TLPT-Net: A point transformer network dedicated to semantic segmentation of transmission line corridors

Abstract

The reliable operation of transmission lines is an essential prerequisite for ensuring the quality of people’s lives. With the development of technology, the digital transformation of transmission lines is a significant development proposition at present. Obtaining point clouds through Unmanned Aerial Vehicle (UAV) equipped with Light Detection and Ranging (LiDAR) provides important support for the digital transformation of large-scale transmission lines. However, existing research on semantic segmentation of point clouds in transmission line corridors still has limitations, such as the lack of segmentation research on insulator strings, a significant but small device, and insufficient attention to deep-learning models. Therefore, we proposed the Transmission Line Point Transformer Network (TLPT-Net), dedicated to semantic segmentation of transmission line corridors, which divided transmission line corridors into five types of point clouds: ground, ground line, insulator string, pylon, and transmission line. We considered the morphological and quantitative characteristics of point clouds of different categories in transmission line corridors, and we proposed Mixed Linear Edge Attention module, Topology layer, and Rectangular Grid Pooling strategies. In addition, we proposed some targeted model optimization strategies, including modifying patch grouping rules, modifying loss function, and modifying serialization encoding rules. The experimental results showed that the mean Intersection over Union (mIoU), mean Accuracy (mA), and Overall Accuracy (OA) indicators of TLPT-Net in the segmentation task of the transmission line corridor point cloud on the test set were 96.47%, 98.42%, and 99.52%, respectively. Compared to the existing state-of-the-art models for point cloud segmentation, TLPT-Net has improved mIoU by at least 3.48% and 3.44% in the validation and test sets, respectively. This work can provide support for digital operation and maintenance of transmission lines, such as tree barrier hazard assessment and safety distance evaluation for robot live-line work.