Abstract: Semantic segmentation of point clouds in transmission corridor remains a challenging task due to strong noise interference, the difficulty of accurately identifying linear structures, and severe class imbalance, all of which result in unstable segmentation performance. To address these issues, this paper proposes a robust semantic segmentation method for transmission corridor point clouds. The proposed approach integrates three key modules: k-nearest neighbor (KNN)-based noise suppression, ellipsoidal structure-aware convolution, and the Focal Loss function. These modules collaboratively enhance the model's robustness to noise, structural completeness in feature representation, and discrimination capability for minority classes. Extensive experiments were conducted on a real-world 500 kV high-voltage transmission corridor dataset and the publicly available ECLAIR dataset. Experimental results demonstrate that the proposed method achieves superior performance compared to state-of-the-art approaches such as KPConv and RandLA-Net in terms of overall accuracy (OA) and mean intersection over union (mIoU). Specifically, on the 500 kV dataset, the proposed method improves the mIoU by 9.5% over RandLA-Net and by 5.9% over KPConv. The method significantly enhances the recognition of linear components and small-category objects, exhibiting excellent accuracy, stability, and generalization capability. This work was supported by the Technical Foundation of State Grid Corporation of China under Grant (52230025000F-191-ZN).

Key words: transmission corridor, semantic segmentation, robustness, airborne lidar point cloud, class imbalance