Semantic segmentation of transmission line scene point clouds based on RandLA-HiAtt

LiDAR point cloud semantic segmentation is a core technology for intelligent grid inspection of power transmission lines. However, point clouds in this scenario are characterized by extreme sparsity, complex topology, and severe class imbalance, which significantly limit the performance of existing models. To address these limitations, we propose an improved model named RandLA-HiAtt, based on RandLA-Net. The core enhancements of the model are twofold: First, a hierarchical attention mechanism is designed, incorporating a serial channel-spatial attention module at shallow layers to enhance local geometric feature extraction, and a channel-convolutional attention module at deeper layers to strengthen global semantic discrimination. Second, a multihead attention pooling module and a triple feature fusion module are introduced to preserve complex spatial topology while effectively mitigating the semantic attenuation problem within the network. Experiments on a self-built 500 kV transmission line datasets show that the proposed model achieves intersection over union (IoU) of 98.54%, 99.87%, and 87.81% for segmenting pylons, power lines, and insulator strings, respectively. The overall mean intersection over union and overall accuracy for the complete scene reach 91.46% and 96.18%, representing improvements of 3.27% and 1.82% over the baseline RandLA-Net, respectively. To validate its cross-scenario generalization capability, the model is trained and evaluated on public point cloud datasets such as Semantic3D and S3DIS. The results demonstrate that the model maintains excellent segmentation performance, further confirming its effectiveness and robustness.