基于PointNet++的建筑物室内部件分割方法  

Interior Component Segmentation Method for Buildings Based on PointNet++

作　　者：高飞 GAO Fei(Jiangxi University of Science and Technology,Ganzhoui 341000,China)

机构地区：[1]江西理工大学土木与测绘工程学院,江西赣州341000

出　　处：《长江信息通信》2025年第2期192-195,共4页Changjiang Information & Communications

摘　　要：三维激光点云能够直观完整地表达建筑物室内物体的总体构造,但是室内点云数据具有无序、庞大的特点。扫描获得的室内三维点云数据若直接用于室内物体部件分割需要耗费大量的时间,增加项目的运行成本。因此,如何准确快速的分割室内物体是点云分割的研究热点。点云中具有三维坐标信息、RGB色彩信息和深度信息等,通过对这些信息的分析研究,有助于对室内物体进行部件分割。为此,文章提出了一种基于PointNet++改进的建筑物室内部件分割方法F-PointNet++,该方法主要分为3个步骤,首先,对数据集进行预处理,为原始点云数据加入曲率信息;其次,改进PointNet++算法,加入曲率作为特征,重构深度网络;最后,通过F-PointNet++网络训练,使用平均准确度(eval accuracy)和平均交并比(mIoU)作为评估指标,验证F-PointNet++网络在室内点云部件分割问题上的性能。结果表明,该方法能够有效提高PointNet++在室内部件分割上的能力,提取精度良好。3D laser point cloud can intuitively and completely express the overall structure of indoor objects in buildings,but indoor point cloud data has the characteristics of disorder and vastness.If the indoor 3D point cloud data obtained from scanning is directly used for indoor object component segmentation,it will consume a lot of time and increase the operating cost of the project.Therefore,how to accurately and quickly segment indoor objects is a research hotspot in point cloud segmentation.Point clouds contain three-dimensional coordinate information,RGB color information,and depth information.By analyzing and studying this information,it is helpful for component segmentation of indoor objects.Therefore,this article proposes an improved building interior component segmentation method F-PointNet++based on PointNet++.The method mainly consists of three steps.Firstly,the dataset is preprocessed to add curvature information to the original point cloud data;Secondly,improve the PointNet++algorithm by incorporating curvature as a feature to reconstruct the deep network;Finally,the performance of the F-PointNet++network in indoor point cloud component segmentation problems was validated through training with the F-PointNet++network,using eval accuracy and mIoU as evaluation metrics.The results indicate that this method can effectively improve the ability of PointNet++in indoor component segmentation,with good extraction accuracy.

关 键 词：深度学习 三维彩色点云 室内部件 部件分割 

分 类 号：K90[历史地理—人文地理学]