An on-board vision based system for drowsiness detectionin automotive drivers
Anirban Dasgupta • Anjith George •
S. L. Happy • Aurobinda Routray • Tara Shanker
Published online: 3 August 2013
� Indian Institute of Technology Madras 2013
Abstract This paper proposes a system for on-board
monitoring the loss of attention of an automotive driver,
based on PERcentage of eye CLOSure (PERCLOS). This
system has been developed considering the practical on-
board constraints such as illumination variation, poor illu-
mination conditions, free movement of driver’s face, limi-
tations in algorithms etc. A novel framework for PERCLOS
computation is reported in this paper. The system consists of
an embedded processing unit, a camera, a near infra-red
lighting system, power supply, a set of speakers and a voltage
regulation unit. The image based algorithm is based on the
PERCLOS as an indicator of the loss of attention of the
driver. The authenticity of PERCLOS as an indicator of
drowsiness has been validated using EEG signals.
Keywords Real-time algorithm � PERCLOS �On-board testing � NIR lighting � Electroencephalography
1 Introduction
A major cause of road accidents is the loss of attention in
automotive drivers, which has been a concern for transpor-
tation safety over the years. Hence, on-board monitoring of
the alertness level of the driver is necessary to prevent such
occurrences. The work in [1] highlights such attempts.
Alertness level in drivers can be assessed using different
measures such as Electroencephalogram (EEG) [2–4] ocular
features [5], blood samples [5], speech [6, 7], skin conduc-
tance [8] etc. The EEG based method has been reported to be
the most authentic cue for estimating the reduction in alert-
ness level [2, 3]. However, being a contact based method, its
feasibility of implementation becomes impractical. In this
paper, image cues have been considered for implementation
on virtue of being non-contact based method. EEG based
method have been used for off-board validation of the vision
based algorithm. The first contribution to this paper is
development of a novel framework for estimating the PER-
centage of eye CLOSure (PERCLOS) in real-time thereby
overcoming certain issues prevailing in literature. The sec-
ond contribution is validation of the algorithm using EEG
signals by conducting off-board experiments.
Several approaches have been reported to develop sys-
tems for on-board monitoring the alertness level in auto-
motive drivers. The work in [9] reports a system named as
Driver AssIsting System (DAISY) as a monitoring and
warning aid for the driver in longitudinal and lateral control
on German motorways. The warnings are generated based
on the knowledge of the behavioural state and condition of
the driver. However, this is a model based approach and
proper on-board testing and validation has not been dis-
cussed. Singh and Papanikolopoulos [10] proposed a non-
invasive vision-based system for the detection of fatigue
level in drivers. The system uses a camera that points
directly towards the driver’s face and monitors the driver’s
eyes in order to detect micro-sleeps. The system is reported
to detect the eyes at 10 fps and track at 15 fps. The
accuracy of the system is about 95 % with small movement
of head bearing a tolerance on head tilt up to 30�. However
in practical driving scenario, this system has drawbacks
where driver’s head movement introduces inefficiency of
the proposed system. Robotics Institute in Carnegie Mellon
University has developed a drowsy driver monitor named
A. Dasgupta (&) � A. George � S. L. Happy � A. Routray
Indian Institute of Technology Kharagpur, Kharagpur, India
e-mail: [email protected]
T. Shanker
DeitY, NewDelhi, India
123
Int J Adv Eng Sci Appl Math (April–September 2013) 5(2–3):94–103
DOI 10.1007/s12572-013-0086-2 IIT, Madras
Copilot [11], which is a video-based system which esti-
mates PERCLOS. However the system has not been
addressed to act robust against illumination variation, head
rotation etc. The co-pilot system uses bright pupil dark
pupil method for the detection of eyes which requires
active illumination method. A similar system called Driver
Assistance System (DAS) has been developed by the group
at the Australian National University [12]. It uses a dash-
board-mounted head-and-eye-tracking system to monitor
the driver. The Distillation algorithm has been used to
monitor the driver’s performance. Feedback on deviation in
lane tracking is provided to the driver using force feedback
to the steering wheel which is proportional to the amount of
lateral offset estimation by the lane tracker. However in the
systems [10–12], the issues such as illumination variation
and head rotations have not been adequately addressed.
Smith et al. [13] described a system in which the driver’s
head is detected using colour predicates and tracked using
optical flow. The system described is reported to be robust
against face rotations, as well as head and eye occlusions.
