Road Anomaly Detection - Model Training
Complete training pipeline for road anomaly detection using YOLOv8, optimized for NVIDIA RTX 2050 (4GB VRAM) and Raspberry Pi 4 deployment.
🚀 Quick Start
# 1. Setup environment
chmod +x setup_training.fish
./setup_training.fish
# 2. Activate environment
source .venv/bin/activate.fish
# 3. Verify your dataset
python verify_dataset.py
# 4. Train model
python train_road_anomaly_model.py
# 5. Package for Raspberry Pi
python package_for_rpi.py
📁 Project Structure
model/
├── setup_training.fish # Environment setup (run first)
├── train_road_anomaly_model.py # Main training script
├── verify_dataset.py # Dataset validation
├── inference.py # Run inference on images/video
├── package_for_rpi.py # Create RPi deployment package
├── plan.txt # Detailed training plan
└── README.md # This file
../dataset/dataset/ # Your dataset (YOLO format)
├── data.yaml
├── train/images/ train/labels/
├── valid/images/ valid/labels/
└── test/images/ test/labels/
🎯 RTX 2050 Optimization
The training is pre-configured for 4GB VRAM:
| Parameter | Value | Reason |
|---|---|---|
| Batch Size | 8 | 4GB VRAM safe |
| Image Size | 416 | Edge-optimized |
| AMP | Enabled | 30-50% faster |
| Workers | 4 | 8-core CPU / 2 |
| Cache | Disabled | Save RAM for GPU |
If you get OOM (Out of Memory):
Edit train_road_anomaly_model.py:
BATCH_SIZE = 4 # Reduce from 8
IMAGE_SIZE = 320 # Reduce from 416
📊 Dataset Format
YOLO format with normalized coordinates:
<class_id> <x_center> <y_center> <width> <height>
Example label file:
0 0.5 0.5 0.2 0.15
1 0.3 0.7 0.1 0.08
Classes (default):
- 0: pothole
- 1: crack
- 2: bump
- 3: obstacle
- 4: road_damage
🔧 Scripts
1. Setup Environment
./setup_training.fish
Installs PyTorch with CUDA, Ultralytics, and all dependencies.
2. Verify Dataset
python verify_dataset.py [path/to/dataset]
Checks dataset structure, validates annotations, shows class distribution.
3. Train Model
python train_road_anomaly_model.py
Trains YOLOv8n model with RTX 2050 optimizations, exports to TFLite.
4. Run Inference
# Image
python inference.py --model road-anomaly-detection/train_*/weights/best.pt --source image.jpg
# Video
python inference.py --model best.pt --source video.mp4
# Camera
python inference.py --model best.pt --source camera
# TFLite model
python inference.py --model best.tflite --source image.jpg
5. Package for Raspberry Pi
python package_for_rpi.py
Creates deployment package with TFLite model and detection script.
📦 Output
After training:
road-anomaly-detection/
└── train_YYYYMMDD_HHMMSS/
├── weights/
│ ├── best.pt # PyTorch model
│ └── best.tflite # TFLite for RPi ⭐
├── training_report.txt # Performance metrics
├── results.png # Training curves
├── confusion_matrix.png # Per-class accuracy
└── sample_predictions/ # Test predictions
🍓 Raspberry Pi Deployment
# Transfer to RPi
scp -r rpi_deployment_*/ pi@raspberrypi.local:/home/pi/
# On RPi
cd rpi_deployment_*/
pip install -r requirements.txt
python detect.py --source 0 # Webcam
Expected performance:
- RPi 4: 10-12 FPS at 416x416
- RPi 5: 15-20 FPS
📈 Expected Results
| Metric | Target |
|---|---|
| mAP@50 | ≥ 0.80 |
| Precision | ≥ 0.75 |
| Recall | ≥ 0.75 |
| Model Size | ≤ 10 MB |
| RPi FPS | 10-12 |
🛠 Troubleshooting
CUDA Out of Memory
# In train_road_anomaly_model.py
BATCH_SIZE = 4
IMAGE_SIZE = 320
Training Slow
# Check GPU utilization
nvidia-smi
# Should see 85-95% GPU usage
# If low, increase batch size
No GPU Detected
# Check NVIDIA driver
nvidia-smi
# Check PyTorch CUDA
python -c "import torch; print(torch.cuda.is_available())"
# Reinstall PyTorch
uv pip install torch --index-url https://download.pytorch.org/whl/cu121 --force-reinstall
📚 References
📄 License
MIT License