Drowsiness Detection using CNN & Transfer Learning

This project implements a Drowsiness Detection System using Convolutional Neural Networks (CNNs) and Transfer Learning with MobileNetV2.
It can predict whether a person is drowsy (eyes closed, yawning) or alert (eyes open, not yawning) in real time using webcam feedback.

The model supports two approaches:

1. Single-task classification → Classifies into 4 categories (open, closed, yawn, no_yawn).

---

📂 Dataset

We use the Yawn-Eye Dataset (NEW).
The dataset contains labeled images of:

• Open eyes
• Closed eyes
• Yawning
• Not yawning

Dataset Structure

data/
│── train/
│   ├── Open/
│   ├── Closed/
│   ├── Yawn/
│   └── No_Yawn/
│── test/

---

⚙️ Installation

Clone this repository and install dependencies:

git clone https://github.com/yourusername/drowsiness-detection.git
cd drowsiness-detection
pip install -r requirements.txt

---

🚀 Training the Model

Run the Jupyter Notebook:

jupyter notebook drowsiness-detection.ipynb

The notebook supports:

• Data preprocessing & augmentation
• Transfer learning with MobileNetV2
• Class balancing using weights
• Callbacks (early stopping, learning rate scheduling, checkpointing)
• Evaluation (accuracy, loss curves, confusion matrices)

---

🎥 Real-Time Drowsiness Detection

Once trained, you can run the webcam demo:

python webcam_demo.py

This will:

• Capture live video
• Detect face, predict eye state & yawning
• Overlay results on the video feed

Press q to exit.

---

📊 Results

• Single-task model: Improved validation accuracy.
• Optimized training (with dropout, L2, fine-tuning, augmentation): Improved generalization.

---

🔧 Troubleshooting

• If validation accuracy is much lower than training → increase augmentation, dropout, or reduce fine-tuned layers.
• If webcam feed is laggy → reduce input size from 224x224 to 128x128.
• If dataset is imbalanced → use class_weight during training.

---

📦 Requirements

See requirements.txt.

---

📜 License

MIT License.
Feel free to use and modify this project for research and learning.