Advanced EEG Processing for the Detection of Drowsiness in Drivers
Griet Goovaerts
1,2
, Ad Denissen
3
, Milica Milosevic
1,2
, Geert van Boxtel
4
and Sabine Van Huffel
1,2
1
KU Leuven, Department of Electrical Engineering (ESAT), STADIUS Center for Dynamical Systems,
Signal Processing and Data Analytics, Leuven, Belgium
2
iMinds Future Health Department, Belgium
3
Brain, Body and Behavior Group, Philips Research, Eindhoven, The Netherlands
4
Department for Psychology, Tilburg University, Tilburg, The Netherlands
Keywords:
EEG, Drowsiness Detection, Blind Source Separation.
Abstract:
Drowsiness is a serious problem for drivers which causes many accidents every day. It is estimated that drowsi-
ness is the cause of four deaths and 100 injuries per day in the United States. In this paper two methods have
been developed to detect drowsiness based on features of ocular artifacts in EEG signals. The ocular artifacts
are derived from the EEG signals by using Canonical Correlation Analysis (BSS-CCA). Wavelet transforms
are used to automatically select components containing eye blinks. Sixteen features are then calculated from
the eye blink and used for drowsiness detection. The first method is based on linear regression, the second on
fuzzy detection. For the first method, the drowsiness level is correctly detected in 72% of the epochs. The sec-
ond method uses fuzzy detection and detects the drowsiness correctly in 65% of the epochs. The best results
are obtained when using one single eye blink feature.
1 INTRODUCTION
Drowsiness is the strong desire to sleep felt before
actually falling asleep (Geetha and Geethalakshmi,
2011). When someone is drowsy, he is less alert and
will perform less efficiently tasks that require a lot
of concentration (Borghini et al., 2012). It is a nat-
ural process which normally causes no extraordinary
problems but it can be dangerous while performing
a task that requires mental attention such as driving.
Drowsy drivers are estimated to be the cause of 7%
of all traffic accidents. Additionally, 16% of all acci-
dents with at least one casualty involve a tired driver
(A.A.A. Foundation, 2010), causing more than four
deaths and 100 injuries per day in the United States
(Borghini et al., 2012). Interviews with drivers indi-
cate that, while they are not aware of the moment they
fall asleep, they are aware of the fact that they are be-
coming drowsy (Reyner and Horne, 1998). Detect-
ing the level of drowsiness and giving an alarm sig-
nal when this level becomes too high would thus be a
way to decrease the number of accidents and danger-
ous situations caused by drowsy driving.
Drowsiness has an effect on many physiological
parameters, such as EEG rhythms, heart rate vari-
ability and EOG signals (Borghini et al., 2012). For
drowsiness detection in drivers, the best base for de-
tection is the use of EOG signals. (Borghini et al.,
2012) indicate that characteristics of EOG signals sig-
nificantly change while performing a task that re-
quires visual attention such as driving. Often, features
such as velocity, duration and amplitude are derived
from the EOG signal and variations in those features
are used to detect drowsiness (Yue, 2011) (Svensson,
2004) (Picot et al., 2011). Different detection ap-
proaches have been used, using for example Support
Vector Machines (Hu and Zheng, 2009), multiple re-
gression (Verwey and Zaidel, 2000) and artificial neu-
ral networks (Vuckovic et al., 2002). In this paper
a simpler approach using linear regression and fuzzy
detection is explored. The validation of drowsiness
detection is done by comparing the detected drowsi-
ness level with a subjective rating given by the subject
himself. The rating can be given on the Karolinska
sleepiness scale (KSS) (Kaida et al., 2006) which has
values from one to nine. Often, a scale with less gra-
dations is used to make detection easier.
In this paper, drowsiness is detected using the ocu-
lar artifacts present in the EEG signal. The brain sig-
nal is first separated from the artifact signal using a
blind source separation method. Out of all BSS meth-
ods, Canonical Correlation Analysis is chosen (BSS-
205
Goovaerts G., Denissen A., Milosevic M., van Boxtel G. and Van Huffel S..
Advanced EEG Processing for the Detection of Drowsiness in Drivers.
