Recognition of the Shape and Location of Multiple Power Lines Based on Deep Learning With Post-Processing

Abstract

Power line collisions pose a significant threat to the safety of drones. This is because it is difficult for drone pilots to recognize power lines at long distances, even on sunny days, and power lines are less visible in rainy or foggy weather. Therefore, power line detection is necessary for safe drone flight. This article proposes an algorithm that can recognize various shapes and locations of multiple power lines while improving the recognition performance of power lines compared to previous studies. YOLO, a deep learning technology used for object detection, is used to recognize power lines as multiple bounding boxes, and center points of these bounding boxes are sorted and integrated. This algorithm improves the power line detection performance by excluding incorrectly detected power lines and restoring undetected parts of the power lines. The performance of the proposed method was evaluated using the intersection-over-union (IoU) and F1-score, which were 0.674 and 0.528, respectively. This performance was superior to those of U-Net, LaneNet and BiSeNet V2 which are deep learning technologies for segmentation. The proposed method was mounted on the embedded system of the test drone, and tests were conducted indoor and outdoor. Then, the average frames per second (FPS) value was calculated as 10.05. Various shapes and locations of multiple power lines can be recognized in real-time using the power line recognition method proposed in this paper.

Keywords

• Power line detection
• continuous object
• segmentation
• agricultural spraying drone
• unmanned aerial vehicle (UAV)