Real-Time Driver Microsleep Detection Using Lightweight MobileViT and Haar Cascade

Abstrak

Microsleep is a brief and involuntary loss of awareness that increases the risk of driving accidents. This study proposes a real-time microsleep detection system using Haar Cascade for eye localization and the lightweight MobileViT-XXS model for eye-state classification. The model was trained on a public dataset and achieved a training accuracy of 99.49%, a test accuracy of 98%, and a real-time accuracy of 90–94%. A microsleep event is detected when the eyes remain closed for ≥ 2 seconds. While the method performs well under controlled conditions, real-time testing revealed technical limitations such as sensitivity to lighting variation, non-frontal head pose, and motion, which affect detection stability and represent common robust-vision challenges. Despite these limitations, the system runs efficiently on CPU-only hardware and demonstrates strong potential as a lightweight early-warning system to support driving safety. Future research may explore expanding dataset diversity, improving environmental adaptation, and deploying the system on embedded or mobile platforms to enhance robustness and scalability.