Eye occlusions are detected by the likelihood of rotational
occlusion of eyes. The algorithm gives an approximate 3D
gaze direction with a single camera. The size of head is
assumed constant among different persons so the gaze
direction and can be estimated without prior knowledge of
the distance between head and camera. However the
accuracy of the system is illumination dependent. Zhu and
Ji [14] presented a real-time system for on-board moni-
toring the driver using active infrared illumination. Visual
cues such as PERCLOS, Average Eye Closure per Second
(AECS), 3D head orientation facial expressions are fused
with a Bayesian framework to obtain the estimation of
fatigue level. Bergasa et al. [15] developed a system which
monitors the driver’s fatigue level with six different
parameters—PERCLOS, eye closure duration, blink fre-
quency, head nodding frequency, face position, and fixed
gaze. These different cues are combined together using a
fuzzy classifier to get the alertness level of the driver.
Bright-pupil dark-pupil method is used for the detection of
pupil. However, the method being active illumination
based, requires special hardware design for robust perfor-
mance. In [16], Eriksson and Papanikolopoulos proposed a
system to locate and track the eyes of the driver. They used
a symmetry-based approach to locate the face for sub-
sequent detection and tracking of eyes. Template matching
has been used for eye state classification. Sacco and Far-
rugia [17] presented a real-time system for classification of
alertness of drivers. Viola-Jones frame work has been used
for the detection of eyes, and mouth, followed by tracking
of these features using template matching. The final deci-
sion about fatigue level is taken by considering three
parameters—PERCLOS, average eye closure interval and
the degree of mouth opening. SVM is used to fuse these
three different features. However the performance with
head rotations have not been considered. Moreover,
detection of features during night time and effect of illu-
mination has not been addressed in this work.
From the literature review, it is apparent that the existing
systems has limitations against illumination variation, head
rotation poor illumination for assessing driver fatigue. The
present system is a solution which addresses the existing
problems in the assessment of driver’s level of fatigue.
In this paper, a vision based system is proposed which
has been tested systematically on-board conditions for
both day as well as night driving conditions. The frame-
work of estimating the PERCLOS value has been modified
from the previous frameworks to make the system robust
against illumination variation, head rotation poor illumi-
nation etc. The system comprises of a single board com-
puter running on Intel Atom Processor, a camera, a set of
speakers, a near infra-red (NIR) lighting system, power
supply and a voltage regulation unit. The embedded sys-
tem has 1 GB RAM and executes at a clock speed of
1.6 GHz. The algorithms have been cross-validated using
EEG based technique. The algorithm has been validated
using EEG signals.
This paper is organised as follows. Section 2 describes the
vision based algorithm for estimating the state of alertness of
the driver. In Sect. 3, the on-board testing has been described
briefly. Section 4 describes the EEG based validation of the
image based algorithm. Section 5 describes the limitations
of the system. Section 6 concludes the paper.
2 Vision based algorithm
The vision based algorithm is based on the PERCLOS [18]
value of the driver. PERCLOS may be defined as the
percentage of eyelid closure over the pupil over a given
time interval as shown below.
P ¼ Ec
Et
� 100 % ð1Þ
In (1), P is the PERCLOS value, Ec is the number of
closed eye count over a predefined interval while Et is the
total eye count in the same given interval. Literature [18]
states that a higher value of P indicates higher loss of
attention and vice versa. In this paper, the PERCLOS value
is computed for each minute on a running average window
of 3 min duration [19]. Based on a threshold of 15 % [15],
the driver’s alertness state is classified as drowsy or alert.
It is evident from the definition of PERCLOS, that its
estimation precedes face and eye detection. The following
section describes the methodology for face and eye detection.
Int J Adv Eng Sci Appl Math (April–September 2013) 5(2–3):94–103 95
123
2.1 Face detection
The face detection is carried out using a classifier based on
Haar-like features [20]. A robust classifier is trained using
the optimal sets of parameters obtained from our earlier
work [21]. A real-time framework is proposed in which the
runtime performance has been improved by down sampling
the frames. An optimal down sampling scale factor of 5 for
the specific application with camera resolution of
640 9 480 has been obtained experimentally as a tradeoff
between accuracy and runtime. Taking approximate dis-
tance of driver from camera and minimum face size after
down sampling into consideration, the algorithm has been
optimized so that Haar classifier performs accurately with
less time complexity.
