双重注意力机制的电力走廊点云语义分割  

Dual attention for power corridor point cloud semantic segmentation

作　　者：李建[1] 王健[1,2] 王雷 李敏 杨立克 赵艺龙 LI Jian;WANG Jian;WANG Lei;LI Min;YANG Like;ZHAO Yilong(College of Geodesy and Geomatics,Shandong University of Science and Technology,Qingdao 266590,China;Qingdao Key Laboratory of Beidou Navigation and Intelligent Spatial Information Technology Application,Qingdao 266590,China;School of Geomatics and Urban Spatial Informatics,Beijing University of Civil Engineering and Architecture,Beijing 102612,China;State Grid Shandong Electric Extrahigh Voltage Company,Jinan 250118,China;Operation and Management Department of Shan jin Mining Co.,Ltd.,Songxian County,Luoyang 471400,China)

机构地区：[1]山东科技大学测绘与空间信息学院,山东青岛266590 [2]青岛市北斗导航空间信息技术重点实验室,山东青岛266590 [3]北京建筑大学测绘与城市空间信息学院,北京102612 [4]国网山东省电力公司超高压公司,山东济南250118 [5]洛阳市嵩县山金矿业有限公司运营管理部,河南洛阳471400

出　　处：《测绘通报》2025年第4期127-133,共7页Bulletin of Surveying and Mapping

基　　金：国家重点研发计划(2022YFB3903501);山东省自然科学基金(ZR2023MD027)。

摘　　要：电力走廊的点云场景具有独特性,采用深度学习方法进行语义分割时面临诸多挑战,场景中存在严重的数据倾斜问题。此外,在处理架空电力线和杆塔时,由于局部半径内点数量不足,模型难以提取足够的局部特征,从而降低了对相似对象的分割精度。为此,本文设计了一种基于双重注意力机制的两阶语义输电线路语义分割方法。首先,在数据精简阶段,利用电力传输设备与背景之间的高程差异,通过非深度学习方法有效去除大量背景点,加速训练过程并显著缓解数据倾斜问题。然后,提出兼顾全局特征与局部特征的双重注意力模型,提升了相似对象的区分度,且提高了点云分割的精度。经测试,本文的数据精简方法可去除63%以上的背景点,解决部分数据倾斜问题;提出的双重注意力网络对地线、导线和绝缘子的分割效果优于其他方法。The point cloud scene of power corridors is unique and faces a number of challenges when using deep learning methods for semantic segmentation.There is a serious data skew problem in the scene,with background objects such as ground and vegetation occupying the majority,which may lead to degradation of network performance.In addition,when dealing with point cloud data of overhead power lines and power towers,the model has difficulty in extracting enough local features due to the insufficient number of points within the local radius,which reduces the segmentation accuracy of similar objects.For this reason,this paper designs a two-order semantic transmission line semantic segmentation method based on the dual attention mechanism.Firstly,in the data streamlining stage,the elevation difference between the power transmission equipment and the background is utilized to have remove a large number of background points,thus reducing the training burden of the subsequent neural network,accelerating the training process,and significantly alleviating the data skew problem.Then,the dual-attention model,which takes into account both global and local features,is proposed to enhance the differentiation of similar objects and improves the accuracy of point cloud segmentation.After testing,the data streamlining method in this paper can remove more than 63%of background points and partially solve the data skewing problem;the proposed dual-attention network outperforms other methods for segmentation of ground wires,conductors and insulators.

关 键 词：点云 电力走廊 深度学习 双重注意力机制 语义分割 

分 类 号：P208[天文地球—地图制图学与地理信息工程]