融合图注意力的多分辨率点云补全  

Multi-resolution point cloud completion fusing graph attention

作　　者：潘李琳 邵剑飞[1] PAN Lilin;SHAO Jianfei(School of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650504,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,昆明650504

出　　处：《激光技术》2023年第5期700-707,共8页Laser Technology

基　　金：国家自然科学基金资助项目(61732005)。

摘　　要：为了解决3维点云补全中难以提取点云局部特征信息的问题,提出了融合图注意力的多分辨率点云补全网络结构。采用了生成对抗网络框架处理数据的方法,生成器通过图注意力层构建点云图结构,融合不同分辨率的特征信息后加上网格数据,结合折叠操作重构缺失结构并输出逐级补全的点云数据;判别器判别点云真伪,通过反馈以提高准确度并优化生成器,使得生成数据具有精细的几何结构,近似于真实点云;在形状数据集上,将本文中的方法与其它4种方法进行比较,通过实验验证和理论分析,取得了最优的结果。结果表明,该方法能够有效地补全点云形状的缺失部分,得到完整且均匀的点云形状,相较于点分形网络性能提高约1.79%,对于实测数据的补全处理也达到了预期效果;所提出的点云补全网络结构,在提取点云全局形状特征的同时更好地提取了其局部几何特征信息,使得补全出的点云形状更加精细。该研究为智慧城市3维建模提供了参考。In order to solve the problem that it is difficult to extract the local feature information of point cloud in 3-D point cloud completion,a multi-resolution point cloud completion network structure based on fusion graph attention was proposed.The method of data processing with generative adversarial network framework was adopted.The structure of the point cloud image was constructed by the generator through the graph attention layer,the feature information of different resolutions with grid data was fused,and the folding operation was combined to reconstruct the missing structure and output the stepwise completed point cloud data.The truth and falsity of the point cloud was discriminated by the discriminator.The accuracy was improved through feedback,and the generator was optimized,so that the generated data has a fine geometric structure,which is similar to the real point cloud.The proposed method was verified experimentally and analyzed theoretically with four related methods on the shape dataset,and the optimal results were obtained.The results show that the proposed method can effectively complete the missing part of point cloud shape and obtain a complete and uniform point cloud shape,the network performance is improved by about 1.79%compared with point fractal network,the proposed method also achieves the expected effect on the completion of the measured data.The proposed point cloud completion network structure not only extracts the global shape features of point cloud,but also better extracts the local geometric feature information of point cloud,making the completed point cloud shape more refined.This study provides a reference for 3-D modeling of smart cities.

关 键 词：激光技术 点云补全 生成对抗网络 图注意力 折叠操作 3维点云 

分 类 号：TN958.98[电子电信—信号与信息处理] TP391.41[电子电信—信息与通信工程]