DOI: 10.5220/0004800102050212
In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2014), pages 205-212
ISBN: 978-989-758-011-6
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
CCA), since (Vergult et al., 2007) and (Borga and
Knutsson, 2001) show that CCA performs better in
separating brain and muscle signals and equally well
for brain and ocular signals than alternatives such as
ICA. Furthermore, it performs an order of a magni-
tude faster than ICA (Borga and Knutsson, 2001).
2 DATA COLLECTION
The data used in this project are obtained from an
experiment that was conducted by Philips Research
in collaboration with the Dutch Organization for Ap-
plied Scientific Research (TNO) in February 2011.
20 male experienced drivers and 7 back-up sub-
jects between 25 and 45 years old were selected and
had to be present in the testing center for one full day.
During this day they finished three different drives in
a driving simulator and filled in a number of ques-
tionnaires about their sleep quality. In the morning,
the participant drove for one hour in the simulator.
During this time, the baseline measurement was per-
formed: a measurement of the participant’s biosig-
nals when he is alert. In the afternoon session, two
more drives of 3.5 hours had to be finished where the
participant was driving in monotonous traffic condi-
tions. Every 5 minutes a questionnaire appeared on
the screen where the subject had to indicate his state
of alertness at that moment.
Different types of signals are recorded throughout
the experiment. The first type are biosignals. They in-
clude EEG, EOG, heart rate and respiration measure-
ments. The sampling rate of all biosignals is 1024 Hz.
The EEG signals were measured with 32 electrodes
using the International 10-20 system (Homan et al.,
1987). In addition, the user perception questionnaires
assess how tired the subject feels. The Karolinska
sleepiness scale is used, a 9-graded scale with 1 being
extremely alert and 9 extremely sleepy (Kaida et al.,
2006). A digital version of KSS is shown on a screen
every 5 minutes, and the participant has to indicate his
state of alertness.
Not every participant managed to finish all three
drives, and for two participants the KSS-rating was
not recorded. Therefore, the final dataset consists of
the (bio)signals of 19 subjects.
3 METHODOLOGY
The complete drowsiness detection consists of five
steps. First, the EEG signal is decomposed in dif-
ferent components to separate brain signal and EOG
signal. The EOG components are selected and EOG
Figure 1: Illustration of method for component selection.
features are extracted from these components. Then,
the drowsiness detection is performed. Two methods
have been developed: one based on linear regression
and one on fuzzy detection. Finally, both methods
are validated by comparing the results with the KSS-
rating.
3.1 Component Decomposition
A Blind Source Separation method (BSS), BSS-CCA,
is used to decompose the signal in different compo-
nents arising from different sources:
X(t) = A ·s(t) (1)
X(t) is the measured EEG signal, s(t) are the different
sources and A is a mixing matrix that represents the
distribution of the different sources over all channels.
BSS-CCA will generate components that are mutu-
ally uncorrelated but with maximal temporal correla-
tion within each component. The complete algorithm
is described by De Clercq et al. (De Clercq et al.,
2006).
The signal is first divided in subsequent epochs of
1 second. One second is experimentally determined to
be the optimal window length for this dataset. BSS-
CCA decomposes this epoch in components. If the
EEG signal has n channels, the result will be n com-
ponents.
3.2 Component Selection
The identification of the EOG components is done
using a wavelet-based method, based on Krishnaveni
et al. (Krishnaveni et al., 2006). Figure 1 illustrates
the different steps.
First, each component is scaled with a coefficient
c:
c
i
= A(1, i) (2)
A is the mixing matrix from Equation (1), i is the in-
dex of the components. The first row of the mixing
matrix corresponds to electrode Fp1. Since eye blinks
are best visible on the signal of this electrode, the am-
plitude of the artifact-components will enlarge, lead-
ing to easier detection. This can be seen on Figure 1:
the first row is an epoch containing three eye blinks
BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
206
and the second row is one of the scaled components,
where the actual eye blinks are more clearly visible.
