Driver drowsiness detection

Real-time driver drowsiness detection using computer vision and an integrated alarm system.

Language

Python

Stars

2

Forks

1

Last push

27 Sep 2026

License

MIT

Problem

Fatigue builds gradually and drivers are poor judges of it. A monitor has to catch prolonged eye closure and head drops without firing on every blink, yawn, conversation or glance at a mirror.

Solution

MediaPipe Face Mesh tracks 478 facial landmarks per frame. Eye and mouth aspect ratios and a solvePnP head pose feed a time-based analyser that turns frames into events: blinks, long blinks, micro-sleeps, yawns, nods and face loss.

Acute signs (continuous closure, sustained head drop) are combined with a capped history score (PERCLOS, yawns, nods, long blinks), so history alone can never trigger the alarm. A hysteresis decision layer then drives interchangeable alarm outputs: a speaker, a Raspberry Pi GPIO buzzer or relay, or an Arduino over USB serial.

Architecture

1. CameraUSB webcam, Pi camera or recorded video, read on a background thread
2. Pre-processingAuto-gamma and CLAHE for uneven lighting
3. Face meshMediaPipe, 478 landmarks
4. MetricsEAR, MAR and head pitch/yaw via solvePnP
5. Temporal analysisBlinks, PERCLOS, yawns, nods and face loss, all time-based
6. Scoring and decisionAcute plus capped history score, with hysteresis and minimum alarm time
7. OutputsAlarm back-ends, live dashboard and CSV session log

Features

• Five-second per-driver calibration of eye and mouth thresholds and neutral head pose
• Frame-rate-independent timing, so results do not depend on camera FPS
• Face loss is treated as 'driver not visible', never as drowsiness
• Eye data is ignored when the head is turned more than 30 degrees
• Hardware abstraction: new alarm outputs implement on() and off()
• Every threshold lives in config.yaml and is validated at start-up
• Camera-free simulated drive and an automated test suite

Results

Default decision thresholds from the project configuration. The automated tests verify that the logic behaves as designed; real-world accuracy has not been measured yet, and the README defines the evaluation protocol for it.

Limitations

• The largest detected face is treated as the driver.
• The Arduino serial protocol has no command acknowledgement yet.
• Dark sunglasses and heavy occlusion make eye measurements unreliable.
• Thresholds and weights still need evaluation across people and conditions.