基于PointNet++的工井点云语义分割模型研究  

Research on Semantic Segmentation Model of Well Point Cloud Based on PointNet++

作　　者：刘丹丹 胡伟 王丽欢 赵健 任雨 王迪 余容 LIU Dandan;HU Wei;WANG Lihuan;ZHAO Jian;REN Yu;WANG Di;YU Rong(Power China Guizhou Electric Power Engineering Co.,Ltd.,Guiyang 550000,Guizhou,China;State Grid Hebei Economic Research Instittute,Shijiazhuang 050000,Hebei,China)

机构地区：[1]中国电建集团贵州电力设计研究院有限公司,贵州贵阳550000 [2]国网河北省电力有限公司经济技术研究院,河北石家庄050000

出　　处：《电力大数据》2024年第2期77-86,共10页Power Systems and Big Data

摘　　要：受地下工井空间狭窄、环境复杂的影响,采集的工井点云数据存在空间分布不规则且不均匀、数据量大、难以实现多目标高效的语义分割等问题。有鉴于此,本文提出了一种基于PointNet++的工井点云语义分割方法。首先,采集地下工井点云数据并加入语义标签,制作模型训练需要的数据集;其次,为提高点云分割任务的性能,引入一种基于深度学习的PointNet++网络模型,并利用多分辨率分组(multi-scalegrouping,MSG)和随机输入(random inputdropout,DP)策略,实现了地下工井地面、顶、爬梯、墙、电缆线和支架的语义分割;最后,采用精确度、召回率、交并比和F1分数作为评价指标对分割效果进行评价。结果表明,与PointNet网络模型相比,本文方法各类别的评价指标均得到了显著的提升,并且地下工井地面、井顶、井墙和电缆线等类别的评价指标均超过80%,显示分割性能良好,有利于地下电缆工井场景的多目标快速精准分割,为地下工程精细化管理奠定了基础。Due to the narrow space of underground wells and complex environment,the distribution of collected well point cloud data is so irregular and uneven in space and the amount of data is so large that it is difficult to achieve multi-objective and efficient semantic segmentation.In view of this,this paper proposes a PointNet++based semantic segmentation method for well point cloud.Firstly,the data of underground well point cloud is collected and semantic labels are added to make the data set required for model training.Secondly,in order to improve the performance of the point cloud segmentation task,a deep learning-based PointNet++network model is introduced,and strategies of multi-scale grouping(MSG)and random input dropout(DP)are used to implement the semantic segmentation of ground,roof,ladder,wall,cable and support.Finally,accuracy,recall rate,the ratio of intersection over union and F1 score are used as evaluation indexes to evaluate the segmentation effect.The results show that compared with PointNet network model,the evaluation indexes of all categories of the proposed method have been significantly improved,and the scores of evaluation indexes of well surface,well top,well walls and cables are all over 80%,indicating good performance of the segmentation,which is conducive to multi-objective fast and accurate segmentation of underground cable well scene,and lays a foundation for the fine management of underground engineering.

关 键 词：PointNet++ 工井点云 语义分割 DP MSG 

分 类 号：TM75[电气工程—电力系统及自动化]