AlertDrive

A comparative study of machine learning approaches for driver drowsiness detection using the NTHU-DDD dataset.

Overview

Drowsy driving is one of the leading causes of road fatalities globally. AlertDrive benchmarks three model families for real-time, vision-based drowsiness detection. We implement and evaluate a classical SVM baseline using HOG features, a standalone MobileNetV2 convolutional neural network fine-tuned on individual face frames, and a hybrid CNN-LSTM architecture that combines MobileNetV2 spatial feature extraction with temporal sequence modeling via LSTM. All models are trained and evaluated on the NTHU-DDD dataset under two protocols — subject-dependent and subject-independent — to assess both in-distribution performance and generalization to unseen drivers. Results show that the CNN-LSTM achieves the best generalization with a test AUC of 0.791 in the subject-independent setting, while all models perform substantially better when test subjects overlap with training data.

AlertDrive benchmarks three model families for real-time, vision-based drowsiness detection:

Model	Type	Temporal Modeling
SVM	Classical ML (HOG features)	No
MobileNetV2	Pretrained CNN	No
CNN-LSTM	Spatiotemporal Deep Learning	Yes

Each model is evaluated under two protocols — subject-dependent and subject-independent — to measure both raw performance and generalization to unseen drivers.

---

Table of Contents

• Overview
• Report
• Project Structure
• Dataset
• Setup
• Training
  • SVM
  • MobileNetV2
  • CNN-LSTM
• Results
  • Subject-Independent
  • Subject-Dependent
• Pipeline
• Ethical Considerations
• References
• Authors

---

Report

The full IEEE-format paper is available on Overleaf:
AlertDrive: A Comparative Study of ML Approaches for Driver Drowsiness Detection

---

Project Structure

AlertDrive/
├── models/
│   ├── train_svm.py              # HOG + SGDClassifier pipeline
│   ├── train_mobilenet.py        # MobileNetV2 two-phase fine-tuning
│   └── train_cnnlstm.py          # CNN-LSTM spatiotemporal model
├── splits/
│   ├── train.csv                 # Training split (47,436 samples)
│   ├── val.csv                   # Validation split (8,006 samples)
│   └── test.csv                  # Test split (11,079 samples)
├── reports/
│   └── images/                   # Confusion matrices, ROC curves, training plots
├── AlertDrive_IEEE_Report.tex    # LaTeX source for IEEE paper
└── README.md

---

Dataset

NTHU Driver Drowsiness Detection (NTHU-DDD)

• 66,521 preprocessed face images from 4 subjects under varying eyewear conditions
• Binary labels: drowsy (74.1%) / not drowsy (25.9%)
• 27 unique subject-eyewear-action video groups

Split	Samples
Train	47,436
Validation	8,006
Test	11,079

The dataset is not included in this repository. Request access from the NTHU-DDD authors.
Place preprocessed images under processed/drowsy/ and processed/notdrowsy/, then run the split script.

---

Setup

git clone https://github.com/NafisaTasnimR/AlertDrive.git
cd AlertDrive
pip install torch torchvision scikit-learn opencv-python pandas matplotlib pillow

Training was run on Google Colab with GPU acceleration. Models are saved to Google Drive at AlertDrive/models/.

---

Training

SVM (HOG + SGDClassifier)

python models/train_svm.py

• Extracts 7,056-dim HOG features (128x128 window, 16x16 blocks, 9 bins)
• SGDClassifier with modified_huber loss, L2 alpha=0.001
• Balanced class weights, early stopping (patience=5), up to 30 epochs

MobileNetV2

python models/train_mobilenet.py

Two-phase training:

• Phase 1 (5 epochs): frozen backbone, train classifier head only — lr=1e-3
• Phase 2 (10 epochs): full fine-tuning — lr=1e-4

CNN-LSTM

python models/train_cnnlstm.py

• MobileNetV2 feature extractor fed into a 2-layer LSTM (256 hidden units)
• Sequence length: 16 frames, stride: 8 frames (50% overlap)
• Batch size: 16 sequences, gradient clipping enabled

---

Results

Subject-Independent (generalization to unseen drivers)

Model	Val Acc	Val AUC	Test Acc	Test AUC
CNN+LSTM	69.53%	0.727	64.91%	0.791
MobileNetV2	77.18%	0.848	57.71%	0.608
SVM	55.28%	0.575	38.34%	0.350

Subject-Dependent (test subjects seen during training)

Model	Val Acc	Val AUC	Test Acc	Test AUC
MobileNetV2	94.20%	0.989	81.42%	0.912
CNN+LSTM	93.80%	0.957	79.88%	0.888
SVM	90.88%	0.906	82.57%	0.824

Key finding: CNN-LSTM has the smallest generalization gap — only 4.6% validation-to-test accuracy drop vs. 19.5% for MobileNetV2 — confirming that temporal modeling captures more generalizable drowsiness signatures.

---

Pipeline

Raw Video Frames (NTHU-DDD Dataset)
        |
        v
Face Detection & Cropping
        |
        v
Resize (224x224) + Normalize + Data Augmentation
        |
        +-------------------------+
        |                         |
        v                         v
Subject-Dependent Split   Subject-Independent Split
        |                         |
        +-------------------------+
        |
        +---------------+---------------+
        |               |               |
        v               v               v
      SVM          MobileNetV2      CNN-LSTM
  (HOG + SGD)   (Frame-Based CNN)  (CNN + LSTM)
        |               |               |
        +---------------+---------------+
        |
        v
Evaluation: Accuracy | AUC | Confusion Matrix | ROC Curve

---

Ethical Considerations

• Privacy: All facial video should be processed on-device in any real deployment and never transmitted externally.
• Bias: The NTHU-DDD dataset covers a limited subject pool and may not generalize across age, gender, or skin tone diversity.
• Safety: This system is a supplementary safety tool, not a replacement for driver judgment. False negatives carry real safety risk and must be handled with appropriate thresholding and fail-safes.

---

References

• Tüfekci et al., Driver Drowsiness Detection Using Deep Learning Methods, Sensors 2022
• Lyu et al., LSTM-Based Drowsiness Detection, IEEE ROBIO 2018
• Shen et al., Drowsiness Detection Based on Multi-Stream CNN, IEEE ICPR 2020
• Abtahi et al., YawDD: A Yawning Detection Dataset, ACM MMSys 2014

---

Authors

Department of Computer Science and Engineering
Islamic University of Technology, Gazipur, Bangladesh

Name	Email
Sanjana Afreen	sanjanaafreen@iut-dhaka.edu
Adrita Ahsan	adritaahsan@iut-dhaka.edu
Nishat Tasneem Oishee	nishatoishee@iut-dhaka.edu
Nafisa Tasnim	nafisatasnim22@iut-dhaka.edu
Jarin Subha Sneha	jarinsneha@iut-dhaka.edu