Abstract: Aiming at the existing problem that the point cloud extraction accuracy of transmission lines is not high and cannot meet the needs of autonomous and refined inspection by unmanned aerial vehicles, an improved PointNet++ semantic segmentation method for transmission tower point cloud is proposed, which realizes the segmentation of wires, ground wires, drainage lines, insulators and towers. Firstly, the parameters of the classic PointNet++ model are adjusted to make the model more suitable for the point cloud data of transmission towers in terms of feature extraction quantity and receptive field; then, the kernel point convolution is used as the point cloud feature extraction algorithm to further improve the model's ability to detect point cloud features; finally, for the data imbalance problem in the point cloud data, the focal loss is used as the loss function, so that the categories with a small proportion can be fully trained. In order to verify the effectiveness of the proposed method, experiments are carried out on the point cloud dataset composed of 2284 transmission towers. The experimental results show that the average F₁ value of the improved algorithm reaches 97.26%, which is 3.95 percentage points higher than that of the classic PointNet++.