StripFormer-3D: A Structure-Prior and Dual-Strip Neighborhood Attention Transformer for Power Corridor Point Cloud Segmentation

Point cloud segmentation of outdoor power facilities is essential for intelligent and automated inspection in large-scale transmission corridors. However, slender structures such as power lines and poles remain challenging due to complex tower geometry, severe point-density variation, and cluttered backgrounds, which often lead to discontinuities and missed small objects. We hypothesize that explicitly injecting structural priors and anisotropic neighborhood interactions can enhance both the recognition and connectivity of slender facilities in outdoor scenes. To this end, we propose StripFormer-3D, a structure-aware Point Transformer that integrates structural-aware positional encoding (SAPE) using principal directions and anisotropy, bi-strip neighborhood attention (BSA) that models longitudinal and transverse strip neighborhoods with gated fusion, and connectivity smoothing regularization (CSR) to enforce local prediction consistency. Experiments on a self-collected UAV-LiDAR dataset and a public power-facility dataset demonstrate that StripFormer-3D achieves higher segmentation accuracy and better structural coherence than strong Transformer baselines, with particularly improved performance on slender facility classes. These results indicate that combining structural priors, anisotropic attention, and connectivity regularization provides a reliable solution for power-corridor point cloud segmentation, supporting practical automated inspection.