Semantic Segmentation of Distribution Network Point Clouds Based on NF-PTV2

• DOI: 10.3390/electronics14040812
• Publication Date: 2025-02-19T13:36:46Z
• Authors: Long Han; Bin Song; Shaocheng Wu; Deyu Nie; Zhenyang Chen; Linong Wang

An on-site survey is the primary task of working live in distribution networks. However, traditional manual method not only very intuitive but also inefficient. The application 3D point cloud technology has opened up new avenues for surveys life This paper focused on Point Transformer V2(PTV2) model segmentation network clouds. Given its deficiencies boundary discrimination ability and limited feature extraction when processing clouds networks, an improved Non-local Focal Loss-Point V2 (NF-PTV2) was proposed. With PTV2 as core, this incorporated Non-Local attention to capturing long-distance dependencies, thereby compensating model’s shortcomings extracting features large-scale objects with complex features. Simultaneously, Loss function introduced address issue class imbalance enhance learning small samples. experimental results demonstrated that overall accuracy (OA) dataset reached 93.28%, mean intersection over union (mIoU) 81.58%, (mAcc) 87.21%. In summary, NF-PTV2 proposed article good performance can accurately identify various objects, which, some extent, overcomes limitations model.