Research on Non-contact Fatigue Monitoring Method with Unsupervised Respiration Estimation

In scenarios where the facial area is partially obscured, fatigue assessment of drivers is a current challenge. In comparison to complex physiological devices and expensive supervised training, this paper proposes a non-contact fatigue detection method achieved through unsupervised training. Firstly, the photoplethysmography method is employed to obtain the driver's respiratory waveform. Subsequently, a time-frequency fusion network is utilized for respiratory rate monitoring. Finally, an analysis of respiratory changes under fatigue conditions is conducted to achieve the alerting effect. Experimental results on a self-built dataset indicate a high correlation between the proposed respiratory signal acquisition method and real signals, with average R and ICC values exceeding 0.98. In fatigue analysis experiments, it was observed that drivers experienced an average decrease of 3.1 bpm in respiratory values upon transitioning from wakefulness to drowsiness, with a corresponding increase in standard deviation by 1.15 bpm. Based on this criterion and the method proposed in this paper, the F1 value for fatigue warning reached 0.91, demonstrating a certain feasibility and practical value.