Each component is then decomposed with the Haar
wavelet up to 4 levels. The Haar wavelet is used since
Salwani and Jasny show that it gives the best results
in detecting ocular artefacts compared to other mother
wavelets (Salwani and Jasmy, 2005) . The reconstruc-
tion of the approximation coefficients results in a step
function with a rising edge when the eye closes and a
falling edge when the eye opens, as is shown on the
third row of Figure 1. When there is no ocular artifact
present in the component, the approximation coeffi-
cients will have low values. The artifact components
are then identified by comparing the maximal approx-
imation coefficient with a threshold of 80. If the max-
imum is larger than 80, the component is identified as
an EOG component. The value of 80 has been deter-
mined experimentally. When the epoch contains no
artifacts, the component with the highest approxima-
tion coefficient is selected. This way, for each epoch
j, a new signal X
EOG
(t) is constructed:
X
EOG, j
(t) =
k
i=1
A
j
(1, i) · s
i, j
(t) j (3)
In this equation, j is the index of the epoch, A
j
and
s
j
are respectively the mixing matrix and components
found in epoch j, i is the index of the component and
k are the number of components identified as EOG.
When the signals for all epochs are concatenated,
they form a signal with the same length as the orig-
inal EEG signal but with only one row. The newly
formed signal, from now on referred to as ’artifact
signal’, contains all artifacts found by BSS-CCA and
only minimal EEG information. It is thus ideal for the
derivation of different eye blink features.
3.3 Feature Calculation
First, the time instance when the artifact is at peak
amplitude is detected using the wavelet-based method
described in Section 3.2. The approximation coeffi-
cients are now compared with a threshold to detect
the individual eye blinks. Now, the threshold is deter-
mined for each subject individually to make detection
more accurate. The actual moment of the peak ampli-
tude is then found by looking for a local maximum in
the artifact signal 0.15 s around the detected blink.
Four points are detected for each eye blink, which
are indicated on Figure 2: the onset (red) and end
(blue) of the eye blink, the half-rise time (green) and
the half-fall time (orange). First, a moving average fil-
ter with length ten samples smooths the artifact signal.
The first derivative of the smoothed signal consists of
a part larger than zero followed by a part smaller than
Figure 2: Eye blink with detected points (Svensson, 2004).
zero, corresponding to the rising and falling flank of
the eye blink. The onset and end are then found by
looking for zero-crossings in the derivative.
For each eyeblink, 16 features are now derived
from these values that can be grouped in three groups:
amplitude features, time features and velocity fea-
tures. The different features and their calculation are
listed in Table 1. During the experiment, the subject
only had to score his drowsiness on the KSS-scale
once every ve minutes. Therefore, all eyeblinks in
each five minute interval are grouped together and two
values for each feature are calculated: the mean value
and standard deviation over a five minute interval.
This means that in the end there are 32 features.
3.4 Drowsiness Detection
Training data are used to determine which features are
affected by drowsiness and to set the different param-
eters needed for the prediction. A sufficient amount of
training data is needed since the training data need to
contain enough variation in drowsiness levels. In this
case, the data of the second drive are used as training
data, the data of the third drive as test data.
The calculation of all 32 features is done for the
training data and Pearson’s product-moment correla-
tion coefficient between the KSS-rating and each fea-
ture is then computed (Fenton and Neil, 2012). The
significance of the correlation coefficient is measured
by its p-value. When p is close to zero, there is sig-
nificant correlation. The p-values of all correlation
coefficients are calculated and sorted in increasing or-
der. Both methods use maximally the m features with
a p-value smaller than 0.2. Typically, p-values of 0.01
or 0.05 are used, but this value was increased to assure
that at least one feature is chosen so drowsiness can
be predicted. It is also possible to use less features,
in that case the n features with the lowest p-values are
selected.
3.4.1 Detection based on Linear Regression
Hargutt and Kruger show that there is a linear
relationship between various EOG parameters and
AdvancedEEGProcessingfortheDetectionofDrowsinessinDrivers
207
Table 1: EOG features and their calculation.
