Temporal facial features based fatigue detection system for low-cost devices

Driver fatigue has long been recognized as a major cause of severe traffic accidents. However, achieving precise real-time fatigue detection for in-cabin drivers using low-cost devices remains a significant challenge. To tackle this issue, we propose a fatigue detection system for drivers based on temporal facial features. In the initial phase, we integrate a high-precision landmark model trained using a teacher-student distillation model with the You Only Look Once 5 face (YOLO5face) model to rapidly and accurately extract facial landmarks. Subsequently, a series of specific criteria are defined for classifying sequential parameters related to eye movement, mouth activity, and head posture. Extensive validation on both self-built dataset and four publicly available datasets demonstrates that both eye and head poses can be accurately detected, achieving an impressive accuracy rate of 97.1% in yawning detection. Furthermore, the system meets real-time detection requirements, operating at inference speed of 8–20 ms on a standard CPU. The source code is available at: https://github.com/Benpowder/Temporal-Facial-Features-Based-Fatigue-Detection-System .