Analysis and modeling of transmission lines based on point clouds are crucial for applications such as intelligent line inspection and lightning protection analysis. A common approach involves voxelizing the point cloud data and then applying 3D convolutional neural networks for feature extraction. However, such methods often sacrifice precise information during the voxelization process. To address this issue, this paper proposes utilizing kernel density estimation (KDE) to generate a smooth density distribution from collected point cloud information. Subsequently, sampling from this density distribution is performed to minimize information loss. The proposed method is validated in downstream classification tasks, demonstrating superior classification accuracy compared to voxelization-based methods.