Research On 3D Point Cloud Attribute Compression Method Based On Geometric Transformation Prediction

With the continuous development of laser scanning technology,as one of the main representations of 3D scenes,point clouds have been widely used in many fields.The huge data volume of point cloud brings a good visual experience on the one hand,but also brings certain storage challenges on the other hand.For example,a 3D human body model has millions of points,which makes data transfer very difficult with limited bandwidth.Therefore,the role of point cloud compression becomes particularly important.At this stage,the current mainstream coding schemes cannot make full use of the correlation between point cloud attributes,which leads to a significant decrease in coding performance.Therefore,these coding methods have certain limitations.Graph transformations,on the other hand,preserve more fundamental information about the true 3D structure and the correlation between points,which works well for disordered and irregular point cloud data.To solve the existing problems in point cloud attribute coding,this paper proposes a point cloud attribute compression scheme based on predictive transformation.This scheme uses graph transformation to process prediction residuals,which significantly improves the compression efficiency of point cloud attributes.This thesis utilizes the spatial correlation between adjacent points and adjacent blocks to eliminate spatial redundancy.One of the main contributions: a bidirectional intra prediction scheme is proposed.There are 9 prediction modes in this scheme,including 4 unidirectional prediction modes,4 bidirectional prediction modes and DC prediction modes.In the bidirectional prediction mode,each mode is a combination of three unidirectional prediction modes according to a certain distance proportional weight.This prediction method extends the block-based intra prediction module.Experimental results show that the point cloud attribute compression algorithm for bidirectional intra-frame prediction proposed in this thesis reduces the BD-rate by 14.62%on average and improves the compression performance.Another contribution of this thesis is that for each coding block,an optimized graph transformation method is adopted.Specifically,the coding performance of point cloud attribute compression is improved by assigning weights to the Laplacian matrix.In graph transformations,the parameter values of the Gaussian kernel function used to assign weights are crucial.Here,in this paper,the key information such as geometric distance and attribute information of point cloud is comprehensively considered for graph transformation.This improved radial basis function model refers to the idea of bilateral filtering in image processing.Use a normalization factor to balance information from two different dimensions.Therefore,the geometric information and attribute information of the point cloud can be fully combined to participate in the encoding to achieve better transformation efficiency.This scheme can achieve better results in point cloud attribute compression.The experimental results show that the average code rate of this scheme on the upper body dataset is reduced by about 10.36%.In terms of visual quality,it can effectively improve the precision of point cloud model compression and reconstruction.