Efficient Semantic Segmentation of Transmission Line Corridor Point Clouds via Transformer With Vector Aggregation

Ensuring precise segmentation of point clouds is essential for intelligent inspection in transmission line corridors. The massive scale, unordered distribution, and complex structures of these point clouds present significant challenges. Effective segmentation requires both local detail extraction and global context modeling. However, most existing methods learn local and global features independently, where local features are extracted from scalar information and global transformers require high computational resources. To solve these issues, this article introduces the Voxelized Random Sampling and Shifted Transformer with Vector Aggregation Network (PointVSV), a segmentation network designed specifically for power transmission corridors. PointVSV incorporates a voxelization and random sampling strategy (VRS) to retain structural detail while improving efficiency. Building on this foundation, the shifted patch transformer module (SPTM) and vector aggregation module (VAM) work together to combine robust global context modeling with fine-grained local feature extraction, enabling highly accurate semantic segmentation. To evaluate the performance of PointVSV, the EHV3D dataset was constructed for transmission line corridors, and comprehensive experiments were conducted. On this dataset, PointVSV achieves 98.0% overall accuracy (OA) and 97.0% mean Intersection over Union (mIoU), while using only 55% of the parameters of Point Transformer and delivering 2.8 × higher inference throughput. It also demonstrates strong generalization on the Toronto-3D dataset, with IoU scores of 88.2% for utility lines and 81.7% for poles. Based on PointVSV segmentation results, we further develop methods for single conductor extraction and safety distance detection, advancing intelligent inspection workflows.