DrowSAFE: Real-Time Driver Drowsiness Detection System

The Problem

Drowsy driving is one of the most dangerous and underreported causes of fatal road accidents. Commercial driver monitoring systems — like those in Mercedes, BMW, or Tesla — cost thousands of euros and are locked into specific vehicles. Fleet-grade standalone systems require cloud connectivity and recurring fees.

DrowSAFE is the open-source answer: a fully self-contained drowsiness detection system that anyone can build and install in any vehicle.

How It Works

DrowSAFE runs a continuous detection pipeline at ~30 fps:

The entire pipeline runs on the Pi 5's quad-core Cortex-A76 CPU. No GPU, no AI accelerator, no internet connection required.

The Science: EAR, MAR, and PERCLOS

MediaPipe Face Mesh detects 468 facial landmarks per frame. DrowSAFE uses these to compute three drowsiness signals:

Eye Aspect Ratio (EAR) — introduced by Soukupová & Čech (2016) — measures the ratio of vertical to horizontal eye span from 6 landmarks per eye. When eyes are open, EAR is typically 0.28–0.35. During a blink it drops sharply; sustained low EAR indicates drowsiness.

EAR = (|p1–p5| + |p2–p4|) / (2 × |p0–p3|)

Mouth Aspect Ratio (MAR) applies the same geometry to the mouth. A MAR above threshold held for 15+ consecutive frames registers as a confirmed yawn.

PERCLOS (Percentage of Eye Closure) is the US NHTSA gold-standard drowsiness metric — the fraction of frames where eyes are more than 80% closed over a rolling 60-second window. PERCLOS > 15% reliably indicates impaired alertness.

Head Pose is estimated in real time via OpenCV's solvePnP using 6 canonical 3D facial landmarks. A forward pitch above threshold — the characteristic microsleep nod — contributes to the fatigue score.

These four signals are combined into a single composite fatigue score (0–100):

PERCLOS : 40%

Instantaneous EAR : 25%

Head pose (nod) : 20%

Yawn frequency (MAR) : 15%

Alert State Machine

The fatigue score drives a three-level state machine with hysteresis — the score must drop well below a threshold before the alert level downgrades. This prevents flickering between states on noisy frames

# Hysteresis prevents flickering — score must drop below these to downgrade
WARNING_HYSTERESIS  = 30   # Drop below 30 to return to ALERT 
CRITICAL_HYSTERESIS = 55   # Drop below 55 to return to WARNING

Here's what each alert level looks like on the display:

Live Detection — Real Results

The following screenshots are taken directly from the running system on the Raspberry Pi 5, with the Camera Module 3 NoIR active and MediaPipe landmarks overlaid.

WARNING state — score 43, EAR 0.188, MAR 1.677, head pitch -27.3°:

CRITICAL state — score 81, EAR 0.029, eyes fully closed:

Both screenshots confirm the system running at a stable 30.0 fps on the Pi 5 with the full MediaPipe pipeline active. The EAR value of 0.029 in the CRITICAL screenshot shows near-complete eye closure correctly triggering the alarm.

The Dashboard

The 7" Raspberry Pi Touch Display runs a fullscreen Pygame dashboard readable at a glance.

Left panel (60%): Live camera feed with MediaPipe landmark overlay (toggled by SHOW_LANDMARKS in config).

Right panel (40%): Fatigue score (large, colour-coded), score progress bar, live EAR / MAR / head pitch values, alert level badge.

Bottom banner: Full-width strip showing alert message — changes colour and text with alert level.

Why Edge AI?

Most commercial driver monitoring systems either require an expensive vehicle or send video to a cloud server. DrowSAFE is different:

• No internet required — works anywhere, including tunnels and rural areas
• Zero latency — no network round trips, detection is instantaneous
• Complete privacy — the driver's face never leaves the device
• One-time cost — no subscription, no API fees

MediaPipe's Face Mesh model is optimised for ARM and runs at a stable 30 fps on the Pi 5 without any GPU or AI accelerator.

Hardware Assembly

Step 1 — Flash the OS

Flash Raspberry Pi OS (Legacy, 64-bit) — Debian 12 Bookworm — using Raspberry Pi Imager.

⚠️ Important: As of April 2026, the default Pi OS image installs Debian 13 Trixie with Python 3.13. MediaPipe is not yet available for Python 3.13. Select "Raspberry Pi OS (Legacy, 64-bit)" explicitly under "Raspberry Pi OS (other)" in the Imager.

⚠️ Important: As of April 2026, the default Pi OS image installs Debian 13 Trixie with Python 3.13. MediaPipe is not yet available for Python 3.13. Select "Raspberry Pi OS (Legacy, 64-bit)" explicitly under "Raspberry Pi OS (other)" in the Imager.

Enable SSH and set your credentials in the Imager customisation screen before flashing.

Step 2 — Connect the Display

Connect the 7" Touch Display to the Pi 5 via the DSI ribbon cable. The display is powered directly from the Pi's USB-C port.

Step 3 — Connect the Camera

The Pi 5 uses a mini CSI connector but the Camera Module 3 ships with a standard 15-pin FPC cable. Use a Standard-to-Mini adapter cable (300mm recommended).

Connect to the Pi 5's CAM/DISP 0 port.

Step 4 — Wire the Buzzer

Connect an active piezo buzzer to GPIO 18:

Software Setup

Step 1 — Clone and install

git clone https://github.com/YOUR_USERNAME/drowsafe.git
cd drowsafe
bash scripts/setup_drowsafe.sh

Expected verification output:

✓  OpenCV               4.11.0
✓  MediaPipe            0.10.18
✓  ONNX Runtime         1.24.4
✓  Pygame               2.6.1
✓  NumPy                1.26.4
All packages OK!

Step 2 — Test without camera (simulation mode)

python src/main.py --simulate

Runs the full dashboard with a synthetic 90-second fatigue cycle — cycles automatically through all three alert levels without any camera or MediaPipe.

Step 3 — Launch live

bash scripts/run.sh

Activates the venv, sets DISPLAY=:0 for SSH sessions, and starts the pipeline.

Calibration

Every face is different. Use the built-in calibration tool to measure your baseline EAR and MAR values:

bash scripts/calibrate.sh

Output

EAR-L    EAR-R   EAR-AVG      MAR    PITCH
  0.312    0.298    0.305    0.312    1.2°   ← mouth closed, eyes open
  0.089    0.076    0.083    0.712   18.4°   ← eyes closed, yawning

Set EAR_THRESHOLD to ~65% of your open-eye EAR. Set MAR_THRESHOLD between your resting and yawn MAR values. All thresholds are in config/config.py.

Configuration

All parameters in a single config/config.py:

CAMERA_FLIP          = True    # Flip 180° if camera mounted upside down

# Detection thresholds — calibrate to your face
EAR_THRESHOLD        = 0.22    # Eye considered closed below this
MAR_THRESHOLD        = 0.75    # Yawn detected above this
HEAD_PITCH_THRESHOLD = 20      # Forward nod angle in degrees

# PERCLOS — NHTSA standard
PERCLOS_WINDOW_SEC   = 60
PERCLOS_THRESHOLD    = 0.15    # >15% closed = drowsy

# Alert levels
WARNING_SCORE        = 40
CRITICAL_SCORE       = 70
WARNING_HYSTERESIS   = 30
CRITICAL_HYSTERESIS  = 55

# GPIO
BUZZER_PIN           = 18