Research On 3D Point Cloud Completion Network Based On Deep Learning

With the rapid development of computer 3D vision,3D point cloud data is widely used in industrial processing,automatic driving and other scenes.However,due to occlusion,angle limitation and other reasons,the lack of geometric spatial information often occurs,which brings difficulties to point cloud classification,component or scene segmentation and other applications.In recent years,with the development of deep learning technology about 3D point cloud,many solutions have been provided for this task.However,due to the limitation of GPU computing power,the existing deep learning network model has problems such as weak learning ability and insufficient performance in completing details.Therefore,this paper proposes a ’ two-step ’ completion method from incomplete point cloud to skeleton reconstruction,geometric skeleton to complete point cloud.The main research work is summarized as follows :(1)Aiming at the problem that the existing deep learning network model has weak completion ability under limited computing power,a skeleton reconstruction network(SR-Net)for point cloud completion is designed to learn and establish the mapping from incomplete point cloud to geometric skeleton.The improved point cloud skeleton point aggregation algorithm is used to extract the skeleton of the complete point cloud,and use it as a supervisory signal.In structural,we improvement Edge Conv encoder,fusion pyramid structure and Folding Net folding operation decoder are proposed.The final skeleton reconstruction visualization results show that it is feasible to learn the point cloud skeleton intensively under limited computing power and reconstruct the skeleton from the incomplete point cloud first.(2)Aiming at the problem that the traditional algorithm is not suitable for filling the closed single and island holes in the geometric skeleton,a point cloud completion network based on skeleton reconstruction(SRC-Net)is designed.After fusing skeleton points and incomplete points,the neighborhood filtering algorithm is used to remove outliers.In structural,we improvement CAE encoder,a decoder with a graph topology inference module and a graph filtering module to infer the complete point cloud.Finally,the completion visualization results show that the completion method in this paper can generate a more uniform and smooth distribution of high-quality complete point clouds,and can achieve more detailed completion effects.The experiments show that SR-Net can effectively reconstruct the geometric skeleton from the incomplete point cloud in advance,and RC-Net can recover the fine,uniform and complete point cloud from the skeleton point cloud.Therefore,the ’ two-step ’ point cloud completion method in this paper,from coarse to fine in the vertical direction,from skeleton to surface in the horizontal direction,realizes the fine completion of incomplete point cloud,and provides a new idea and method for point cloud deep learning completion,which has certain guiding significance.