Signal Processing - November 2016 - 43

versa (except a few locations, such as "I" and "J"). A compreIn the EEG method, two (or three including the active
hensive review of various EEG processing techniques used to
ground) to more than 100 electrodes are located on the
monitor driver's mental workload, fatigue, and drowsiness is
scalp, each measuring an aggregate of electric voltage fields
provided in [31].
from millions of neighboring neurons. EEG signals can be
Finally, it is noteworthy that any facial activity or eye
expressed in terms of a number of rhythmic activities, which
movement can cause interference in the EEG signals, which
are usually divided into the following frequency bands: Delta
are commonly referred to as artifacts. For example, the blue
(< 4 Hz), Theta (4-8 Hz), Alpha (812 Hz), Beta (12-30 Hz),
highlighted parts in Figure 7 are the artifacts caused by eye
and Gamma (> 30 Hz) [42]. When no major cognitive or
blinks. The common practice in EEG analysis is to remove
motor task is performed, large populations of neurons are
these artifacts to study the underlying brain activity. However,
synchronized and result in steady rhythmic activities. In conin the context of driver monitoring, these artifacts can provide
trast, during cognitive or motor tasks, the synchronization of
side information regarding the facial activities, in particular
these populations usually decreases (in some cases increases),
eye movements. Therefore, one area with great potential for
which results in a decrease (or increase) in the power of corfuture research is to exploit EEG artifacts for extraction of fearesponding oscillatory rhythms.
tures, such as blink duration/frequency, eye closures, and even
The waking EEG is mostly characterized by a desynchroeye gaze direction.
nized pattern, causing rapid, high-frequency waves in the beta
range. When people are awake in a quiet
relaxed state, the increased synchrony of
Data fusion
There is great potential
underlying neural activity in nonstimulatEach of the three measurement categories
for the development
ed brain regions results in patterns of alpha
discussed in the previous sections has its
of new fusion
waves particularly at the posterior part of
own pros and cons, which have to be taken
the brain. In driver monitoring systems,
into account in the design of smart driver
techniques for
one of the most difficult tasks is to detect
monitoring systems. The main limitation
driver monitoring.
the onset of drowsiness, often referred to
of both vehicle-based measures and facial/
as nonrapid eye movement (NREM) sleep
body expressions is that they depend on facstage N1 [43]. However, in EEG signals, this transition can be
tors that tend to manifest themselves at the late stages of fatigue
easily identified by the changes of alpha and theta waves. If
or drowsiness (low arousal) or mental overload (high arousal),
the driver passes this stage and enters the next NREM sleep
whereas physiological measures can provide early indicators
stage (N2) [43], the eye movements will cease, the HR slows
of changes in arousal. However, since current technologies for
down, and body temperature decreases, preparing the body
physiological measurement require direct attachment of sento enter deep sleep. At this stage, background EEG oscillasors to driver's body, these measures might be more intrutions decrease below 5 Hz. Furthermore, these slow oscilsive than other categories. Similarly, some drivers may have
lations will be superimposed by periodic, transient EEG
concerns with having a camera analyzing their facial/body
patterns called sleep spindles and K-complexes. These synexpressions while driving. If the driver's concerns regardchronization changes at different regions of the brain and
ing privacy, ease of use, and nonintrusiveness are taken into
their corresponding effects on the EEG characteristics in
account in the design of new sensing technologies such that
time/frequency/space domains can be used for driver state
drivers embrace them, the combination of these three categomonitoring. Some examples of commonly used methods
ries can provide a multifaceted driver monitoring system that
for extraction of discriminative features from EEG signals
can benefit from the advantages of each of these categories
include: parametric spectral estimation (e.g., autoregressive/
and provide higher accuracy and performance. Furthermore,
moving-average), nonparametric spectral estimation (e.g.,
information provided by intelligent transportation systems
Fourier/wavelet transform), bandpass filtering, and spatial fil[44] as well as information about driver's interactions with
tering (e.g., common spatial patterns, independent component
his surroundings, such as conversation with other passengers,
analysis, surface Laplacian filtering).
can also be utilized. Combining multiple sources of informaFigure 6 provides an example of the changes in an EEG
tion calls for the development of new data fusion techniques.
feature (the ratio between powers in Alpha and Theta bands
For example, [45] uses a fuzzy Bayesian framework to comfor T7-O1 bipolar EEG channel) and the driver's sleepiness
bine the information obtained from facial features, ECG,
level recorded in a simulator experiment. In Figure 6, both
photoplethysmography, temperature, and a three-axis accelgraphs in (a) and (b) illustrate the relationship between the
erometer to monitor driver drowsiness. Still, fusion techchanges in sleepiness level and the EEG signal. Note that the
niques for driver monitoring systems are in their early stages
sleepiness scale is a subjective scale with limited granularity
and are not fully explored yet. Potential reasons include: 1)
(five discrete levels) whereas EEG provides a nonsubjective
relatively recent success in computationally efficient realcontinuous measure of sleepiness. As shown in the lower plot
time analysis of video data, 2) relatively recent advancements
of Figure 6(a), corresponding to the 3-minute moving averin nonintrusive measurement techniques for physiological
age window (dashed line), the brain wave power ratio shows
signals, and 3) the complex nature of sensing technologies
an increasing trend with increasing sleepiness level, and vice
in driver monitoring that usually requires multidisciplinary
IEEE SIgnal ProcESSIng MagazInE

November 2016

Table of Contents for the Digital Edition of Signal Processing - November 2016