Group Feature Calculation
Amplitude
Peak amplitude Peak amplitude
Rise value Peak amplitude – onset amplitude
Fall value Peak amplitude – end amplitude
Onset difference Onset amplitude – end amplitude
Time
Total length End time – onset time
Rise length Peak time – onset time
Half-rise length Peak time – half-rise time
Half-fall length Half-fall time – peak time
Fall length End time – peak time
Half length Half-fall time – half-rise time
Time between two blinks
(peaktime)
t
Velocity
Rise velocity
Rise value
Rise length
Half-rise velocity
0.5Rise value
Hal f rise length
Fall velocity
Fall value
Fall length
Half-fall velocity
0.5Fall value
Hal f f all length
Total velocity
Rise value
Total length
drowsiness stages (Hargutt and Kruger, 2001). For
each of the selected features, this linear relationship
between the feature and the KSS rating is modeled by
fitting them with a polynomial of degree 1:
KSS = a
i
· f eature + b
i
(4)
The process of detecting the eye blinks and calcu-
lating the features is now repeated for the test data.
Only the features that had significant correlation in
the training data are calculated. The drowsiness is
then predicted by using linear regression to predict
the drowsiness based on each feature and taking the
average over all predictions:
KSS =
1
n
n
i=1
KSS
i
(5)
=
1
n
n
i=1
a
i
· f eature
i
+ b
i
(6)
i is the feature’s index and n is the number of fea-
tures that are used to predict the drowsiness, which
can range from 1 to m. Increasing n will increase the
number of features used, but each extra feature will
have a less strong correlation with the KSS rating than
the previous features.
3.4.2 Fuzzy Detection
The second method is based on a paper by Picot et al.
(Picot et al., 2011). For each feature that is used, two
parameters a and b are calculated:
a = µ
i
+ 0.25µ
i
(7)
b = µ
i
0.25µ
i
(8)
where i is the index of the feature and µ
i
the mean
value of feature i in the training data. The feature
calculation is repeated for the test data, and the cal-
culated features are converted to fuzzy indicators: a
number between 0 and 1 that indicates the drowsy
state. When this number is closer to one, the drowsi-
ness level will be higher. When the correlation be-
tween the feature and the KSS rating is positive, the
fuzzy drowsiness indicator is calculated by:
D( f
i
) =
0, if f
i
a
f
i
a
ba
, if a f
i
b
1 if f
i
b
(9)
When the correlation is negative, Equation (9) be-
comes
D( f
i
) =
1, if f
i
a
f
i
b
ba
, if a f
i
b
0 if f
i
b
(10)
In both equations (9) and (10), D is the drowsiness in-
dicator, i the index of the feature and f
i
the value of
feature i. Finally, the mean of the drowsiness indica-
tors is taken and converted to a number between one
and nine:
KSS = 8 · (
1
n
n
i=1
D
i
) + 1 (11)
n is again the number of features used to predict the
drowsiness, which can be varied to include more or
less features.
3.5 Validation
Finally, the prediction results are compared with the
KSS rating given by the subjects to examine the qual-
BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
208
ity of the prediction. The comparison can be pre-
sented by constructing a confusion matrix. The values
on the diagonal are the most important values, since
they represent a perfect prediction. However, since
KSS is a scale with nine gradations, results where
the prediction differs one value from the true rating
can also be considered good results. Therefore, from
now on the distinction will be made between perfect
results (e.g. the predicted value is equal to the real
value) and good results (e.g. the difference between
the predicted and true value is one).
4 RESULTS
The eye blinks present in the signal are first counted
manually and the result is compared with the num-
ber of eye blinks detected by the methods described
in Section 3. The developed methods detect in total
85.7 ± 8.1% of all eye blinks. The different features
and their correlation with the KSS rating are then cal-
culated. Features for which the correlation coefficient
is small enough are used for drowsiness detection.
The ten features that are most often used to detect
drowsiness are listed in Table 2.
Table 2: Features included in most subsets.
Feature %
Mean half-rise speed 89.4
Mean rise speed 84.22
Mean fall value 78.9
Std half-fall length 78.9
Mean half-fall speed 78.9
Mean rise value 73.6
Mean half-rise length 73.6
Mean rise-length 73.6
Mean peak amplitude 73.6
Std total length 73.6
The first column is the name of the feature and the
second column the percentage of subsets that includes
the feature. When a feature is included in a lot of sub-
sets, it means that the correlation between the feature
and the KSS rating is significant for a lot of subjects.
