Deep Learning-based Power Line Corridor Intrusion Detection

In recent years, research into wide-area, high efficiency methods for inspecting obstacles in power line corridors has attracted widespread attention. Leveraging the cost effectiveness and ease of operation of drones, installing visible light digital cameras on these unmanned aerial vehicles (UAV s) to capture images of power line corridors and then detect potential hazards has become an emerging method of power line corridor obstacle inspection. During the inspection operation, drones fly along the power line's direction, capturing stereoscopic images of the corridor with a certain overlap. Using depth estimation and image segmentation networks, the three-dimensional reconstruction of power lines and surrounding potential invasive objects, such as trees and buildings, can be completed. Based on the 3D reconstruction, obstacles can be automatically identified and located by calculating the spatial distance between the point clouds of potential invasive objects and power lines. However, because power line corridors are often covered by dense vegetation, it presents significant challenges for depth estimation and image segmentation. Additionally, in the process of reconstructing power lines in 3D, accurately reconstructing them is difficult with conventional stereo matching methods since power lines appear as thin lines with singular texture in images. This paper, using drone imagery and targeting the positioning and quantitative evaluation of obstacles within power line corridors, studies methods of dense point cloud extraction for power line corridors, automatic 3D coordinate measurement of power lines, and point cloud measurement of obstacles