Drowsiness Detection System
A Vision Transformer classifies eye state from live camera input at 99.2% test accuracy — 98.4% on people it never saw in training. A closed-eye counter escalates from an on-screen Warning to a Critical alarm on an Arduino module of LEDs and a buzzer; on scripted video, every posed eye closure raised a Warning within a median 0.27 s. Each 720p frame takes 22 ms on a laptop GPU.
- Python
- PyTorch
- OpenCV
- ViT
- Arduino
- Role
- Design & build — personal project
- Year
- 2026
- Code
- View repository on GitHub