The majority of the most frequently included features
are calculated as the mean value over a five-minute
interval. Only two features are the standard deviation
over a five-minute interval.
4.1 Method 1: Linear Regression
Figure 3 (a) shows the results for the drowsiness de-
tection using the first method for one subject. The
predicted values are shown in red, the true values in
Figure 3: Results for drowsiness detection using method 1
(left) and method 2 (right).
blue. One thing is immediately clear: the predicted
values and true values have the same trend over time,
but there appears to be a constant difference of four
between the red and blue values. This difference,
or bias, is observed in almost all results. If the bias
is added to the predicted values, the results improve
tremendously: in the example from Figure 3 orig-
inally only 12% of the results were predicted well
(meaning a difference of one between the predicted
and the true value) and none of the values were pre-
dicted perfectly. If a bias of four is added, 75% of the
results are good and 54% of the results are perfect.
For each subject, the bias is determined visually and
added to the predicted values.
In Equation (6), n, the number of features used
to predict drowsiness can be varied. In Table 3 (a)
the average percentage of good results is shown for
different numbers of features used.
Table 3: Effect of number of features of results of (a) linear
regression and (b) fuzzy detection for one subject.
(a) Linear regression
Number % good
1 75.4
2 37.9
3 25.6
4 14.7
5 9.13
(b) Fuzzy detection
Number % good
1 71.4
2 68.1
3 63.4
4 61.8
5 60
The best results are obtained using only one fea-
ture. The accuracy decreases consistently with rising
n. Therefore, the number of features is fixed at one
and the complete dataset is processed likewise. The
confusion matrix is shown in Table 4, the sum of the
good and perfect results is shown at the end of each
row.
The drowsiness prediction works well for large
KSS values. When the true value is one or two how-
ever, the majority of the predictions are wrong.
The results for individual subjects are good as
well. On average, 72 ± 18% of the predictions are
good. The percentage of perfect results is a lot lower:
29 ± 11%
AdvancedEEGProcessingfortheDetectionofDrowsinessinDrivers
209
Table 4: Confusion matrix for linear regression expressed in percentages.
Predicted value
1 2 3 4 5 6 7 8 9
1 0 1.85 22.22 25.92 25.92 14.81 7.40 1.85 0 1.85
2 0 8.33 8.33 58.33 25 0 0 0 0 16.66
3 0 3.22 35.48 45.16 0 3.22 6.45 3.22 3.22 83.86
4 0 0 24.50 32.35 19.60 8.82 2.94 5.88 5.88 76.45
True 5 0 0.76 6.87 22.13 22.13 26.71 14.50 2.29 4.58 70.97
value 6 1.76 1.76 4.42 7.07 13.27 26.54 38.05 6.19 0.88 77.86
7 2.87 2.15 0.71 1.43 7.19 12.94 33.09 30.93 8.63 76.96
8 0.69 1.38 1.38 1.38 3.47 4.86 25 29.86 31.94 86.8
9 3.33 1.11 1.11 2.22 6.66 14.44 10 11.11 50 61.11
4.2 Method 2: Fuzzy Detection
The second method again has a bias, as can be seen on
Figure 3(b), although the bias is smaller. Table 3(b)
shows that using one feature also gives the best result
in this case.
Table 5 shows the confusion matrix.
The predictions for a KSS rating of one are not
good: only 6.45% is predicted well. the results for
the other values are a lot better, the majority of predic-
tions differ maximum one value from the true rating.
After examining the results of all subjects, 34 ± 12%
of all predictions are perfect and 65±12.2% are good.
