Research On The Completion Method Of 3D Point Cloud Dat

In the field of 3D computer vision,unstructured point cloud data has been widely used due to its convenient collection method and simple data structure.Although 3D scanning equipment can collect high-precision point cloud data,the acquired data is usually sparse and incomplete due to factors such as the inherent resolution limitation of the scanning equipment,the reflective properties of the surface of the object,or the occlusion of the line of sight.Therefore,it is necessary to complete the missing point cloud data for downstream tasks.Most of the traditional point cloud completion methods are used to complete the small-scale missing on the surface of the object,and the completion effect on the data with lost shape structure is poor.In order to complete the incomplete point cloud data,this paper uses the method of deep learning to construct a deep learning network framework that directly takes 3D point cloud data as input for two complementary objects,the shape of a single point cloud object and the indoor point cloud scene.Output fine single point cloud objects and complete indoor point cloud scenes.A multi-resolution point cloud completion network structure,fusing graph attention,is constructed in this paper to address the challenge of extracting local feature information from 3D point clouds.In general,the method of generating confrontational network framework to process data is adopted.The generator in the network constructs the point cloud structure through the graph attention layer,fuses the feature information of different resolutions and adds two-dimensional grid data,and reconstructs the missing points by combining the folding operation.Structure and output the point cloud data that is completed step by step;the discriminator distinguishes the authenticity of the point cloud,improves the accuracy and optimizes the generator through feedback,so that the generated data has a fine geometric structure and is similar to the real point cloud.The constructed point cloud completion network was verified experimentally and theoretically analyzed with four related methods on a large shape dataset,and achieved optimal results.This paper’s method has been demonstrated to be capable of effectively supplementing the absent portion of the point cloud form,resulting in a full and consistent point cloud shape,which boosts performance by approximately 1.79%in comparison to the point fractal network.Crafting a network structure to finish a point cloud can not only accurately reflect the overall characteristics of the point cloud,but also probe its local geometric features deeply,refining the shape of the ultimate point cloud.This paper constructs a point cloud scene completion network framework based on semantic instance joint segmentation,aiming to address the challenge of using neural networks to directly process point cloud scene data in indoor point cloud scene data completion.Inputting the absent scene point cloud information into the scene completion network,the object components of the scene are segmented through the utilization of a semantic instance joint segmentation strategy.The self-encoding structure’s completion network complements the indoor furniture segmented by the scene’s shape.Finally,the restored furniture shape is integrated into the scene data to obtain a complete 3D point cloud scene data.This paper advances the autoencoder structure-based point cloud completion network in order to maximize its repair effect.By executing input alteration and feature transformation on the input point cloud data,a perceptron with shared weight extracts the feature information of the point cloud.Maximum pooling operation is used to generate the feature code word,and the decoder is then fed two-dimensional grid data.The perceptron’s folding operation is utilized by the decoder to transform grid data into a full point cloud shape.The experimental results verify that the network can effectively complete the missing shapes,and the constructed scene completion network can effectively complete the missing data in the indoor space.This paper builds a point cloud shape completion network and an indoor point cloud scene restoration completion network for the accomplishment of 3D point cloud object shapes and point cloud scenes.A deep neural network framework for point cloud completion is suggested in this paper,based on the elucidation of 3D point cloud data completion techniques both domestically and abroad.A quantitative comparison and elucidation of the associated techniques,along with an experimental visualization effect of the algorithm in question,is presented.