EMG Gesture Recognition on Edge

**Developed by the Edge AI Team

Overview

This repository contains a machine learning pipeline for classifying hand gestures using 8-channel forearm EMG signals. The project focuses on taking a model from raw data exploration to optimized hardware deployment.

We trained a Multilayer Perceptron (MLP) on time-domain engineered features and applied Static 8-bit Quantization (INT8) to the final model. The quantized model is exported to ONNX format for efficient, low-latency inference on edge devices like the Raspberry Pi Zero 2 W.

---

Dataset

The dataset used for training the model. You can download it directly here: Download EMG Dataset (.csv files)

---

Repository Contents

• Notebooks: Full pipeline including EDA, preprocessing, feature extraction, model training, Optuna hyperparameter tuning, and ONNX export.
• Deployment: run_inference.py script to benchmark and run the quantized ONNX model locally on a Raspberry Pi.
• Models: Exported FP32 and INT8 .onnx models for edge evaluation.

---

Edge Deployment Guide (Raspberry Pi Zero 2 W)

The following steps were used to deploy and test the model locally on the Raspberry Pi.

1. Clone the Code Repository (GitHub)

First, clone this GitHub repository to your Raspberry Pi to get the most up-to-date deployment scripts (run_inference.py and MLResourceUse.py):

git clone https://github.com/LonghornNeurotech/Edge_EMG_Classification.git
cd Edge_EMG_Classification

2. Download Model and Data Artifacts (Google Drive)

Next, download your heavy model binaries (.onnx files) and the evaluation dataset (.npy files) from Google Drive directly into the folder you just cloned:

pip install gdown --break-system-packages
echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bashrc && source ~/.bashrc

gdown --folder "https://drive.google.com/drive/folders/YOUR_DRIVE_LINK_HERE" -O ./

3. Setup Python Virtual Environment

To avoid externally-managed-environment (PEP 668) restrictions on the latest Raspberry Pi OS, create and activate a virtual environment:

python3 -m venv emg_env
source emg_env/bin/activate

4. Install Dependencies

Install the required ONNX Runtime and NumPy packages inside the virtual environment:

pip install onnxruntime numpy

5. Run Inference

Navigate to the downloaded folder, ensure run_inference.py is present, and execute the benchmark:

cd Edge_EMG_Classification
python run_inference.py

Tip: If you push new code to GitHub and want to instantaneously overwrite your Pi's run_inference.py without dealing with git merges or re-cloning, simply pull the raw file directly:

wget -O run_inference.py https://raw.githubusercontent.com/LonghornNeurotech/Edge_EMG_Classification/main/run_inference.py

---

Benchmark Results (Raspberry Pi Zero 2 W)

Running inference on the test set (4,497 samples) yielded the following performance metrics:

Model Version	Accuracy	Total Time	Avg Latency/Sample
FP32 Baseline (emg_mlp_model.onnx)	90.50%	4.68 seconds	1.04 ms
INT8 Quantized (emg_mlp_model_quantized.onnx)	90.37%	1.94 seconds	0.43 ms

---

Conclusion

Static 8-bit quantization achieved a 2.4x speedup (cutting latency from 1.04ms to 0.43ms) with a negligible accuracy drop of only 0.13%.