针对电力场景下因激光雷达扫描范围有限、电力设备部件相互遮挡等因素导致电力设备部件点云缺失的问题,提出了基于点特征传递的电力设备部件激光点云补全网络PPC-Net(Power Point cloud Complete Net).该网络使用多尺度特征融合编码器提取不同尺度输入残缺点云的全局和局部特征,以避免多维度映射特征导致的电力设备部件细节特征丢失问题,并引入EdgeConv加强对点云邻域信息提取;在精细完整点云生成阶段提出DT模块整合父级点到子级点的特征传递,以保留生成点云的局部特征;设计平滑优化模块,经三级平滑采样算法输出分布均匀、表面平滑的电力设备部件完整点云.在自建电力设备部件点云数据集ELE及公开数据集PCN上实验表明,PPC-Net对残缺的电力设备部件点云有较好的补全效果,并在一般形状点云上有良好泛化性.
LiDAR point cloud completion network for power equipment components based on point feature transform
Aiming at the problem of defective point cloud of power equipment components due to limited scanning range of LiDAR and mutual occlusion of power equipment components in power scenario,a power equipment compo-nent LiDAR point cloud completion network Power Point Cloud Complete Net(PPC-Net)based on point feature trans-form is proposed in this paper.A multi-scale feature fusion encoder is used to extract global and local features of de-fective point clouds at different scales to avoid the problem of losing detailed features of power equipment components caused by multi-dimensional mapping,and EdgeConv is used to enhance the extraction of neighborhood information from point clouds.Then,the DT module is proposed to integrate feature transfer from parent to child points during the generation stage of fine and complete point clouds in order to preserve the local features of the generated point cloud.Next,a smooth optimization module is designed to output a complete point cloud of power equipment components with uniform distribution and smooth surface through three-level smooth sampling algorithm.Experiments on the self-built power equipment component point cloud dataset ELE and the public dataset PCN show that PPC-Net has a good com-pletion effect on defective power equipment component point clouds and good generalization on the general shape point clouds.
NETL
NSTL
万方数据
