An improved point cloud denoising method based on gradient field

3D point clouds obtained through scanning devices have been widely used in many applications such as autonomous driving, three-dimensional modeling and cultural heritage conservation. Point cloud are often corrupted with the amount of noise due to the limitation of the devices. It affects the performance of point cloud based 3D vision tasks such as surface reconstruction, detection, and segmentation. This paper focus on the importance of point cloud smoothness for downstream tasks. A novel regularization method is proposed to solve the smoothness problem in the momentum gradient ascent of point cloud execution, which is essentially iterative refinement of intermediate point cloud. The experimental results show that compared with the previous advanced methods, both the error values of CD and P2M obtained by the proposed method are reduced. Furthermore, the proposed method not only obtains better denoising results, but also significantly improves the smoothness of the point cloud after denoising, and provides more accurate point cloud data for downstream tasks.