一种鲁棒的多特征点云分类分割深度神经网络  

A Robust Deep Neural Network for Multi-Feature Point Cloud Classification and Segmentation

作　　者：田钰杰 管有庆[1] 龚锐 TIAN Yujie;GUAN Youqing;GONG Rui(School of Internet of Things,Nanjing University of Posts and Telecommunications,Nanjing 210023,China)

机构地区：[1]南京邮电大学物联网学院,南京210023

出　　处：《计算机工程》2021年第11期234-240,共7页Computer Engineering

基　　金：江苏省高校自然科学研究计划项目(05KJD520146)。

摘　　要：现有的深度学习方法在处理点云分割任务时,难以有效地学习点云的局部特征,存在分类分割精度低和鲁棒性差的问题。构建深度神经网络RMFP-DNN用于多特征点云分类分割。分别利用自注意力模块和多层感知机提取点云的局部特征和全局特征,并将两者相互融合,提高分类分割的准确率和鲁棒性。实验结果表明,RMFP-DNN平均分类准确率和整体分类准确率分别为88.9%和92.6%,与PointNet、PointNet++、DGCNN等方法相比,准确率较高且鲁棒性较好。The existing deep learning-based methods for point cloud classification and segmentation usually fail to learn the local features of point clouds,which limits their accuracy and robustness.To address the problem,a robust deep neural network,RMFP-DNN,is proposed for multi-feature point cloud classification and segmentation.The network employs a self-attention module to extract the local features of point clouds,and uses the Multi-Layer Perceptron(MLP)to learn the global features of point clouds.On this basis,the extracted local and global features are fused to improve the accuracy and robustness of classification and segmentation.Experimental results show that the average classification accuracy and overall classification accuracy of RMFP-DNN are 88.9%and 92.6%respectively.Compared with PointNet,PointNet++and DGCNN,RMFP-DNN achieves higher accuracy and better robustness.

关 键 词：深度学习 点云 鲁棒性 自注意力 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]