Advanced YOLO-based Real-time Power Line Detection for Vegetation Management

TL;DR

The paper tackles real-time, automated detection of power lines and surrounding vegetation from UAV imagery to enhance safety and resilience. It introduces PL-YOLOv8, an enhanced YOLOv8 model that employs a directional block and an $OBB(x,y,w,h,\theta)$-based head with ProbIoU loss for tight, orientation-aware localization, enabling high-speed inference suitable for onboard UAV processing. A vegetation encroachment metric is proposed by combining a Greenness Index (GI) derived from GRVI with a Tree-Grass Differentiation Index (TGDI), producing a quantitative risk score around detected lines. Validated on the TTPLA-derived TTPLA-Tile-OBB dataset, the approach achieves $AP_{50}=78.24\%$ and $FPS=97$ for the n-size model, while demonstrating an area-under-curve of $0.83$ for encroachment detection, signaling strong practical potential for real-time UAV-based vegetation management and risk mitigation.

Abstract

Power line infrastructure is a key component of the power system, and it is rapidly expanding to meet growing energy demands. Vegetation encroachment is a significant threat to the safe operation of power lines, requiring reliable and timely management to enhance the resilience and reliability of the power network. Integrating smart grid technology, especially Unmanned Aerial Vehicles (UAVs), provides substantial potential to revolutionize the management of extensive power line networks with advanced imaging techniques. However, processing the vast quantity of images captured by UAV patrols remains a significant challenge. This paper introduces an intelligent real-time monitoring framework for detecting power lines and adjacent vegetation. It is developed based on the deep-learning Convolutional Neural Network (CNN), You Only Look Once (YOLO), renowned for its high-speed object detection capabilities. Unlike existing deep learning-based methods, this framework enhances accuracy by integrating YOLOv8 with directional filters. They can extract directional features and textures of power lines and their vicinity, generating Oriented Bounding Boxes (OBB) for more precise localization. Additionally, a post-processing algorithm is developed to create a vegetation encroachment metric for power lines, allowing for a quantitative assessment of the surrounding vegetation distribution. The effectiveness of the proposed framework is demonstrated using a widely used power line dataset.

Figures (15)

• Figure 1: Outline of PL-YOLO based power line detection for vegetation management.
• Figure 2: Enhanced YOLO structure: PL-YOLOv8 with directional block.
• Figure 3: OBB task: (a) OBB representation. (b) Power line detection example using the OBB concept.
• Figure 4: Directional block containing 8 directional high-pass filters in its first stage.
• Figure 5: 2D filter rotation example when $n=5$. (a) Filter $f_{\theta=0^\circ}(i,j)$. (b) Filter $f_{\theta=25.56^\circ}(i,j)$.