powerline-vision 1.0.0

Automated detection of overhead power line cables from aerial imagery using YOLOv8

PowerLine Vision — Automated Detection of Overhead Conductors from Aerial Imagery

Overview

This project develops an end-to-end deep learning pipeline for the automatic detection of overhead power lines and transmission towers from aerial imagery. The system addresses a core challenge in electricity distribution asset management: identifying and mapping low-voltage conductors at scale, including those that are partially obscured or difficult to locate manually.

The pipeline combines YOLOv8 object detection with classical computer vision preprocessing (OpenCV) and produces confidence-scored bounding box detections suitable for downstream GIS integration and asset mapping workflows.

Live Demo: https://karanm5-powerline-vision.hf.space

---

Motivation

Electricity distribution networks span thousands of kilometres. Manual inspection and mapping of overhead conductors is time-intensive, expensive, and prone to gaps in coverage. Automated detection from aerial and UAV imagery offers a scalable alternative — enabling utilities to maintain accurate, up-to-date asset records and identify infrastructure that may be hidden by vegetation or environmental occlusion.

This project is directly motivated by real-world infrastructure intelligence challenges in the UK electricity distribution sector.

---

Dataset

This project uses the TTPLA Dataset (Transmission Towers and Power Lines from Aerial imagery), a publicly available annotated dataset containing aerial images of overhead power line infrastructure.

• Source: TTPLA Dataset — GitHub
• Classes: Power lines, Transmission towers
• Format: COCO-style annotations, converted to YOLO format for training
• Split: 70% train / 15% validation / 15% test

To download and prepare the dataset:

python data/download_data.py

---

Project Structure

powerline-vision/
├── README.md
├── requirements.txt
├── configs/
│   └── config.yaml              # Training and model hyperparameters
├── data/
│   ├── download_data.py         # Dataset download and preparation script
│   └── processed/               # YOLO-format annotations after conversion
├── notebooks/
│   ├── 01_data_exploration.ipynb   # Dataset statistics and visualisation
│   └── 02_results_analysis.ipynb  # Post-training evaluation and analysis
├── src/
│   ├── dataset.py               # Dataset loading, conversion, augmentation
│   ├── train.py                 # Training pipeline
│   ├── evaluate.py              # Evaluation: mAP, precision-recall, F1
│   └── utils.py                 # Shared utilities and visualisation helpers
├── scripts/
│   └── inference.py             # Run detection on new images or video
├── results/
│   └── plots/                   # Training curves, detection visualisations
└── tests/
    └── test_dataset.py          # Unit tests for data pipeline

---

Model Architecture

The detection backbone is YOLOv8n/s (Ultralytics), fine-tuned from COCO pretrained weights on the TTPLA dataset. YOLOv8 was selected for:

• State-of-the-art real-time detection performance
• Strong transfer learning capability from diverse pretraining
• Native support for custom dataset fine-tuning
• Efficient inference suitable for edge deployment scenarios

Preprocessing uses OpenCV for image normalisation, contrast enhancement (CLAHE), and augmentation (horizontal/vertical flip, mosaic, HSV jitter).

---

Results

Metric	Value
mAP@0.5	0.688
mAP@0.5:0.95	0.561
Precision	0.758
Recall	0.647
Inference speed (GPU)	4.9ms per image (Tesla T4)
Validation images	359
Instances evaluated	2,738

Trained for 50 epochs on YOLOv8s fine-tuned from COCO weights. Dataset: aerial power line imagery (cable detection). Hardware: Tesla T4 GPU.

---

Quickstart

1. Install dependencies

pip install -r requirements.txt

2. Download and prepare data

python data/download_data.py

3. Train

python src/train.py --config configs/config.yaml

4. Evaluate

python src/evaluate.py --weights results/best.pt --data configs/config.yaml

5. Run inference on a new image

python scripts/inference.py --source path/to/image.jpg --weights results/best.pt

---

Training Results

Training curves and detection visualisations from the completed run:

Sample detections on validation images:

---

Limitations and Future Work

• Detection performance degrades under heavy occlusion from tree canopy — a known challenge in UK distribution networks
• LiDAR point cloud fusion is identified as a key extension to improve detection of hidden conductors; this is a planned next phase of development
• GIS export of detected conductor coordinates is in scope for a future release

---

Author

Karan | MSc Data Analytics (Distinction), Aston University
LinkedIn | GitHub

---

Acknowledgements

• TTPLA dataset authors for making aerial power line imagery publicly available
• Ultralytics for the YOLOv8 framework