, while the latter was moved along the normal vector estimated by the PCA. Next, a bilateral filter was developed based on vertex brightness, vertex position and normal vector. During image processing, the grayscale of the current point was replaced with the weighted mean of the grayscales of its neighborhood points. The weight function is related to the distance and grayscale difference between the current point and neighborhood points. The effectiveness of our method was proved through experiments on actual point clouds. The results demonstrate that our bilateral filter can retain the sharp features of point cloud data, in addition to removing the small-scale noises.