Unsupervised Point Cloud Segmentation for Power Corridor Line Inspection

In recent years, learning based methods are introduced to identify the elements in the point cloud, which are easily captured with UAV and LiDAR. However, supervised learning needs great efforts in data labeling, and cannot well adapt to the point clouds in different places. In this paper, we propose an unsupervised point cloud segmentation method that can can distinguish identical elements from the point cloud. Firstly, with the collected point cloud, pre-processing including denoising and downsampling is conducted. After that, cloud elements such as transmission lines, towers, buildings, roads and vegetation are extracted with unsupervised algorithms including DBSCAN, and other point cloud processing algorithms such as cloth filtering and PCA. To be specific, intrinsic features, e.g., the evaluation, the density, the curvature, of the point clouds are used in the unsupervised learning and clustering. Finally, with the safety distance rules, potential danger points that locate close to the transmission lines are detected, and a danger point report is further generated. In addition, based on the algorithm, a point analysis and management system is developed. With a B/S architecture, the system provide the function of point cloud uploading, viewing, analysis, and danger point prediction. The research results is hoped to be applied in grid enterprise to improve the intelligent level of power grid management and maintenance.