Driver Drowsiness Detection

• Most existing driver drowsiness detectors rely heavily on eye-based features (EAR, blink rate) and completely fail when the driver wears sunglasses or has occluded eyes, which is a very common real-world scenario during daytime driving, leading to dangerous false negatives.
• Our system solves this by combining a classical geometric model (using mouth-based MAR & MOE that work even with sunglasses) with a high-accuracy three-branch CNN (face + left/right eyes).

Dataset

• This project uses the YawDD dataset (Yawning Detection Dataset). The classes are alert and drowsy

Drowsiness Detection Features (Facial Landmarks)

• EAR (Eye Aspect Ratio): Measures eye opening. Lower values indicate closed or blinking eyes.

• MAR (Mouth Aspect Ratio): Measures how widely the mouth is open (yawning).

• Eye Circularity: Detects squinting by measuring how circular the eye contour is.
  Values close to 1 → wide-open circular eye; lower values → squinted/closed.

• MOE (Mouth-over-Eye Ratio): Highly robust drowsiness indicator, especially when eyes are occluded by sunglasses.
  When EAR drops or becomes unreliable (e.g., sunglasses), MAR alone drives MOE high during yawning → strong drowsy signal even with fully covered eyes.

CNN Architecture

• Backbone: A ResNet-18 backbone (pretrained on ImageNet) is used for the main face/scene branch to extract global facial features.
• Eye branches: Two lightweight CNN branches (one per eye) process tighter eye crops to capture local eye patterns under occlusion.
• Fusion & classifier: Features from the three branches are concatenated and passed through fully-connected layers with dropout and a final sigmoid/logit for binary Alert vs Drowsy prediction.

Folder Structure

Driver-Drowsiness-Detection/
├── cnn_resnet18.ipynb              # CNN training 
├── DataPreprocess_FeatureExtract.ipynb  # Dataset extraction & classical feature engineering
├── ensemble_model.ipynb            # Real-time ensemble inference
├── model_evaluation.ipynb          # Evaluation scripts & analysis
├── models/
│   └── drowsiness-97_10.pth        # Trained CNN checkpoint (committed)
│   └── best_drowsiness_model.pkl   # Saved KNN model
│   └── feature_scaler.pkl    
├── README.md                       

Quick Setup & Usage

1. Create a Python environment and install required packages.

pip install torch torchvision opencv-python scikit-learn dlib matplotlib numpy

2. Open the notebooks in Jupyter / VS Code and run the cells in order:

• DataPreprocess_FeatureExtract.ipynb — prepare dataset, extract classical features.
• cnn_resnet18.ipynb — training and architecture notes for the CNN branch.
• model_evaluation.ipynb — evaluate model performance on held-out data.
• ensemble_model.ipynb — demo notebook that runs the real-time ensemble (webcam).

Results Summary

• KNN (classical features): ~72% test accuracy on the processed dataset (features extracted from frames like EAR, MAR, Circularity, MOE).
• CNN (ResNet18-based): ~97% accuracy on visible-eyes data.

Demo Output

Built by: Esha Bidkar, Carol Chopde, Prachi Chavhan, Niharika Hariharan