基于改进PointNet++的输电线路关键部位点云语义分割研究

Research on Semantic Segmentation of Point Cloud for Key Parts of Transmission Lines Based on Improved PointNet++

输电线路的关键部位包括塔身、导线、绝缘子、避雷线以及引流线,无人机精细化导航的首要任务是构造输电线路的点云地图并从中分割出上述部位。为解决现有算法在输电线路的绝缘子、引流线等精细结构分割时精度低的问题,通过改进PointNet++算法,提出了一种面向输电线路精细结构的点云分割方法。首先,基于无人机机载激光雷达在现场采集的点云数据,构造了输电线路点云分割数据集;其次,通过对比实验,筛选出在本输电线路场景下合理的数据增强方法,并对数据集进行了数据增强;最后,将自注意力机制以及倒置残差结构和PointNet++相结合,设计了输电线路关键部位点云语义分割算法。实验结果表明:该改进PointNet++算法在全场景输电线路现场点云数据作为输入的前提下,首次实现了对引流线、绝缘子等输电线路中精细结构和导线、杆塔塔身以及输电线路无关背景点的同时分割,平均交并比(mean intersection over union,mIoU)达80.79%,所有类别分割的平均F_(1)值(F1 score)达88.99%。

The key components of a power transmission line include tower structure,conductor,insulator,lightning arrester,and grounding wire.The primary task of precise navigation for unmanned aerial vehicle is to construct a point cloud map of the transmission line and to segment the aforementioned components from it.To solve the problem of low accuracy in existing algorithms for segmentation of fine structures such as insulators and drainage lines in transmission lines,we propose a point cloud segmentation method for fine structures of transmission lines by improving the PointNet++algorithm.First,the point cloud data collected by unmanned aerial vehicle airborne LiDAR on site are constructed as a point cloud segmentation dataset for power transmission lines.Then,a reasonable data augmentation method in this transmission line scenario is selected through comparative experiments and applied to this dataset.Finally,the self attention mechanism and inverted residual structure have been applied in the PointNet++algorithm,completing the design of the semantic segmentation algorithm for key point clouds in transmission lines.Under the premise of using point cloud data as input on the entire scene transmission line site,the experimental results show that the improved PointNet++algorithm achieves simultaneous segmentation of fine structures,wires,tower bodies,and irrelevant background points in transmission lines such as drainage lines and insulators.The average intersection over union(mIoU)reaches 80.79%,and the average F_(1) score for all category segmentation reaches 88.99%.