Unmanned Aerial Vehicle (UAV) Vision-based Detection of Power Line Poles by CPU-based Deep Learning Method

More and more power supply companies use Unmanned Aerial Vehicles (UAV) for power line inspection. UAVs allow almost immediate inspection of power lines after extreme weather events. However, the current UAV vision based damage assessments have still been performed manually, which is time-consuming, poor efficient, and low accurate. In this work, a fast CPU-based detection model is presented for detecting normal and abnormal power line poles from the UAV vision data after typhoon striking. Three types of poles including two types of normal poles and one type of abnormal poles are considered. The detection process is designed in two stages. The first stage is to generate candidate boxes of poles based on the YOLO-Lite model, and the second stage is to filter the background candidate boxes based on the classification model of the SPP (Spatial Pyramid Pooling) network structure. The combined model achieves a detection precision of 75.80%, an increase of 26.85% compared to the YOLO-Lite model alone, and reaches a recall of 57.33%. The combined poles detection model runs at 9 FPS (Frames Per Second) on a CPU-only computer.