Journal article · 2026
Real-Time UAV-Based Oil Pipeline and Visual Anomaly Detection Using YOLOv26n: A Dataset and Edge-Deployment Study
Drones (MDPI)
A compact detection model running on the drone's own onboard computer finds pipelines and visible signs of leaks in aerial images in real time, at 18 frames per second and 92% detection accuracy (mAP@0.5).
Why it matters
Oil and gas pipelines run for hundreds of kilometres across desert and industrial land. Ground patrols and manned flights are slow and expensive. Drones can survey faster, but only if the detection runs on the drone itself, in real time, without sending video back to a server.
What we did
- Built a dataset of 6,127 drone images over desert, semi-urban and industrial areas, labelled for two classes: pipeline and anomaly or leak.
- Fine-tuned an off-the-shelf YOLOv26n detector, without changing its architecture.
- Deployed it on an NVIDIA Jetson Xavier NX onboard computer with TensorRT, and measured the complete pipeline from image to detection.
What we found
- 92.4% mAP@0.5 on held-out test images: 94% for pipelines and 86% for anomalies and leaks.
- Better than YOLOv10, YOLOv8 and YOLOv5 baselines trained on the same data.
- 18 frames per second on the onboard computer, fast enough for live inspection.
About the data
The images show critical infrastructure, so they are not publicly released. The paper documents the full training and deployment settings.
Cite this work
Hatem Magdy Keshk, Ayman M. Abdallah (2026). Real-Time UAV-Based Oil Pipeline and Visual Anomaly Detection Using YOLOv26n: A Dataset and Edge-Deployment Study. Drones (MDPI). https://doi.org/10.3390/drones10040255
@article{keshk2026realtime,
title = {Real-Time UAV-Based Oil Pipeline and Visual Anomaly Detection Using YOLOv26n: A Dataset and Edge-Deployment Study},
author = {Hatem Magdy Keshk and Ayman M. Abdallah},
journal = {Drones (MDPI)},
year = {2026},
doi = {10.3390/drones10040255},
}