Real-time traffic sign detection for legacy hardware. Runs at 14 FPS on a 2015 MacBook CPU and 89 FPS on T4 GPU.
Modern detection models assume you have a GPU. Most edge devices don't. This project squeezes YOLOv11 onto an old Intel Broadwell CPU by:
- Quantizing to INT8 (38% smaller, 6× faster)
- Threading video I/O separately from inference
Result: usable real-time detection without any GPU.
| Metric | Value |
|---|---|
| FPS (416px) | 14.0 |
| FPS (640px) | 7.4 |
| Latency | <75ms |
| mAP@0.5 | 97.1% |
| Model Size | 3.2MB (INT8) |
| Metric | Value |
|---|---|
| FPS (640px) | 89 |
| Throughput | 3× real-time |
| mAP@0.5 | 96.9% |
┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
│ Video Reader │────▶│ Frame Queue │────▶│ YOLO Inference │
│ (Thread 1) │ │ (Buffer=5) │ │ (Main Thread) │
└─────────────────┘ └─────────────────┘ └─────────────────┘
- Reader thread decodes frames in background
- Queue buffers 5 frames for timing jitter
- Inference runs at max throughput
git clone https://github.com/Sasaank79/Driving-Narrator-Legacy-Edge.git
cd Driving-Narrator-Legacy-Edge
pip install -r requirements.txt
# Fast mode (416px)
python scripts/deploy_720p.py --video your_video.mp4 --conf 0.25
# Precision mode (640px)
python scripts/deploy_720p.py --video your_video.mp4 --imgsz 640src/
├── detector.py # YOLO wrapper (PyTorch/ONNX/OpenVINO)
└── utils.py # Threading, FPS counter
scripts/
├── deploy_720p.py # Main inference script
└── benchmark.py # Throughput testing
notebooks/
└── GPU_Benchmark_T4.ipynb # Colab GPU benchmarking
docker build -t driving-narrator .
docker run -v /path/to/video.mp4:/app/input.mp4 driving-narratorpython -m pytest tests/ -vAGPL-3.0. Uses the LISA Traffic Sign Dataset.