Abstract
We developed an EEG-based probabilistic model, which effectively predicts drowsiness levels of thirty-two
subjects involved in a moving base driving simulator experiment. A hierarchical Gaussian mixture model (hGMM)
with two mixture components at the lower hierarchical level is used. Each mixture models data density distribution
of one of the two drowsiness cornerstones/classes represented by 4-second long EEG segments with low and high
drowsiness levels. We transfer spectral contents of each EEG segment into a compact form of autoregressive model coefficients.
The Karolinska drowsiness scoring method is used to initially label data belonging to individual classes. We demonstrate good
agreement between Karolinska drowsiness scores and the predicted drowsiness, when the hGMM is applied to continuously
monitor drowsiness over the time-course of driving sessions. The computations associated with the approach are fast enough to
build up a practical real-time drowsiness monitoring system.
Go back