Road Damage Detection - Project Overview
Mission Statement
Our AI-powered drone system revolutionizes road infrastructure maintenance by providing automated, efficient, and accurate road damage detection. By flying over roads and analyzing surface conditions in real-time, we enable proactive maintenance and improve road safety.
The Problem
Traditional road inspection methods face several challenges:
- Time-consuming: Manual inspections require significant labor and time
- Costly: Professional inspectors and specialized equipment are expensive
- Inconsistent: Human assessment can vary between inspectors
- Dangerous: Inspectors must work near traffic or on damaged roads
- Reactive: Damage often discovered after it becomes severe
- Limited coverage: Budget constraints limit inspection frequency
Our Solution
An autonomous drone equipped with the hardware listed below, running an AI model that currently detects potholes. The current implementation does not record GPS data; detections are saved locally via the AI pipeline in camera_control.py.
Hardware Components (Drone 3)
- Frame: SpeedyBee BEE35 Pro 3.5” CineWhoop Frame Kit
- Flight Controller: Flywoo GOKU F722 PRO V2 (STM32F722, 216MHz, 512kB Flash) 55A Stack 3-6S AM32 (Betaflight v4.5.2)
- Camera: Caddx Ratel Pro 1500TVL Analog (FPV)
- Video Transmitter: SpeedyBee TX800 VTX
- Channel: 5847 MHz (Band: BOSCAM/RichWave, Channel: 7)
- VTX Antenna: Foxeer Lollipop 4 RHCP
- Receiver: Radiomaster XR1 ELRS Dual Band RX (Firmware: ExpressLRS 3.6.0)
- Binding Phase: drone3
- Motors: Axisflying C206 2006 2500KV 4-6S
- Propellers: Gemfan 90mm D90-5 3.5” Ducted 5-Blade Propeller
- GPS: Matek M10Q-5883 GPS with Compass
Types of Road Damage Detected
1. Potholes (current model)
Description: Bowl-shaped depressions in the pavement surface, varying in size and depth.
Causes:
- Water infiltration through cracks
- Freeze-thaw cycles
- Traffic loading on weakened pavement
Severity levels:
- Low: Diameter < 200mm, depth < 25mm
- Medium: Diameter 200-500mm, depth 25-50mm
- High: Diameter > 500mm, depth > 50mm
System Workflow
┌─────────────────────────────────────────────────────┐
│ 1. FLIGHT PLANNING │
│ - Define road segments to inspect │
│ - Generate autonomous flight path │
│ - Set altitude and speed parameters │
└────────────────┬────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────┐
│ 2. AUTONOMOUS FLIGHT │
│ - Drone follows pre-programmed route │
│ - Maintains consistent altitude (5-10m) │
│ - Camera continuously captures images │
└────────────────┬────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────┐
│ 3. REAL-TIME IMAGE CAPTURE │
│ - Analog video frames captured (1500TVL) │
│ - Timestamp for temporal tracking │
└────────────────┬────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────┐
│ 4. AI DETECTION & CLASSIFICATION │
│ - Images processed by YOLO/TensorFlow Lite │
│ - Google Coral accelerates inference │
│ - Potholes detected and classified by severity │
│ - Detection screenshots saved │
└────────────────┬────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────┐
│ 5. DATA STORAGE & LOGGING │
│ - Detection screenshots saved │
│ - Metadata: timestamp and confidence score │
└────────────────┬────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────┐
│ 6. REPORT GENERATION (PLANNED) │
│ - Interactive map with damage locations │
│ - Statistical summary of road condition │
│ - Priority list for maintenance │
│ - Before/after tracking over time │
└─────────────────────────────────────────────────────┘
Technical Approach
Image Acquisition
- Altitude: 5-10 meters above road surface
- Speed: 2-5 m/s for optimal image quality
- Overlap: 60-80% forward overlap between images
- Capture source: Caddx Ratel Pro 1500TVL analog camera
- Processing resolution: Typically 720p/1080p (depends on capture pipeline)
- Frame rate: 5-10 FPS
AI Model Architecture
Option 1: YOLOv5/YOLOv8
- Single-stage detector for real-time processing
- Excellent speed-accuracy tradeoff
- Well-suited for edge devices
- Can detect multiple damage types simultaneously
Option 2: Faster R-CNN
- Two-stage detector for higher accuracy
- Better for small object detection (small cracks)
- Requires more computational resources
- May need post-processing optimization
Option 3: EfficientDet
- Balanced approach between speed and accuracy
- Efficient architecture for mobile/edge deployment
- Good performance on varied scales
Training Dataset
Required dataset characteristics:
- Minimum 5,000 labeled images for potholes
- Various lighting conditions (sunny, cloudy, shadows)
- Different road surface types (asphalt, concrete)
- Multiple viewing angles and altitudes
- Balanced representation of severity levels
Potential public datasets:
- RDD2020: Road Damage Dataset (global dataset)
- SDNET2018: Concrete crack dataset
- Custom dataset collection with our drone
Performance Metrics
- mAP (mean Average Precision): Overall detection accuracy
- Precision: Ratio of correct detections to all detections
- Recall: Ratio of detected damages to all actual damages
- F1-Score: Harmonic mean of precision and recall
- Inference time: Time to process one image (target: < 100ms)
- FPS: Frames processed per second (target: > 10 FPS)
Expected Benefits
For Municipalities
- Cost savings: 40-60% reduction in inspection costs
- Increased coverage: Inspect 10x more road length per day
- Better planning: Data-driven maintenance prioritization
- Historical tracking: Monitor road degradation over time
For Road Safety
- Early intervention: Fix problems before they worsen
- Hazard prevention: Identify dangerous potholes quickly
- Reduced accidents: Better maintained roads = safer driving
For Research
- Open-source platform: Replicable system for research
- Dataset creation: Contribute to road damage detection research
- Algorithm benchmarking: Test different AI approaches
- Infrastructure analytics: Study road degradation patterns
Challenges & Solutions
| Challenge | Solution |
|---|---|
| Variable lighting conditions | Image augmentation, HDR capture |
| Small crack detection | Shorter altitude, sharper optics, multi-scale detection |
| Real-time processing constraints | Edge TPU acceleration, model optimization |
| Location tagging | Not implemented yet (no GPS logging) |
| Shadow/water confusion | Additional spectral bands, temporal analysis |
| Battery life vs coverage | Efficient flight planning, battery swap stations |
| Wind stability for imaging | Gimbal stabilization, weather-aware scheduling |
Project Milestones
Month 1-2: Hardware setup and initial testing
- Assemble drone and test flight capabilities
- Install Raspberry Pi and Google Coral
- Camera calibration and image quality testing
Month 2-3: AI model development
- Collect/acquire training dataset
- Train initial damage detection models
- Evaluate model performance and iterate
Month 3-4: System integration
- Integrate AI inference with drone camera
- Implement autonomous flight with ArduPilot
- Develop data logging and storage system
Month 4-5: Field testing and refinement
- Test on actual roads with known damage
- Refine detection algorithms based on results
- Optimize flight parameters (altitude, speed, overlap)
Month 5-6: Documentation and presentation
- Create comprehensive technical documentation
- Develop demonstration materials
- Prepare final project presentation
Future Extensions
- Multi-spectral imaging: Detect subsurface damage
- 3D reconstruction: Create detailed surface topology
- Predictive maintenance: ML models to predict future damage
- Traffic integration: Coordinate with traffic signals for safe flights
- Fleet management: Multiple drones for large-scale surveys
- Mobile app: Real-time monitoring and reporting interface
| ← Back to Home | Next: Hardware Setup → |