THRUST-AID: Clearing Object Detection Model
THRUST-AID YOLOv8-based Clearing Object Detection model will help investigate the surroundings of the powerline based on RGB aerial data and indicate potential threats
In the EU, the total length of overhead power lines reaches staggering 5.3 million km that need to be routinely inspected and maintained. However, current manual diagnostics practices are highly inefficient and prone to human error: walking crews focus on reactive maintenance, are inaccurate and slow, while massive amount of data collected by aviation is simply impractical to be analyzed by personnel.
Unattended prohibited activities in the protective zone are among the top risk-generating factors to the health of the overhead powerlines. This YOLOv8-based detection model will help investigate the surroundings of the powerline based on RGB aerial data and indicate potential threats. Among the detected objects are construction debris, heavy machinery, fallen trees, piling, storage of timber and other objects, industrial installations, farming activities, ground digging, etc.
These foreign objects are related to a wide spectrum of risks, including limited access for maintenance teams (due to blocked passage), fire hazard (if flammable materials are piled), short-circuit hazard (light objects can be windblown on conductor), physical damage to infrastructure (directly or via soil damage) or safety and health-related risks to people if safety regulations are not followed.
Timely detection of risk-inducing objects can prevent potential fatalities and damage to infrastructure. In combination with high-efficiency large-scale aerial data acquisition, this foreign object detection model can be a real game-changer in the aerial inspection and diagnostics of powerlines.
Model. This foreign object detection model is developed utilizing YOLOv8, a state-of-the-art computer vision framework renowned for its exceptional object detection capabilities.
Training materials. The model has been pretrained on COCO dataset and finetuned using extensive professionally annotated imagery dataset of more than 4000 ultra-high-resolution (<2 cm/px) oblique aerial photographs, collected from the Lithuania’s transmission power grid.
Training results. The model has been trained to detect a very wide variety of objects in the clearing area, from heavy machinery to timber piles or disturbed soil, that range widely in size and appearance. 89.7% average precision was reached. The model does not detect powerline infrastructure elements (pylons, etc.) and vegetation.
Input data. This detection model gives the best results when analyzing ultra-high-resolution oblique aerial imagery collected at large scale (e.g., using long-range fixed-wing drones) of powerlines in non-urbanized territories.
Testing. We provide a trained YOLOv8 model file (clearing-object-detection.pt). Please see README file and refer to Ultralytics webpage for more details on how to run prediction and other functions using the model.
This "THRUST-AID: Clearing Object Detection Model" is part of a project that has received funding from the European Union’s Horizon 2020 research and innovation programme under I-NERGY grant agreement No 101016508.
