malansi/EMG-Gesture-Recognition

EMG Gesture Recognition — Myo Armband

Real-time hand gesture classification from EMG signals using CNN + LSTM. Built for prosthetic arm control and drone navigation.

Results (Real-Time on Myo Armband)

Gesture	Accuracy
rest	100%
wave_in	100%
fist	98%
pinch	97%
open_hand	92%
wave_out	85%
Average	95%

Model Architecture

Input: 150 samples × 8 EMG channels (750ms @ 200Hz) ↓ CNN Block: Conv1D(64) → BatchNorm → ReLU Conv1D(128) → BatchNorm → ReLU → MaxPool Conv1D(256) → BatchNorm → ReLU → MaxPool ↓ Bidirectional LSTM (128 units × 2 layers) ↓ Dense(128) → Dense(6) ↓ Output: 6 gesture classes

Hardware

• Sensor: Myo Armband — 8 EMG channels, 200Hz
• Platform: Apple Silicon (MPS) / CPU
• Latency: < 100ms end-to-end

Dataset

• 4 recording sessions
• 6 gestures × 6 rounds each
• ~250,000 EMG samples total
• Per-subject global normalization

Key Challenges Solved

1. Data Leakage — block-level train/test split (no overlapping windows)
2. Normalization Mismatch — fixed global stats saved and reused at inference
3. Class Imbalance — weighted CrossEntropy loss
4. Real-time Stability — hysteresis voting system for actuator control

Files

File	Description
models/best_model_v4.pt	Trained PyTorch model weights
models/norm_mean.npy	Normalization mean (required at inference)
models/norm_std.npy	Normalization std (required at inference)
results/confusion_matrix_v4.png	Confusion matrix
results/training_curves_v4.png	Training loss & accuracy curves
code/train.py	Full training pipeline
code/realtime.py	Real-time inference with Myo
code/guided_test.py	Guided accuracy evaluation

Demo

Watch the model running in real-time on the Myo Armband:

https://youtube.com/shorts/uGS7Rv67E7w

Usage

import torch
import numpy as np

# Load model
model = EMG_CNN_LSTM(n_channels=8, n_classes=6)
model.load_state_dict(torch.load('models/best_model_v4.pt'))
model.eval()

# Load normalization stats
norm_mean = np.load('models/norm_mean.npy')
norm_std  = np.load('models/norm_std.npy')

# Normalize input (150 samples × 8 channels)
window = (window - norm_mean) / norm_std

# Predict
x = torch.tensor(window.T.copy(), dtype=torch.float32).unsqueeze(0)
with torch.no_grad():
    probs = torch.softmax(model(x), dim=1)
    pred  = probs.argmax().item()

Author

Mohammed Alansi AI & Robotics Research — EMG-Based Prosthetic Control