🧠 DriveSense AI — Adaptive Driver Fatigue & Attention Monitor

"Every year, 100,000 people die from drowsy driving. DriveSense AI predicts fatigue before it kills."

🎯 Judge's Pitch

Existing driver monitoring systems detect drowsiness after your eyes close. DriveSense AI uses a multi-factor fatigue fusion model that combines PERCLOS, blink anomalies, head pose drift, yawn frequency, and circadian rhythm data to predict dangerous fatigue levels 10 minutes before they happen — giving drivers time to save their own lives. It's the difference between a near-miss and a funeral.

💡 Inspiration

We started with a question: Why do driver monitoring systems only react instead of predict?

openpilot has a driver-facing camera that checks if you're looking at the road. But alertness isn't binary. Fatigue is a gradual physiological process — it shows up in your blink patterns 20 minutes before you feel sleepy. Your head starts micro-drifting. Your yawns cluster together.

We built DriveSense AI to catch what humans can't feel happening to themselves.

🚗 What It Does

DriveSense AI is a real-time driver fatigue monitoring and prediction dashboard that:

🔍 Tracks 6+ biometric signals — PERCLOS (eye closure percentage), blink rate, head yaw/pitch, yawn frequency, gaze direction, mouth openness
🧠 Runs a multi-factor AI fatigue model — Weighted fusion of all signals through an exponential moving average pipeline
📈 Predicts fatigue 10 minutes ahead — Linear regression on trend data projects when you'll cross danger thresholds
🏆 Computes a real-time Safety Score (0–100, A–F grade) — With component breakdowns showing exactly what's degrading
🚨 3-tier alert system — Info nudges → Warning chimes → Critical "PULL OVER" full-screen overlay with audio
🌙 Circadian rhythm modeling — Knows that 3 AM and 2 PM are biologically dangerous times
☕ Coffee break simulation — Models temporary alertness recovery after a rest stop
🗺️ Gaze attention heatmap — Visualizes where the driver is looking over time

4 Realistic Driving Scenarios

Scenario	Description
🌙 Night Highway	Monotonous highway at 11 PM — fatigue builds fast
🌅 Morning Commute	Rush hour at 7:30 AM — stop-and-go keeps you alert
😴 Post-Lunch Dip	Afternoon drive at 2 PM — the circadian dip strikes
🚛 Long Haul Trucker	6-hour interstate haul — relentless fatigue accumulation

🏗️ How We Built It

Architecture

┌──────────────────────────────────────────────────┐
│                  Dashboard UI                     │
│  (index.html + glassmorphism CSS)                │
│                                                   │
│  ┌──────────┐ ┌───────────┐ ┌────────────────┐  │
│  │ Driver   │ │ Chart.js  │ │ Alert Feed     │  │
│  │ Face     │ │ Timeline  │ │ + Toast        │  │
│  │ Canvas   │ │ + Gauge   │ │ Notifications  │  │
│  └────┬─────┘ └─────┬─────┘ └──────┬─────────┘  │
│       └──────────────┼──────────────┘             │
│               ┌──────▼──────┐                     │
│               │  app.js     │ ← Main Controller   │
│               │  (60fps)    │                     │
│               └──┬───┬───┬──┘                     │
│          ┌───────┘   │   └────────┐               │
│     ┌────▼───┐ ┌─────▼────┐ ┌────▼────┐         │
│     │AI Eng. │ │Simulation│ │Scoring  │         │
│     │Fatigue │ │Scenario  │ │Safety   │         │
│     │Model   │ │Generator │ │Score    │         │
│     └────────┘ └──────────┘ └─────────┘         │
│                                                   │
│     config.js    utils.js    alerts.js            │
└──────────────────────────────────────────────────┘

Tech Stack

Technology	Why We Chose It
Vanilla JS	Zero build step, instant demo, max portability. Opens from a file.
Canvas API	Real-time face visualization at 60fps without DOM overhead
Chart.js 4	Beautiful animated charts with 5 lines of config
CSS Glassmorphism	Premium dark theme that wows judges instantly
Web Audio API	Alert sounds without loading audio files
requestAnimationFrame	Smooth, battery-efficient render loop

Key Technical Decisions

Multi-factor fusion over binary detection — Real fatigue is a spectrum. Our weighted model captures gradients that threshold-based systems miss.
Predictive regression — Instead of just reporting current state, we project the fatigue trajectory forward using linear regression on rolling history.
Circadian rhythm modeling — Physiological alertness follows a 24-hour curve. Our model factors in time-of-day to correlate with biological fatigue patterns.
Simulation with realistic noise — Gaussian noise, smooth sine-based drift, and Box-Muller transforms create data that feels organic, not robotic.

