Research And Implementation Of Point Cloud Completion Algorithm Based On Depth Grid Transformation

With the introduction and popularization of a series of advanced technologies and facilities,the acquisition of 3D point cloud data through various sensors and samplers has become increasingly simple.This paper aims at various problems and challenges in the process of point cloud acquisition: 1.The shape of point cloud data is missing,making it difficult to obtain its corresponding shape features.2.Reconstruction difficulties caused by insufficient resolution of bottleneck information in the reconstruction process of point cloud completion.Carry out the research on point cloud completion algorithm based on the deep grid transformation computing framework,aiming to make certain contributions to improve the quality of data collection,algorithm accuracy and other issues:(1)A point cloud feature extraction algorithm based on learnable K near point sampling is proposed.By combining the improved edge convolution of the convolution kernel with the deep residual network connection,the coding block is optimized.Meanwhile,the attention mechanism based on feature weight is introduced to strengthen the feature extraction process,which can overcome the disorder of the point set and extract the features of adjacent points on a two-dimensional scale,thus improving the interpretability and accuracy of the coding results.In the analysis of experimental data,the quantitative experimental results of this model in the benchmark index EMD,CD and F-Score were 2.27,1.65 and 0.66,respectively.(2)A point cloud completion method based on stacked grid transformation combined with style coding is proposed.In this method,the bottleneck information is transformed by affine using the generative property of generative adversarial network,and the convolution kernel is assigned in the reconstruction process of each layer.The approach that introduces style coding injection makes full use of the bottleneck information,which is expressed as a subshape of the overall shape,and all the subshapes are synthesized into the final completed shape by stacking.At the same time,a new completion architecture is proposed,and the farthest point sampling algorithm based on recursive equinoxes is proposed to solve the problem of uneven data distribution in the traditional completion architecture,so as to further improve the integrity and accuracy of high point cloud completion.In the experimental analysis,the benchmark indexes EMD,CD and F-Score of this model improved by 16% compared with other algorithms.(3)Design and build a low-delay data completion system for point cloud completion,which is composed of four subsystems: information management,point cloud completion,personal warehouse and background management.Users can submit the defective point cloud in real time and return the complete point cloud data.Users can also view the uploaded and shared point cloud data through personal warehouse,thus improving the efficiency in the collection process of high point cloud.Meet the application requirements of point cloud acquisition in intelligent devices,user interaction and other fields.In this paper,algorithms and applications related to the design and implementation of multiple point cloud completion are transformed into: 1 journal paper and 2 invention patents in terms of academic achievements.