BlinkWatch is a real-time drowsiness detection system that analyzes eye-blink behavior using MediaPipe FaceMesh, engineers blink-based features, classifies alert vs drowsy states with a Random Forest model, and captures temporal patterns using a TensorFlow LSTM network to trigger smart alarm warnings.
Computer Vision: OpenCV-Python 4.10.0.84, MediaPipe FaceMesh 0.10.14
ML / Deep Learning: Scikit-Learn (Random Forest), TensorFlow-CPU 2.16.1 (LSTM via Keras)
Data Processing: NumPy 1.26.4, Pandas 2.2.2, Joblib 1.4.2
Core Utilities: EAR (eye aspect ratio) calculations, Head pose estimation
Environment & Package Management: Python venv, requirements.txt
Frontend (Future Scope): Streamlit (UI only, not used in current pipeline)
Image1
Image2
Image3
Image4
Image5
- Clone your repo
git clone https://github.com/Shivp45/BlinkWatch.git
cd BlinkWatch
- Create & activate virtual environment
python -m venv cv_env
cv_env\Scripts\activate # windows
source cv_env/bin/activate # Mac
- Install dependencies
pip install -r requirements.txt
- Run data collection (if needed)
python data_collection/record_blinks.py
- Run feature extraction (If new feed was added)
python feature_extraction/extract_features.py
- Train ML baseline model (Random Forest)
python models/train_ml.py
- Train LSTM model (uses temporal 30-frame sequences stored in dataset/)
python models/train_lstm.py
- Evaluate model performance
python models/model_eval.py
- Run real-time drowsiness detection using webcam
python inference/realtime_detect.py
Flow Diagram:
Open Webcam
↓
Capture Frames (Alert + Drowsy)
↓
ESC Pressed → Save blink features (features.csv)
↓
Generate sequences (X.npy, y.npy) [temporal dataset creation]
↓
Train ML model (Random Forest → ml_model.pkl)
↓
Train LSTM model (lstm_model.h5)
↓
Evaluate models (model_eval.py)
↓
Run real-time detection (realtime_detect.py)
↓
Overlay status + Trigger alarm if drowsy