5 DISCUSSION
In this paper, drowsiness is detected based on the arti-
facts in the EEG signal. Equation 3 constructs a new
signal containing all artifacts and a minimum of EEG
signal. The signal will be similar to the EOG signal,
measured by placing electrodes around the eye, since
the EOG signal also contains all EOG artifacts and
minimal EEG signal. It would therefore be possible
to use the described methods with electrodes placed
around the eye. The advantage would be a reduction
in the number of electrodes which would benefit the
user friendliness. However, an advantage of using the
complete EEG signal is that it contains a lot more in-
formation than merely EOG artifacts. This informa-
tion can be either be used for other purposes (such as
stress or concentration monitoring) or can be incorpo-
rated in the drowsiness detection to make the results
more accurate. Since CCA can effectively remove
both ocular and muscle artifacts, further calculations
can immediately be done on the clean EEG signal.
This makes the detection of other drowsiness indica-
tions in the EEG signal easier, since they are often
missed when the signal quality is bad (Simon et al.,
2011). It would be very interesting to repeat the cal-
culations using less electrodes, to see if similar results
can be obtained. This would increase the practical us-
ability, while still recording EEG signals that can be
used for further drowsiness detection.
The optimal number of features to detect drowsi-
ness is shown to be one. In one way, this could be
expected since the features are sorted based on the
significance of their correlation. The feature with the
strongest correlation gives the best results. Adding
next important features didn’t increase the detection
performance, even in the case when the first predic-
tion was incorrect. This indicates that the correlation
of the first feature is a lot more significant than the
other features.
Table 2 shows ”the most popular” features for
each set of training data. Mean speed and mean half-
rise speed are clearly two key features; they are in-
cluded in the majority of the subsets. First it has to be
noted that it is very probable that the rise speed and
half-rise speed are very correlated with each other. If
the speed of the total eye opening increases, it is ex-
pected that the speed at which half the eye opens in-
creases as well. Therefore it might be useful to first
reduce the number of features using principal compo-
nent analysis. PCA can be used to convert a set of
possibly correlated variables to a set of linearly un-
correlated variables (Abdi and Williams, 2010). Sec-
ond, it is curious that the correlation of the rise speed
decreases most with increasing drowsiness. (Yue,
2011), (Svensson, 2004) and (Borghini et al., 2012)
all state that eye parameters such as blink duration
and frequency are influenced most by drowsiness, but
no one of them names velocity as an important pa-
rameter. The reason for this difference might lie in
the nature of the experiment. During the experiment,
the participants had to look at a screen for longer than
three hours. (Rosenfield, 2011) demonstrates that this
may reduce the blink rate and cause Computer Vision
Syndrome, leading to dry and irritated eyes, which
may again influence eye blink behaviour. It is thus
BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
210
Table 5: Confusion matrix for fuzzy detection expressed in percentages.
Predicted value
1 2 3 4 5 6 7 8 9
1 0 6.45 9.67 25.8 12.9 16.12 19.35 3.22 6.45 6.45
2 0 6.25 25 50 6.25 6.25 6.25 0 0 31.25
3 0 1.47 29.41 32.35 23.52 7.35 2.94 1.47 1.47 63.23
4 0 .99 0.99 16.83 31.68 23.76 14.85 3.96 2.97 3.96 72.27
True 5 0 2.06 3.09 19.58 21.64 18.55 14.43 11.34 9.27 59.77
value 6 2.43 1.62 9.75 7.31 12.19 30.89 22.76 10.56 2.43 65.84
7 3.61 1.2 1.8 4.81 8.43 13.85 41.56 16.26 8.43 71.76
8 0.84 0.84 1.68 2.52 4.2 6.72 22.68 42.01 18.48 83.17
9 5.26 0 3.15 2.10 1.05 7.36 10.52 8.42 62.10 70.52
probable that in a natural environment (e.g. while
driving a real car instead of a driving simulator) other
features will be more important. Since using one fea-
ture gives the best prediction results and mean speed
and mean half-rise speed are clearly features that of-
ten have significant correlation with the drowsiness
level, it would be possible to only calculate one of
both features.
Figure 3 shows that there is a difference between
the drowsiness level predicted and the true drowsiness
level. The bias is different for each subject. There
could be two reasons why there is a bias present. First,
the subjects do not rate their drowsiness consistently.
