Research On 3D Point Cloud Compression And Quality Enhancement Technology

3D laser scanning technology is widely used in fields like autonomous driving,digital museum,robot and medicine in recent years.3D point cloud can be obtained by 3D scanning equipment and cameras,which can describe the surface characteristics of the object,and it generally contains geometric and color information.3D point cloud is a collection of massive points,and there is no topological relationship and order between points,so whether the 3D point cloud can be quickly transmitted to the users through the limited network bandwidth is the prerequisite for 3D point cloud's application,and how to effectively compress the 3D point cloud is the key to solve this problem.But the data after lossy compression will have varying degrees of distortion,inducing poor visual effects.Thus,it is of great significance to search the quality enhancement technology of 3D point cloud.Aiming at the large data volume of 3D point cloud,this paper firstly studies 3D point cloud compression technologies based on compressed sensing and down sample encode-up sample reconstruction methods,and then taking G-PCC(Geometry based Point Cloud Compression)as the 3D point cloud compression coding framework,studies the quality enhancement of 3D point cloud data based on deep learning method.This paper makes a deep study on related works,the main works and achievements are as follows:1.Introduce the acquisition methods of 3D point cloud,and analyzed the problems and corresponding solutions of 3D point cloud in the process of obtaining and applying.2.Analyze and study the common 3D point cloud compression algorithms,and propose a 3D point cloud data compression algorithm based on compressed sensing and sampling method.First,the compressed sensing algorithm is studied.By analyzing the distribution characteristics of 3D point cloud,compressed sensing is used for "compress-sampling" and reconstruction.In this procedure,to better dig out the sparse representation of signals,K-SVD algorithm is used.The algorithm learns a general over-complete dictionary based on five point cloud sequences.By analyzing the experimental results,we can see that the algorithm can not only reduce the number of coding bits but also restore the original data effectively.Then we study the sample-based compression algorithm.After analyzing the shortcomings of the existing sampling algorithm,we propose our own complete compression algorithm based on down sample encode-up sample reconstruction methods.By analyzing the experimental results,we can see that the algorithm can effectively reconstruct the original point cloud shape on the premise of retaining the amount of original point cloud data.This algorithm has important reference value for immersive communication.3.A quality enhancement algorithm is proposed for the reconstructed lossy 3D point cloud.First,the 3D point cloud is divided into patches,and the luma component of each point in each patch is extracted.Then the data in each patch is arranged into 2D form according to the custom mode,and sent to the proposed neural network for training.In the test phase we show that our algorithm can improve the quality of reconstructed 3D point cloud.