🎮 Demo

Running Locally

# No build step needed! Just serve static files:
npx -y serve .

# Then open http://localhost:3000

Controls

Scenario Selector — Switch between 4 driving scenarios
Speed Controls — 1×, 2×, 5×, 10× simulation speed
Pause/Resume — Freeze the simulation to inspect state
Reset — Start the scenario fresh

What to Watch For

Start with Night Highway at 5× speed
Watch the fatigue line climb on the timeline chart
Notice the face canvas — eyes start closing, head drifts
Info alerts appear first, then warnings
At ~80% fatigue: PULL OVER overlay with urgent audio
The Safety Score drops from A (green) to F (red) in real-time
Check the gaze heatmap — attention scatter increases with fatigue

🧩 Challenges

Making simulated data feel real — Purely random data looks fake. We used Gaussian noise, Box-Muller transforms, and multi-frequency sine wave composition to create organic physiological patterns.
Alert fatigue — Ironic for a fatigue system! We implemented cooldown timers per alert level to prevent notification spam while still catching every real danger.
Predictive accuracy — Linear regression is simple but effective on monotone fatigue curves. More complex scenarios (coffee breaks, conversation) required modeling recovery dynamics.
60fps Canvas rendering — Drawing a face + indicators + eyes every frame while updating Chart.js and managing alerts required careful throttling of expensive operations.

🏆 Accomplishments

Full real-time fatigue prediction — Not just detection, but a 10-minute lookahead
5-component weighted AI model running at 60fps with no jank
Production-quality UI with glassmorphism, micro-animations, and responsive design
Zero dependencies (except Chart.js CDN) — opens from any file server
Realistic simulation engine that models circadian rhythms, microsleeps, coffee breaks, and yawn clustering
Multi-modal alerts — visual toasts, feed history, audio chimes, and emergency overlay

📚 What We Learned

PERCLOS (Percentage of Eye Closure) is the gold standard metric in fatigue research — we implemented it using a rolling 60-second window
Circadian rhythms create two daily valleys of alertness: 2–4 AM and 1–3 PM — both are peak crash windows
Alert design matters as much as detection accuracy — a system that annoys the driver is a system that gets turned off
Exponential Moving Averages are surprisingly effective for real-time signal smoothing with minimal computational cost
Canvas 2D can render surprisingly expressive face visualizations with basic shape primitives

🔮 Future Improvements

Real camera integration — Replace simulation with MediaPipe Face Mesh for actual driver monitoring via webcam
openpilot plugin — Deploy as a companion app that reads openpilot's driver monitoring camera feed
ML model upgrade — Train an LSTM on real fatigue datasets (like DROZY or UTA-RLDD) for higher accuracy
Heart rate variability — Integrate smartwatch HRV data for cardiovascular fatigue markers
Fleet management dashboard — Multi-driver monitoring for commercial trucking companies
V2X communication — Share fatigue state with nearby vehicles and infrastructure
Personalized baselines — Learn individual driver patterns to customize thresholds

📁 Project Structure

car/
├── index.html              # Main dashboard (single page)
├── README.md               # This file
├── config/
│   └── config.js           # All tunable constants
├── css/
│   └── style.css           # Premium dark glassmorphism theme
└── js/
    ├── utils.js            # Helpers (EMA, noise, color interpolation)
    ├── ai-engine.js        # Multi-factor fatigue AI model
    ├── simulation.js       # Driving scenario simulator
    ├── scoring.js          # Safety score engine (0–100, A–F)
    ├── alerts.js           # 3-tier alert manager + Web Audio
    ├── charts.js           # Chart.js timeline + canvas gauge + heatmap
    └── app.js              # Main controller + face renderer