🧠 Driver Drowsiness Detection using YOLOv5 and LSTM

Real-time drowsiness detection system that uses YOLOv5 for face/eye detection and LSTM for fatigue classification based on temporal patterns. This AI-based solution aims to enhance driver safety by detecting early signs of drowsiness.

---

🚀 Features

• 🎥 Real-time webcam integration
• 👁️ Eye and face detection using YOLOv5
• ⏱️ Temporal sequence modeling with LSTM
• 🧠 Custom labeled dataset (awake/drowsy)
• 📊 Performance monitoring in Jupyter
• 🔧 Modular design for easy training and customization

---

🧠 How It Works

1. YOLOv5 model detects face and eyes from webcam feed.

2. Eye regions are cropped and passed to a trained LSTM classifier.

3. LSTM analyzes a sequence of frames to determine the state:

   • 🟢 Awake
   • 🔴 Drowsy

The system reacts instantly and overlays the prediction label with bounding boxes on the live feed.

---

📈 Model Evaluation Results

🔹 Label Distribution Heatmap

Labels are well-distributed in terms of position and size. Balanced annotation increases detection accuracy.

---

🔹 Precision vs Confidence

Awake class achieves high precision at lower confidence thresholds, while drowsy class stabilizes later.

---

🔹 Precision-Recall Curve

• Awake: 0.948
• Drowsy: 0.861
• mAP@0.5: 0.905

Indicates a reliable and balanced detection model.

---

🔹 Recall vs Confidence

High recall even at low confidence thresholds suggests effective detection capability.

---

🔹 Training Metrics

Loss values decrease steadily and performance metrics increase over epochs, showing good training convergence.

---

🔹 Training Batches

Sample training images with labels. Variations in expression, angle, and lighting enhance model generalization.

---

🔹 Validation Results

Predictions match the ground truth closely, validating model reliability.

---

📁 Project Structure

driver_drownise_detection_Program/
├── data/
├── datasets/
├── labelimg/
├── yolov5/
├── yolov5s.pt
├── train_lstm.py
├── main.py
├── Drowsiness Detection Tutorial.ipynb
├── requirements.txt
└── README.md

---

⚙️ Installation

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

YOLOv5 repo must be cloned under yolov5/:

git clone https://github.com/ultralytics/yolov5.git

---

🦖 Training

YOLOv5

cd yolov5
python train.py --img 640 --batch 16 --epochs 100 --data ../data/data.yaml --weights yolov5s.pt

LSTM

python train_lstm.py --data_dir datasets/ --epochs 50 --batch_size 32

---

📹 Run Real-Time Detection

python main.py

---

📊 Final Thoughts

The project demonstrates that combining object detection with temporal sequence modeling (YOLO + LSTM) can effectively solve real-world problems such as driver fatigue monitoring. Future improvements could include:

• Adding audio alerts
• Deploying on mobile devices with TFLite
• Expanding the dataset with night-time/low-light scenarios

---

📄 License

MIT © 2025 Ali Nebi Er

---

🤝 Acknowledgments

Built with support from OpenAI's GPT-4 and the YOLOv5 community.