🛡️ Pro-Vision v2.0: Advanced Driver Monitoring System

Overview

Pro-Vision is a production-ready, real-time driver drowsiness detection system using edge computing and advanced computer vision. It classifies driver states into AWAKE, DROWSY, or SLEEPING with multi-level alerts.

Key Features

✅ Real-time Detection - 30+ FPS processing
✅ Advanced Analytics - Blink detection, yawn detection, fatigue scoring
✅ Multi-level Alerts - Info/Warning/Critical with audio & visual
✅ Event Logging - Persistent JSON logs with filtering
✅ Production Architecture - Modular, configurable, scalable
✅ Easy Deployment - Single command startup, Streamlit web UI

---

🚀 Quick Start

Prerequisites

• Python 3.9+
• Webcam (for live mode)
• Linux/Mac/Windows

Installation

# Clone or download the repository
cd pro-vision-v2

# Install dependencies
pip install -r requirements.txt

# Run the dashboard
streamlit run app.py

The dashboard opens at http://localhost:8501

---

📋 System Architecture

┌─────────────────────────────────────────┐
│          Video Input (Webcam)           │
└──────────────────┬──────────────────────┘
                   │
                   ▼
    ┌──────────────────────────────┐
    │  detection_engine.py         │
    │  ├─ FaceDetector            │
    │  ├─ DrowsinessClassifier    │
    │  └─ FrameAnnotator          │
    └──────────────────┬───────────┘
                       │
                       ▼
        ┌──────────────────────────┐
        │  alert_system.py         │
        │  ├─ AlertManager        │
        │  └─ EventLogger         │
        └──────────┬───────────────┘
                   │
                   ▼
    ┌──────────────────────────────┐
    │  app.py (Streamlit)          │
    │  ├─ Live Feed               │
    │  ├─ Real-time Metrics       │
    │  ├─ Charts & Analytics      │
    │  └─ Event Logs              │
    └──────────────────────────────┘

---

🔧 Configuration

Edit config.json to tune the system:

{
  "detection": {
    "face_scale_factor": 1.1,        // Higher = faster but less accurate
    "face_min_neighbors": 6,         // Higher = fewer false positives
    "eye_scale_factor": 1.1,
    "eye_min_neighbors": 5,
    "clahe_clip_limit": 2.0          // Image enhancement strength
  },
  "drowsiness": {
    "history_window": 15,            // Frames to average (delay)
    "blink_threshold_frames": 2,     // Frames to consider a blink
    "eye_closed_ratio": 0.4,         // % eyes closed = drowsy
    "drowsy_threshold": 35,          // Score threshold for DROWSY
    "sleep_threshold": 75            // Score threshold for SLEEPING
  },
  "performance": {
    "skip_frames": 2,                // Process every Nth frame
    "frame_width": 640,
    "frame_height": 480,
    "target_fps": 30
  },
  "alerts": {
    "beep_cooldown": 2.0,            // Seconds between alerts
    "warning_threshold": 35,
    "critical_threshold": 75
  }
}

---

📊 Features Explained

1. Real-time Detection

• Face Detection: Haar Cascade (fast, edge-optimized)
• Eye Detection: Haar Cascade in face ROI
• Smile/Yawn Detection: Optional feature for fatigue analysis

2. Drowsiness Scoring

• Rolling 15-frame window tracks eye closure
• Hysteresis prevents false positives
• Score decays when eyes open (recovery time)
• Thresholds: DROWSY (>35) → SLEEPING (>75)

3. Blink Detection

• Distinguishes between normal blinks and eye closure
• Tracks blink rate (blinks/30s)
• Abnormal patterns trigger warnings

4. Multi-level Alerts

Level	Trigger	Response
INFO	Normal operation	Green border, ✅ message
WARNING	Score > 35	Orange border, ⚠️ message
CRITICAL	Score > 75	Red border, 🚨 alert + beep

5. Event Logging

• Persistent JSON log (vision_events.json)
• Filterable by alert level
• Timestamp, score, and message
• Export for analysis

---

🎛️ Dashboard Controls

Sidebar Settings

• Camera Mode: Live Camera or Demo Mode
• Eye Sensitivity: 3-20 (lower = more sensitive)
• Frame Skip: Process every Nth frame (1-5)
• Audio/Visual Alerts: Toggle on/off
• Event Logs: View, filter, export, clear

