TL-PVCNN: A Point Cloud Semantic Segmentation Framework on Transmission Line Scene

The transmission line is an important part of the power system. The point cloud semantic segmentation in transmission line scene can automatically segment the important equipment such as towers and transmission conductors, so that it has an understanding of the power scene. We propose a Transmission Line Point-Voxel CNN (TL-PVCNN) neural network framework for point cloud semantic segmentation in the transmission line scene. TL-PVCNN aims to solve three difficulties, and the three difficulties are information loss caused by uneven point cloud density, segmentation difficulty caused by the complex and changeable geographical environment, and difficulty in network convergence due to large gaps between foreground and background point clouds. To solve these problems, we propose our solutions in three processes, which are point clouds data processing, neural network structure construction, and loss function design. In point clouds data processing, we propose a method named Block Density Importance Sampling (BDIS) to solve the problem of information loss caused by uneven point cloud density. In neural network structure construction, we proposed a module named Attention PV-Conv to solve the problem of segmentation difficulty caused by a complex and changeable geographical environment. In loss function design, we proposed Improved Point Dice Loss to solve the problem of difficulty in network convergence due to large gaps between foreground and background point clouds. TL-PVCNN achieves mIoU of 0.81 on the transmission line point cloud dataset, specifically IoU of 0.83 on towers and IoU of 0.84 on transmission conductors. The experiment result is better than most the point cloud semantic segmentation neural network models.