This project is part of the Master's Program in Computer Science (Field of Study: Intelligent Systems) at Frankfurt University of Applied Sciences, Winter Semester 2025/2026, supervised by Prof. Dr. Christian Baun.
We are developing an AI-powered FPV drone system for automated road damage detection and assessment. The drone flies over roads and uses computer vision and deep learning models to identify road surface damage. The current implementation focuses on pothole detection using YOLOv8, with architecture designed to support additional damage types in the future.
Our drone autonomously patrols road infrastructure and performs real-time analysis of road surface conditions. By combining aerial imagery with AI-powered defect detection, we aim to:
- Automate road inspection: Replace manual inspections with efficient aerial surveys
- Early damage detection: Identify road damage before it becomes severe
- Cost reduction: Reduce inspection costs and improve maintenance planning
- Safety improvement: Minimize road hazards through proactive maintenance
- Data-driven decisions: Provide quantitative data for infrastructure management
- Develop and train AI model for pothole detection using YOLOv8
- Implement real-time detection system optimized for edge devices (Raspberry Pi + Google Coral TPU)
- Create autonomous flight paths for systematic road coverage
- Generate detailed damage reports with GPS coordinates and severity assessment
- Build comprehensive documentation for replication and research
- FPV Drone Frame with Flight Controller
- Raspberry Pi Zero 2 WH (Single-board computer)
- Google Coral USB Accelerator (AI inference)
- GPS Receiver
- Camera System
- FPV Transmitter and Goggles
- ELRS Receiver
- Electronic Speed Controllers (ESC)
- Motors and Batteries
- Flight Controller Firmware: Betaflight / INAV / ArduPilot
- Ground Control Station: QGroundControl
- Operating System: Raspbian / Ubuntu
- AI Framework: YOLOv8 (Ultralytics), TensorFlow Lite, PyTorch
- Programming Languages: Python, C++
Understanding the capabilities and limitations of all drone components.
Developed a YOLOv8 nano-based pothole detection system optimized for edge deployment:
Implementation Highlights:
- Model: YOLOv8n (nano variant) for real-time detection on resource-constrained devices
- Dataset: 4,510 annotated road images from Roboflow Universe
- Train: 3,993 images | Validation: 352 images | Test: 165 images
- Preprocessing: Histogram equalization, resizing to 840×840
- Augmentations: 3× data expansion with blur, noise, and grayscale variations
- Optimization: INT8 quantization reducing model size by ~75% for edge deployment
- Deployment: TensorFlow Lite with Google Coral TPU acceleration (10× faster inference)
- Current Focus: Pothole detection (single-class)
- Extensibility: Architecture supports additional damage types with appropriate training data
Technical Stack:
- Training: Ultralytics YOLOv8, PyTorch, Roboflow API
- Deployment: TensorFlow Lite, INT8 quantization, Edge TPU compiler
- Detection: Real-time video/webcam inference scripts
See ai-model/ directory for complete implementation and AI Documentation for details.
Implementing autonomous flight capabilities using INAV or ArduPilot.
Developing drop mechanisms for delivering inspection equipment.
Creating comprehensive online documentation via GitHub Pages.
Full documentation is available at: GitHub Pages Documentation
- Dominique Conceicao Rosario
- Muhammad Rizki Aulia Rahman
- Houda El Abbassi
# Clone the repository
git clone https://github.com/HoudaElAbbassi/ai-drone-ws2526.git
# Navigate to project directory
cd ai-drone-ws2526
# Install dependencies (detailed in documentation)
pip install -r requirements.txtai-drone-ws2526/
├── docs/ # GitHub Pages documentation
│ ├── hardware/ # Hardware setup guides
│ ├── software/ # Software configuration
│ ├── ai-applications/ # AI implementation guides
│ ├── road-inspection/ # Project overview
│ └── tutorials/ # Step-by-step tutorials
├── ai-model/ # AI model implementation
│ ├── train/ # Training scripts (train.py, validate.py)
│ ├── detect/ # Detection scripts (video, webcam)
│ ├── utilities/ # Model export and utilities
│ └── dataset/ # Training dataset (created when train.py is run)
├── camera_control.py # Camera control script
└── drop-mechanism.py # Payload delivery system
This project is licensed under the MIT License - see the LICENSE file for details.
- Prof. Dr. Christian Baun - Project Supervisor
- Frankfurt University of Applied Sciences
- Master's Program in Computer Science (Intelligent Systems)
For questions or collaboration:
- Professor: Prof. Dr. Christian Baun - christianbaun@fra-uas.de
- Website: www.christianbaun.de
Course: Master Project - Winter Semester 2025/2026 University: Frankfurt University of Applied Sciences Field of Study: Intelligent Systems