多输入特征约束的电力走廊点云分割网络

打开文本图片集

中图分类号：TP391.4 文献标识码：A 文章编号：2096-4706（2025）20-0149-06

Abstract： The correct segmentation of the point cloud of the power corridor is crucial to power inspection tasks.Most existingDeepLeaming methodsofteninputiformationsuchascoordinates，color，intensityandechoasaunifedfeature tothe networkinthepointcloudsegmentationofpowercorriors，whichmakes itdificultforthemodeltolearefectiefeaturesand amplify noiseinterference.Inthisegard，amulti-eaturecontrolmoduleisproposed toacheveweightcontrolofiputfeatures to enhance efectivefeatures and suppressnoise efects.Inaddition，the main network adopts a decoupled structure，which makes themodel paymoreatentionto thelocalgeometric featuresandtheirfusion with thesemanticfeatures beforeandafter downsampling，thus effctively improving thenetwork performance.Theexperimentalresultsonthepowercorrdorscenario show that the proposed multi-feature control module can bring 2%～3% Mean Intersection over Union （mIoU） improvement to the backbone network，and the overall segmentation performance is 4% higher than the PointNeXt model， indicating the advancement and effectiveness of the method.

Keywords： power corridor; point cloud; semantic segmentation; decoupling; multi-feature input

0 引言

随着我国电力系统基础设施的不断完善，架空输电线路总里程持续增长，长距离输电线路巡检的重要性也日益凸显。（剩余9482字）