The problem
Inspection companies fly drones and analyse the images afterwards, which delays repairs. Real-time detection needs a model on the drone (limited power) or a fast link to an edge server; the right split depends on image size, link quality and battery.
Who needs it
Solar farms, electricity grid operators, inspection service companies.
Idea of the approach
- Use public datasets of solar-panel defects and power-line components.
- Train a lightweight detector (YOLO family) and export small/medium versions.
- Measure accuracy, latency and energy on a Jetson/Raspberry Pi (on-board) and on a laptop (edge) with emulated links.
- Build a decision rule for where to run each frame.
What you will build and deliver
- Trained models and an inference demo on recorded flights.
- Trade-off study.
- Paper.
Skills needed
Python, computer vision basics, PyTorch.
Research contribution & where to publish
Contribution: an empirical on-board/edge decision model for inspection workloads. Venues: IEEE ICRA workshops, Drones (MDPI), IEEE Access, International Journal of Electrical Power & Energy Systems.
Key references — start reading here
- J. Redmon, S. Divvala, R. Girshick, A. Farhadi. You Only Look Once: Unified, Real-Time Object Detection. IEEE CVPR, 2016.
- V. N. Nguyen, R. Jenssen, D. Roverso. Automatic autonomous vision-based power line inspection: A review of current status and the potential role of deep learning. International Journal of Electrical Power & Energy Systems, 2018.
- Y. Kang, J. Hauswald, C. Gao, A. Rovinski, T. Mudge, J. Mars, L. Tang. Neurosurgeon: Collaborative Intelligence Between the Cloud and Mobile Edge. ASPLOS, 2017.
Interested?
Send the topic title and the courses you have taken.
Contact Dr. Hamza Djigal