DST-Pointnet++:基于Pointnet++改进的点云分类网络  

DST-Pointnet++:a Novel Point Cloud Classification Network Based on Pointnet++

作　　者：王子璇 任明武[1] WANG Zixuan;REN Mingwu(School of Computer Science&Engineering,Nanjing University of Science&Technology,Nanjing 210094)

机构地区：[1]南京理工大学计算机科学与工程学院,南京210094

出　　处：《计算机与数字工程》2022年第11期2497-2501,共5页Computer & Digital Engineering

摘　　要：点云提供了精确的空间位置信息而被广泛应用于环境感知领域。近年来,越来越多的工作尝试直接以点云作为输入进行特征提取,Pointnet[10]和Pointnet++[11]是这个方向的开创者,但Pointnet++没有考虑点云非均匀采样的问题。研究提出了DST-Pointnet++对其进行改进,通过核密度估计和非线性变换从点云中提取出逆密度因子,将其与原点云特征进行加权,得到了具有密度信息的点云特征,提高了边缘点对局部特征的贡献,改善了因点云分布不均造成的问题。经过在公开数据集上测试对比,结果表明DST-Pointnet++具有更好的准确率和鲁棒性。Point cloud provides accurate spatial location information and it is widely used in environmental perception area.More and more work attempts to extract features directly from point clouds in recent years,Pointnet[10]and Pointnet++[11]are the pioneers in this direction,but Pointnet++does not consider non-uniform sampling.The research proposes DST-Pointnet++to improve it.The paper extractes the inverse density factor from point cloud through kernel density estimation and nonlinear transformation,then weightes it with the original point cloud feature to obtain a new feature with density information.DST-Pointnet++increases the contribution of edge points to local features and improves the problem caused by points cloud non-uniform distribution.After testing and comparison on public data sets,the results show that DST-Pointnet++has better accuracy and robustness.

关 键 词：点云分类 深度学习 Pointnet++ 逆密度 

分 类 号：P413[天文地球—大气科学及气象学]