Real Time Driver Drowsiness and Distraction Detection System

---

Installation

1. Installing Required Packages

# Clone the repository
git clone https://github.com/username/driver-drowsiness.git
cd driver-drowsiness

# Create a virtual environment (recommended)
python -m venv venv

# Activate the virtual environment
# Windows:
venv\Scripts\activate
# macOS/Linux:
source venv/bin/activate

# Install required packages
pip install -r requirements.txt

2. Downloading the ETH-XGaze Model

The ETH-XGaze model is not included in the repository due to its large size. Follow these steps:

Option A: Official Model (Recommended)

1. Visit the ETH-XGaze official page
2. Download the model file (epoch_24_ckpt.pth.tar)
3. Place it in the models/ directory:

   mkdir -p models
   # Copy the downloaded file to the models directory
   cp /download/path/epoch_24_ckpt.pth.tar models/eth_xgaze_model.pth

Option B: ONNX Format (Fast Inference)

# Run the model conversion script
python scripts/convert_ethxgaze_model.py models/eth_xgaze_model.pth models/eth_xgaze_model.onnx --export_onnx

---

Usage

1. Basic Usage

# Start the main application
python run.py

Usage Steps:

1. Make sure your camera is connected
2. Click the "Start" button
3. Keep your face in the camera's field of view
4. Monitor real-time metrics
5. End the analysis with "Stop"

2. Video Analysis

# Video file analysis
python examples/analyze_video.py --input video.mp4 --output results/

# Batch video analysis
python scripts/batch_analyze.py --input_dir videos/ --output_dir results/

You can use the feature either command line or application.

3. Gaze Zone Detection

# Gaze zone test
python tests/test_gaze_zone_detector.py --camera 0 --show_zones

# Gaze zone calibration
python scripts/calibrate_gaze_zones.py

EU regulation C(2023)4523 compliant zones:

Zone ID	Zone Name	Area	Criticality
0	Road Center	Area 2	🔴 Critical
1	Driving Instruments	Area 2	🟡 Driving Related
2	Infotainment	Area 1	🟠 Non-Driving
3	Left Side	Area 2	🟡 Driving Related
4	Right Side	Area 2	🟡 Driving Related
5	Rear Mirror	Area 2	🔴 Critical

Classification Performance Metrics

• Overall Accuracy: 93.23%
• Balanced Accuracy: 83.70%
• Macro-averaged Precision: 91.96%
• Macro-averaged Recall: 83.70%
• Macro-averaged F1-Score: 85.84%

Statistical Reliability Measures

• Cohen's Kappa: 0.9110 (Almost Perfect)
• Matthews Correlation Coefficient: 0.8549
• Krippendorff's Alpha: 0.9110

---

Metrics

EAR (Eye Aspect Ratio)

• Formula: EAR = (|p2-p6| + |p3-p5|) / (2 * |p1-p4|)
• Threshold Value: 0.21 (eyes considered closed below this)
• Usage: Instant eye blink and closure detection

MAR (Mouth Aspect Ratio)

• Formula: MAR = |p14-p18| / |p12-p16|
• Threshold Value: 0.65 (mouth considered open above this)
• Usage: Yawning detection

PERCLOS

• Definition: Percentage of eye closure over a 60-second period
• Warning Threshold: 15%
• Critical Threshold: 20%

KSS (Karolinska Sleepiness Scale)

Score	Status	Description
1-3	🟢 Normal	Fully awake state
4-5	🟡 Mild	Mild signs of fatigue
6-7	🟠 Warning	Beginning of attention deficit
8-9	🔴 Critical	Immediate intervention required

---

Scientific References

• ETH-XGaze: Gaze estimation model (Paper)
• MediaPipe: Face mesh detection (Documentation)