Real-time object detection on embedded Linux — YOLOv8-nano, TensorFlow Lite INT8, and OpenCV on ARM Cortex-A72.
Benchmarks four edge model architectures and ships a C++ inference runner for Raspberry Pi 4 deployment.
YOLOv8-nano achieves 22 FPS within a 3.5 W power budget — 47% faster than MobileNet baseline.
| Metric | YOLOv8-nano | MobileNet baseline |
|---|---|---|
| FPS (Pi 4) | 22 | 14 |
| Power | 3.5 W | ~4.2 W |
| Model size (INT8) | 7 MB | 12 MB |
| mAP loss vs FP32 | < 2% | — |
Edge devices can't run cloud-scale vision models. EdgeVision answers: which compact detector gives the best FPS per watt on ARM?
Production choice: YOLOv8-nano + INT8 quantisation — documented with reproducible benchmarks and a deployment guide.
Camera (V4L2) → OpenCV preprocess → TFLite INT8 YOLOv8-nano → NMS → Display / MQTT
flowchart LR
A[Camera / V4L2] --> B[OpenCV Preprocess]
B --> C[TFLite INT8 YOLOv8-nano]
C --> D[NMS + Labels]
D --> E[Display / MQTT]
edgevision/
├── python/
│ ├── benchmark.py # Compare 4 model architectures
│ └── quantize_model.py # INT8 quantisation workflow
├── cpp/
│ ├── src/inference_main.cpp
│ └── include/edgevision/
├── docs/
│ ├── DEPLOYMENT.md # Raspberry Pi setup
│ └── assets/benchmark.png
└── results/ # Generated after benchmark run
git clone https://github.com/pranav-singh-rathore/edgevision.git
cd edgevision
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python python/benchmark.py --output results
python python/quantize_model.py --output results/quantization.jsoncd cpp && mkdir build && cd build
cmake .. && make -j
./edgevision_inferCross-compile for Pi 4 with aarch64-linux-gnu-g++ and link TensorFlow Lite.
| Model | Size | FPS | Notes |
|---|---|---|---|
| MobileNetV3-SSD | 12.4 MB | 14 | Baseline |
| EfficientDet-Lite0 | 9.8 MB | 16.5 | Good accuracy |
| YOLOv8-nano | 6.2 MB | 22 | Production choice |
| NanoDet-Plus | 4.1 MB | 19.5 | Smallest |
Run python python/benchmark.py to regenerate results/BENCHMARK.md.
Full Raspberry Pi 4 guide: docs/DEPLOYMENT.md
Includes YOLOv8 → TFLite INT8 export, cross-compilation, and power profiling.
Pranav Singh Rathore · LinkedIn · Portfolio
MIT