They for example think they are less tired than they
actually are. Since the subjects rate their drowsiness
after a questionnaire is shown on a screen, it is pos-
sible that the appearance of this questionnaire influ-
ences the perceived drowsiness. If that is the case,
the questionnaire arouses the subject for a short time,
effectively decreasing the drowsiness rating. There-
fore, it would be interesting to obtain a more objec-
tive drowsiness rating by analyzing video recordings
or the driving behaviour of the subjects. Second, the
coefficients of the polynomial that is fitted to the train-
ing data might not be entirely correct. This may be ex-
plained by the fact that there are not many data avail-
able for low KSS values.
The second reason is obviously not relevant for
explaining the offset in the second method since no
linear fit is used. However, there is also a bias no-
ticed in the results of the second method, although the
bias is smaller than in the results of the first method.
Therefore, it is suspected that both reasons apply in
this case.
Finally, the results of both methods can be con-
sidered good, also when compared with other similar
methods. Svensson (Svensson, 2004) detects drowsi-
ness based on changes in blink behaviour and uses a
four graded scale. There, a correspondence with KSS
values of 70% is obtained. Picot et al. (Picot et al.,
2011) obtain an accuracy of 82% when distinguishing
between drowsy and not-drowsy, making the classifi-
cation significantly easier.
It is not easy to choose the best method among the
two tested here, since they both have their strengths
and weaknesses. If the two confusion matrices are
compared, it is clear that the method based on fuzzy
detection gives more perfect results and the linear
regression-based method gives more good results.
The differences are however small. They also show
that the fuzzy detection performs better for small KSS
values (although the result for KSS = 1 is still not
good). However, since the aim is detecting drowsi-
ness, large values are more important. Even more,
since drowsiness varies gradually, a small difference
in the true and predicted values is not very important.
Therefore, in most applications the linear-regression
based method would be preferred. However, when
perfect classifications need to be obtained, one would
probably start with fuzzy detection. In that case, the
method would have to be adjusted to improve the ac-
curacy.
6 CONCLUSIONS
The methods described in this paper are ways of de-
tecting drowsiness based on EOG artifacts present in
the EEG signal. They work semi-automatic: both the
threshold to detect eyeblinks and the bias are deter-
mined individually. The results of the methods are
good when perfect accuracy is not required. On av-
erage 70% of the predictions differ at most one value
from the KSS rating given by the subject. In applica-
tions when perfect accuracy is required however, the
methods should be improved or combined to give bet-
ter results. The results of the drowsiness detection
could be improved by extending the developed meth-
ods. It is possible to incorporate more information
extracted from the EEG in the methods. This way, the
accuracy and more specifically the number of perfect
results would be increased.
AdvancedEEGProcessingfortheDetectionofDrowsinessinDrivers
211
If the methods were to be used in real-life appli-
cations, they need to be used in real-time. Therefore,
the methods would have to be adapted. It would for
example be possible to calculate the features in a slid-
ing window, hereby giving a continuous prediction.
The thresholds for eye blink detection, which are now
defined manually, can be fixed during a test drive. To
increase the practical usability, the number of elec-
trodes should be decreased, since it is virtually impos-
sible to equip drivers with a full EEG cap. Finally, to
reduce the computational load, it can be researched if
the method gives good results when only calculating
one or two features, instead of calculating 32 features
and selecting one of them.
REFERENCES
A.A.A. Foundation (2010). Asleep at the wheel: The preva-
lence and impact of drowsy driving.
Abdi, H. and Williams, L. (2010). Principal component
analysis. Wiley Interdisciplinary Review: computa-
tional statistics, (2):433–459.
Borga, M. and Knutsson, H. (2001). A canonical ap-
proach to Blind Source Separation. Report LiU-
IMT-EX-0062 Department of Biomedical Engineer-
ing, Linkping University.
Borghini, G., Astolfi, L., Vecchiato, G., Mattia, D., and Ba-
biloni, F. (2012). Measuring neurphysiological signals
in aircraft pilots and car drivers for the assessment
of mental workload, fatigue and drowsiness. Neuro-
science and Biobehavioral Reviews, pages 45–57.
