Research On Point Cloud Denoising Algorithm Based On Guidance Information

In recent years,with the popularity of 3D laser scanners and low-cost sensors,3D point cloud data can be obtained more quickly and easily.3D point cloud data has simple,flexible and powerful representation capability.Nowadays,3D point cloud models are more and more widely applied,such as robot,augmented reality,automatic driving,intelligent manufacturing,computer art and so on.However,due to various reasons such as scanning precision is not enough with the device itself,the point cloud data scanned by scanning inevitably contains a lot of noise,so point cloud denoising becomes the foundation and challenge of downstream applications.Therefore,this paper takes scattered point cloud as the research object and proposes a point cloud denoising algorithm from local and non-local information perspectives.The main work and innovations of the paper are as follows:I Curvature weighted guide point cloud denoising algorithm based on local informationThe existing denoising algorithms use the same strategy for all the points in the model when constructing local neighborhood,so that the neighborhood at the feature point is generally composed of multiple smooth regions,which is one of the reasons why the feature is smoothed after denoising.In order to make the point cloud model retain the features of the model while denoising,this paper proposed a curvature weighted guide point cloud denoising algorithm based on local information.The innovations of this algorithm mainly include:(1)A neighborhood reconstruction strategy for feature points is designed.This strategy makes the neighborhood at the feature point have a unique smooth region.The experimental results show that this method can extract the feature points from the noise model in a robust way and is helpful to preserve the features of the model.(2)A weighted guided point cloud denoising algorithm based on curvature information is proposed.On the basis of guided filtering algorithm,different weights are assigned to the points in the smoothing region and the feature region of the model to further preserve the feature details of the model.Experimental results show that this method can not only effectively preserve the features of the model but also be robust to noise intensity.II Guided point cloud denoising based on non-local informationMost existing denoising algorithms make a series of assumptions about noise distribution,which is not enough to deal with the real noise models.Therefore,this paper proposes a guided point cloud denoising algorithm based on non-local information.The innovations of this algorithm mainly include:(1)A method for extracting non-local similar patches from 3D point cloud is designed.It uses the non-local self-similarity of point cloud model to provide training data for model prior learning.(2)An algorithm of GMM clustering learning for patches of ideal models is proposed.The prior knowledge learning of ideal models is realized to guide the local patches clustering of noise model.(3)An approach to the orthogonality dictionary learning of the ideal prior-guided noise model is proposed.The hybrid orthogonal dictionary not only retains the fine structure information of the general model,but also contains the characteristic details of the noise model.The experimental results show that the proposed algorithm is effective for real noise models.In this paper,a point cloud denoising algorithm based on guided information is proposed from the perspective of local information and non-local information.The experimental results show that the algorithms in this paper can preserve the features of the model well while removing the noise,and have certain robustness for noise scale,which have important application value and theoretical value.