A Multi-Modal Driver Fatigue and Distraction Assessment System
We present an estimation of fatigue level within individual operators using voice analysis. One advantage of voice analysis is its utilization of already existing operator communications hardware (2-way radio). From the driver viewpoint it’s an unobtrusive, non-interfering, secondary task. The expected fatigue induced speech changes refer to the voice categories of intensity, rhythm, pause patterns, intonation, speech rate, articulation, and speech quality. Due to inter-individual differences in speech pattern we recorded speaker dependent baselines under alert conditions. Furthermore, sophisticated classification tools (e.g. Support Vector Machine, Multi-Layer Perceptron) were applied to distinguish these different fatigue clusters. To validate the voice analysis predetermined speech samples gained from a driving simulator based sleep deprivation study (N=12; 01.00-08.00 a.m.) are used. Using standard acoustic feature computation procedures we selected 1748 features and fed them into 8 machine learning methods. After each combining the output of each single classifier we yielded a recognition rate of 83.8% in classifying slight from strong fatigue.