De Clercq, W., Vergult, A., Vanrumste, B., Van Paesschen,
W., and Van Huffel, S. (2006). Canonical Correla-
tion Analysis Applied to Remove Muscle Artifacts
From the Electroencephalogram. IEEE transactions
on Biomedical Engineering, 53(12):2583–2587.
Fenton, N. and Neil, M. (2012). Correlation coefficient and
p-values: what they are and why you need to be wary
of them. In Risk assessment and Decision Analysis
with Bayesian Networks. CRC Press.
Geetha, G. and Geethalakshmi, S. (2011). Scrutinizing dif-
ferent techniques for artifact removal from EEG sig-
nals. International Journal of Engineering Science
and Technology, (1).
Hargutt, V. and Kruger, H. (2001). Eyelid movements and
their predictive value for fatigue stages. In Interna-
tional Conference on Traffic and Transport Psychol-
ogy.
Homan, R. W., Herman, J., and Purdy, P. (1987). Cere-
bral location of international 10–20 system electrode
placement. Electroencephalography and clinical neu-
rophysiology, 66(4):376–382.
Hu, S. and Zheng, G. (2009). Driver drowsiness detection
with eyelid related parameters by support vector ma-
chine. Expert Systems with Applications, 36(4):7651–
7658.
Kaida, K., Takahasi, M., Akerstedt, T., Nakata, A., Otsuka,
Y., Haratani, T., and Fukasawa, K. (2006). Valida-
tion of the Karolinska sleepiness scale against perfor-
mance and EEG variables. Clinical Neurophysiology,
7(117):1574–1581.
Krishnaveni, V., Jayaraman, S., Aravind, S., Hariharasud-
han, V., and Ramadoss, K. (2006). Automatic Iden-
tification and Removal of Ocular Artifacts from EEG
using Wavelet Transform. Measurement science re-
view, 6(2):45–57.
Picot, A., Charbonnier, S., and Caplier, A. (2011). EOG-
based drowsiness detection: Comparison between
a fuzzy system and two supervised learning classi-
fiers. Preprints of the 18th IFAC World Congress,
18:14283–14288.
Reyner, L. and Horne, J. (1998). Falling asleep whilst driv-
ing: are drivers aware of prior sleepiness? Int J Legal
Med, 3(111):120–123.
Rosenfield, M. (2011). Computer vision syndrome: a re-
view of ocular causes and potential treatments. Oph-
thalmic and Physiological Optics, 31(5):502–515.
Salwani, M. and Jasmy, Y. (2005). Comparison of few
wavelets to filter ocular artifacts in eeg using lifting
wavelet transform. In TENCON 2005 2005 IEEE Re-
gion 10, pages 1–6. IEEE.
Simon, M., Schmidt, E. A., Kincses, W. E., Fritzsche, M.,
Bruns, A., Aufmuth, C., Bogdan, M., Rosenstiel, W.,
and Schrauf, M. (2011). Eeg alpha spindle measures
as indicators of driver fatigue under real traffic condi-
tions. Clinical Neurophysiology, 122(6):1168–1178.
Svensson, U. (2004). Blink behaviour based drowsiness de-
tecion - method development and validation. Master’s
thesis, University of Link
¨
oping.
Vergult, A., De Clercq, W., Palmini, A., Vanrumste, B.,
Dupont, P., Van Huffel, S., and Van Paesschen, W.
(2007). Improving the Interpretation of Ictal Scalp
EEG: BSS-CCA algorithm for muscle artifact re-
moval. Epilepsia, 48(5):950–958.
Verwey, W. B. and Zaidel, D. M. (2000). Predicting drowsi-
ness accidents from personal attributes, eye blinks and
ongoing driving behaviour. Personality and Individual
Differences, 28(1):123–142.
Vuckovic, A., Radivojevic, V., Chen, A. C., and Popovic,
D. (2002). Automatic recognition of alertness and
drowsiness from eeg by an artificial neural network.
Medical Engineering & Physics, 24(5):349–360.
Yue, C. (2011). EOG signals in drowsiness research. Mas-
ter’s thesis, University of Link
¨
oping.
BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
212