Research On Point Cloud Completion In 3D Reconstruction Based On Deep Learning

In 3D digital modeling,3D point cloud is one of the most widely used data structures.Real-world point clouds are usually collected from 3D sensors such as Li DAR,depth cameras,3D scanners,etc.However,due to factors such as occlusion,light reflection,target surface material and sensor resolution,the point clouds collected in the real world are prone to local sparsity or local missing,which causes loss of geometric structure and semantic information,bringing impact to the accuracy of 3D digital modeling.Since missing point clouds are prone to semantic-level information loss,traditional 3D vision methods based on geometric solving have limitations in incomplete point cloud completion tasks.In recent years,deep learning has been rapidly developed in the field of 3D vision.In this paper,we explore the use of a data-driven deep generative model to fuse semantic and geometric structure information to repair incomplete missing point cloud data with high accuracy.We innovatively propose a deep generative model,the point cloud fractal network PF-Net(Point Fractal Network),which is deployed to the task of missing point cloud completion.(1)In order to cope with the unordered as well as transformation invariance in the abstraction process of unstructured point cloud features,PF-Net abstracts missing point clouds into representational features with high integration of geometric and semantic information by means of a multi-level point cloud feature encoder based on key points.In this paper,we deeply investigate multi-scale point cloud features,multi-resolution point cloud features,3D point cloud key point sampling algorithms,and combine symmetry functions to propose combined multi-layer perceptron for disorderly point cloud feature extraction.(2)In order to restore high-quality missing region point clouds on unstructured missing point clouds,PF-Net first innovatively proposes a multi-level decoder for point clouds based on the pyramid structure to achieve multi-level reconstruction of the representational features of missing point clouds.Secondly,this paper incorporates the ideas of multilevel disordered point set distance metric,generative adversarial network,and point cloud surface density consistency metric on the loss function,so as to optimize the training of PF-Net with high efficiency.(3)For the scarcity of missing point cloud annotation data,this paper proposes to use the planar projection method and random region elimination method to simulate and emulate the missing point cloud data for a large-scale synthetic point cloud dataset to achieve data enhancement of missing point clouds.The above method enables the point cloud fractal network PF-Net to achieve high precision completion of geometric details of missing regions while preserving the spatial distribution of point clouds.In this paper,we design comparative experiments,ablation experiments,and robustness tests to demonstrate the advancement and effectiveness of PF-Net in multiclass point cloud completion tasks.In order to verify the technology in real scenes,this paper proposes a point cloud completion system based on the point cloud fractal network PF-Net for sparse 3D reconstruction scenes and a point cloud completion system for dense 3D reconstruction scenes,and tests and demonstrates the effect of the point cloud completion algorithm in real scenes for sparse point cloud data acquired by Li DAR and dense point cloud data reconstructed based on RGB images,respectively,to verify the possibility and effectiveness of the point cloud fractal network PF-Net migrated to the real world for missing point cloud completion.