A Method for Extracting Power Lines and 3D Reconstruction Based on the Improved PointNet++

To address the issues of noise interference,the class imbalance, and insufficient 3D reconstruction accuracy in airborne LiDAR power line point cloud extraction, this paper proposes a method for power line extraction and 3D reconstruction based on an improved PointNet++. Firstly, the dataset is divided, and then multi-dimensional features such as point cloud spatial information and geometric features are extracted through various means. Next, PointNet++ is improved by introducing local relative height features to enhance vertical dimension information. A three-level feature extraction network is constructed, and Focal Loss is adopted in the training stage to solve the class imbalance problem. Dynamic learning rate adjustment and early stopping mechanisms are implemented with power line IoU as the monitoring metric. Finally, the DBSCAN algorithm is used to fit the parametric equation of the power line, and continuous frame point cloud registration is combined to achieve 3D reconstruction of the power corridor. Experimental results show that on a power line point cloud dataset with 4096 points sampled, this method achieves a power line IoU of $\mathbf{9 4 . 9 \%}$ and an overall mIoU of $\mathbf{9 7 . 2 7 \%}$ on the validation set.