SJSU AI & ML Club — Nuvoton Team 2

Edge AI People Counting on the Nuvoton M55M1 EVB

Deploy a real-time headcount system using a top-down camera and a model accelerated by the Ethos-U55 NPU. All inference runs locally on-device — no cloud or external storage.

Goals

• ≥95% headcount accuracy
• ≥15 FPS inference
• Zero-Cloud: all processing on-device, no data transmitted externally

---

Repo Structure

Nuvoton-Team-2/
├── data/         # dataset structure, metadata, and data instructions
├── ml/           # training, local inference, and conversion scripts
└── deployment/   # export / Vela / board integration notes and scripts

---

Resources

• ML_YOLO Repo — training, export, and quantization pipeline
• ML_M55M1_SampleCode — sample M55M1 ML application reference
• M55M1 BSP — board support package for lower-level firmware integration
• NuEdgeWise
• bdanko/overhead-person-detection — overhead/top-down detection dataset used for initial training tests

---

Quick Start

1. Create and activate virtual environment

python3.11 -m venv .venv
source .venv/bin/activate

2. Install dependencies

pip install --upgrade pip setuptools
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1
pip install ultralytics opencv-python matplotlib pyyaml datasets pillow
pip install numpy==1.26.4 py-cpuinfo seaborn

3. Clone required repositories

mkdir -p repos
cd repos
git clone https://github.com/OpenNuvoton/ML_YOLO.git
git clone https://github.com/OpenNuvoton/ML_M55M1_SampleCode.git
cd ..

4. Download baseline YOLOv8 weights

python -c "from ultralytics import YOLO; YOLO('yolov8n.pt')"

---

See also

• data/README.md for dataset format and labeling expectations
• ml/README.md for dataset conversion, training, and local inference
• deployment/README.md for export / deployment notes

---

Current Status

• Local YOLOv8 ReLU6 training pipeline is working
• Initial training on overhead-person dataset at 192×192 completed successfully
• Local inference demo supports image/video testing with baseline vs custom checkpoints