Assessment of Driver Mental Fatigue Using Facial Landmarks

Driver fatigue is one of the causal factors for traffic accidents. Actions of facial units convey various information from our brains. This paper proposed a comprehensive approach to explore whether pattern of sequences of the driver's facial landmarks changes from the alert start to the fatigue state. A driving-simulator-based experiment was designed and conducted. Videos of 21 participants' faces were recorded during the experiment, together with subjective and others' assessment of the level of alertness. Sequences of eye aspect ratio (EAR) and mouth aspect ratio (MAR) were calculated based on facial landmarks. Totally 21 feature candidates including Percent eye-closure over a fixed time window (PERCLOS), blink rate, statistics of blink duration, closing speed, reopening speed and number of yawns were extracted. A mental fatigue assessment model is proposed after feature selection. Four machine learning algorithms were used to build the assessment model of fatigue. The best performance was achieved by logistic regression, with cross-validation accuracies of 83.7% and 85.4%. This study may contribute to the development of driver fatigue monitoring system for automotive ergonomics.