Main Display

• Live Video Feed with annotated detections
• Real-time Metrics: Status, score, eyes detected, blink rate
• Drowsiness Chart: Score over time
• Blink Rate Chart: Blink frequency trend

---

📈 Performance Optimization

Frame Skipping

Process every 2nd or 3rd frame (configurable) while maintaining smooth UI:

skip_frames = 2
if frame_count % skip_frames == 0:
    detect_faces_and_eyes()

Benefit: 2x-3x faster with minimal latency (50-100ms delay)

CLAHE Enhancement

Contrast Limited Adaptive Histogram Equalization improves detection in poor lighting:

clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8,8))
enhanced = clahe.apply(gray_frame)

ROI-based Detection

Eyes are searched only in face region:

roi = frame[face_y:face_y+face_h, face_x:face_x+face_w]
eyes = detect_eyes(roi)  # Much faster than full frame

---

🔮 Future Enhancements

Phase 2: Deep Learning Integration

# Optional: Replace Haar Cascades with YOLO/RetinaFace
from ultralytics import YOLO
face_model = YOLO("yolov8n-face.pt")
faces = face_model.predict(frame)

Phase 3: Multi-person Support

• Track multiple drivers simultaneously
• Per-person thresholds and profiles

Phase 4: Advanced Analytics

• Fatigue prediction (predicts drowsiness before critical)
• Driving pattern analysis
• Integration with CAN bus (vehicle data)

Phase 5: Cloud Sync

• Upload alerts to backend
• Mobile notifications
• Fleet monitoring dashboard

---

🐛 Troubleshooting

Camera Not Detected

# Check available cameras
python -c "import cv2; print(cv2.VideoCapture(0).isOpened())"

# Try different camera indices
# In app.py, change: cap = cv2.VideoCapture(0)  →  cap = cv2.VideoCapture(1)

Low FPS

1. Increase skip_frames in settings (3-4)
2. Reduce frame_width/frame_height in config
3. Lower face_scale_factor to 1.05

False Positives (False Alerts)

1. Increase eye_min_neighbors in config (7-8)
2. Adjust eye_closed_ratio (0.5-0.6)
3. Increase history_window (20-25)

Can't Find Eyes

1. Ensure good lighting
2. Remove glasses or adjust angle
3. Lower eye_min_neighbors (3-4)
4. Increase eye_sensitivity slider (15-18)

---

📁 Project Structure

pro-vision-v2/
├── app.py                    # Streamlit dashboard (main entry point)
├── detection_engine.py       # Core detection logic
├── alert_system.py          # Alert & logging system
├── config.json              # Tunable parameters
├── requirements.txt         # Python dependencies
├── vision_events.json       # Event log (auto-generated)
└── README.md               # This file

---

🔐 Data Privacy

• No cloud transmission: All processing local
• No persistent video: Frames not stored
• Logs only: Timestamps and status changes
• Opt-out: Disable logging in config

---

📄 License

Free to use and modify. Built for educational use only.

---

🤝 Contributing

Want to improve Pro-Vision?

1. Better Detection: Experiment with different Haar Cascades
2. New Features: Add blink rate alerts, fatigue prediction
3. Performance: Optimize with GPU acceleration (CUDA)
4. UI: Enhance Streamlit dashboard design

---

📞 Support

• Issues: Check troubleshooting section
• Enhancement: Modify config.json for your use case
• Integration: Adapt detection_engine.py for custom pipelines

---

🚀 Deployment

Local Development

streamlit run app.py

Docker (Optional)

FROM python:3.9-slim
WORKDIR /app
COPY . .
RUN pip install -r requirements.txt
CMD ["streamlit", "run", "app.py"]

Cloud Deployment (Streamlit Cloud)

1. Push repo to GitHub
2. Go to streamlit.io/cloud
3. Connect repo and deploy

---

📊 Performance Benchmarks

Metric	Value	Notes
FPS	25-30	640x480, skip_frames=2
Latency	100-150ms	Face detection + scoring
Memory	~200MB	Python + OpenCV
CPU	30-40%	Single core (i7 @ 2.6GHz)

---

Pro-Vision v2.0 - Built for Production ✅

👥 Contributors

• LABBAALLI Hamza
• ID-BOUBRIK Abdelouahed
• MAAROUF Yassine
• IDDAHA Soumaya

Thanks to all contributors who have helped this project!