电子测量与仪器学报

张丹丹,宋睿,任超,刘高华,李帅,李威.二维分割驱动的电力设施三维分割方法[J].电子测量与仪器学报,2026,40(3):81-92

二维分割驱动的电力设施三维分割方法

2D segmentation-driven 3D power facility reconstruction method

DOI：

英文关键词:3D segmentation 2D segmentation 3D reconstruction power facilities

基金项目:国家电网公司总部科技项目（5700-202440335A-2-1-ZX）资助

作者	单位
张丹丹	天津大学福州国际联合学院天津300072
宋睿	中国电力科学研究院有限公司北京100192
任超	天津大学电气与信息工程学院天津300072
刘高华	天津大学电气与信息工程学院天津300072
李帅	4.浙江大学机械工程学院杭州310030； 5.国网瑞嘉（天津）智能机器人有限公司天津300480
李威	国网瑞嘉（天津）智能机器人有限公司天津300480

Author	Institution
Zhang Dandan	School of International Joint Institute, Tianjin University, Tianjin 300072, China
Song Rui	China Electric Power Research Institute, Beijing 100192，China
Ren Chao	School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China
Liu Gaohua	School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China
Li Shuai	4.School of Mechanical Engineering, Zhejiang University, Hangzhou 310030, China; 5.State Grid Ruijia (Tianjin) Intelligent Robot Co., Ltd., Tianjin 300480, China
Li Wei	State Grid Ruijia (Tianjin) Intelligent Robot Co., Ltd., Tianjin 300480, China

摘要点击次数: 332

全文下载次数: 194

中文摘要:

面向电力机器人带电作业与设备巡检等任务需求，三维分割技术将复杂电力场景分解为多个独立的语义或实例级别的三维空间区域，为电力机器人提供了精准的空间认知，是保障电力系统自动化运维、提升带电作业安全性与效率的关键场景理解技术。现有方法因高成本的三维数据采集和模型实时处理能力不足，难以满足实际电力场景的需求。因此，提出了首个无需三维数据输入的低成本、快速三维分割框架，仅需以多视角的单目二维光学图像和目标提示词为输入就可实现电力设施的三维分割，有效解决了现有方法的局限性。首先利用Grounded-SAM2的强泛化特性以及所设计的颜色标注机制，获取输入图像中电力设施的RGB标识化分割掩码；随后，经Spann3r重建模型融合几何及颜色特征生成电力场景点云；最终，依据颜色筛选准则获得电力设施候选点云集，并基于DBSCAN密度约束聚类方法对候选点云进行处理，实现电力设施的实例区分与噪声抑制，输出精准的三维分割结果。实验结果表明，该方法在真实电力场景中验证可实现92.6%以上平均交并比（mIoU）与93.4%以上平均准确率（mAcc）的三维分割精度，为电力机器人智能化运维提供了一种无需三维传感器支持的低成本且快速的解决方案。

英文摘要:

Three-dimensional segmentation decomposes complex power environments into semantic and instance-level regions, enabling precise spatial awareness for power robots in automated and safer live-line operations. However, existing methods face high 3D data acquisition costs and limited real-time processing, hindering practical deployment. To address these challenges, we introduce the first low-cost, rapid 3D segmentation framework using only multi-view monocular images and text prompts, without requiring 3D sensors. Specifically, we leverage Grounded-SAM2’s robust generalization and a custom color-annotation scheme to produce RGB segmentation masks for each view. These masks are then fed into the Spann3r reconstruction model, fusing geometry and color to reconstruct a dense scene point cloud. Candidate point sets corresponding to power equipment are extracted via color filtering, and a density-constrained DBSCAN clustering step separates individual instances while suppressing noise. Experimental validation in real-world power scenarios shows that, using only monocular RGB imagery, our method achieves over 92.6% mean intersection over union (mIoU) and over 93.4% mean accuracy (mAcc) for 3D segmentation. It offers the low-cost, and fast solution for the intelligent operation and maintenance of power robots without 3D sensor support.

查看全文查看/发表评论下载PDF阅读器