融合多维空间信息的三维点云语义分割方法  

3D Point Cloud Semantic Segmentation Method Fused with Multidimensional Spatial Information

作　　者：魏东[1] 张潇瀚 刘欢[1] 孙天翼 李昌恺 WEI Dong;ZHANG Xiaohan;LIU Huan;SUN Tianyi;LI Changkai(Shenyang University of Technology,Shenyang 110870,China)

机构地区：[1]沈阳工业大学,辽宁沈阳110870

出　　处：《长江信息通信》2023年第6期90-93,共4页Changjiang Information & Communications

基　　金：辽宁省教育厅科学基金资助项目(No.LJGD2020006)。

摘　　要：融合多维空间信息的三维点云语义分割方法以较低的时间复杂度处理场景级三维点云。针对局部特征不足导致感知能力下降的问题,通过引入法向量估计和混合池化用以增强局部特征。针对网络结构难以处理大规模三维点云的细微语义对象,通过使用长跳跃多层融合和短跳跃多维特征融合,可以保留多层特征多样性。经实验结果表明,在S3DIS(Stanford Large-Scale 3D Indoor Spaces Dataset)数据集6折交叉验证平均交并比为70.2%。The 3D point cloud semantic segmentation method fused with multi-dimensional spatial information can process scene level 3D point cloud with low time complexity.Aiming at the problem that the lack of local features leads to the decline of perception ability,the normal vector estimation and hybrid pooling are introduced to enhance the local features.For the network structure that is difficult to handle the subtle semantic objects of large-scale 3D point clouds,multi-layer feature diversity can be preserved by using long-hop multi-layer fusion and short-hop multi-dimensional feature fusion.The experimental results show that the average intersection and union ratio of 6-fold cross-validation in the S3DIS(Stanford Large-Scale 3D Indoor Spaces Dataset)dataset is 70.2%.

关 键 词：语义分割 多层融合 法向量估计 多维特征融合 混合池化 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]