2.1.1 Detection of in-plane rotated faces
The classifier mentioned above was found to be robust for
frontal faces. However the issue of detection of rotated
faces still prevailed. Towards achieving robustness for
detection of in-plane rotated faces, an affine transforma-
tion [22] of the input frame has been used. The idea is to
rotate the frame so as to align the face as frontal. The
rotation matrix R is found from the angle of rotation h of
the image matrix as shown in (2). If a point on the input
image is A ¼ x
y
� �and the corresponding point on the
affine transformed image is B ¼ x0
y0
� �, then B can be
related to A using (3).
R ¼ cosh �sinhsinh cosh
� �ð2Þ
A ¼ RB ð3Þ
If frontal face is not detected, a rotational matrix R with
h = ±15� is created, to transform the image using (3). It is
implicit that when a face is detected in a rotated image,
then there is higher probability that face will be detected by
rotating few subsequent frames by the same angle. To
achieve efficiency in time complexity, the algorithm is
optimized such that when a face is detected while rotating
the image by angle hs, then for the next frame the image
will be directly rotated initially to angle hs and then search
for face. If face is not detected then it will search for faces
by rotation of angles hs ± h, hs ± 2 9 h and hs ± 3 9 hThus, the algorithm was found to be robust for every
possible in-plane rotations.
The affine transformation only occurs in the event the
frontal face is not detected by the Haar classifier thereby
preventing appreciable loss in runtime owing to the com-
putational burden of the transform.
2.1.2 Detection of off-plane rotated faces
Affine transformation based method was found to aid the
Haar classifier for in-plane rotated face detection. The goal
of detection of off-plane rotated faces has been achieved by
the usage of Perspective transformation, which is a robust
and computationally inexpensive method for aligning the
head for small off-plane rotations [23]. It is a mapping of
one view of a face to another view. The off-plane rotated
image I x0; y0ð Þ can be related to its original frontal image
Iðx; yÞ using
x0 ¼ a0xþ a1yþ a2
a3xþ a4yþ 1ð4Þ
y0 ¼ a5xþ a6yþ a7
a3xþ a4yþ 1ð5Þ
The coefficients a0–a7 are estimated using non-linear
least squares method as in [23]. On obtaining the
coefficients, the desired perspective mapping is obtained.
For the driving context, cases of off-plane rotations are
observed very rarely and are restricted within ±30� and the
perspective transformation is found to compensate such
rotations.
2.1.3 Performance of face detection
The face detection algorithm is found to provide a hit rate
of around 95 % with a false alarm rate of 2 % under on-
board conditions. Once the face is detected, a region of
interest (ROI) is selected as the upper half of the detected
face region, based on morphology of the face, for eye
detection. This reduces the search space for detection of
eyes.
2.2 Eye detection
For estimation of PERCLOS, accurate detection and
classification of eye state is necessary. As reported in
literature, eye detection may be carried out using pixel to
edge information [24], optimal wavelet packets and radial
basis functions [25], combined binary edge and intensity
information [26], projection functions [27], principal
component analysis (PCA) [28], classifiers based on Haar-
like features [29], support vector machines (SVM) [30],
local binary patterns (LBP) [29] and other methods [31–
33]. The use of block-LBP histogram features is more
efficient in the extraction of information than global his-
togram based LBP [34]. In this paper, block LBP histo-
gram features have been used for real-time eye detection
owing to its illumination invariance property and compu-
tational simplicity.
96 Int J Adv Eng Sci Appl Math (April–September 2013) 5(2–3):94–103
123
2.2.1 LBP features
The LBP operator labels the pixels of an image by thres-
holding the neighborhood of each pixel and considering the
result as a binary number. The use of LBP features for
objet detection has been detailed in [34]. In the present
work, the LBP texture descriptors have been used to build
several local descriptions of the eye and combine them into
a global description. These local feature based methods are
more robust against variations in pose or illumination than
holistic methods. The use of local histograms and pixel
level values of LBP to get the feature vector is given in
Table 1. In the detection phase, LBP histogram features are
obtained in the localized eye region and then projected into
the PCA feature space to get the energy components.
For evaluation of the performance of the algorithm, 300
prerecorded eye images are used for training. The algo-
rithm is tested on the whole database. Figure 1 shows some
eye localization results using block LBP features.
2.3 Eye state classification
For accurate estimation of PERCLOS, the detected eye
needs to be accurately classified into open or closed state.
The aforementioned problem being a binary classification
problem, SVM [31] being reported as a robust binary
classifier, has been used for the purpose.
The weights obtained as described in the eye detection
section, along with the ground truth (open or closed), are
used for the training the SVM. The energy components are
fed into the SVM which returns the class number.
The training of SVM has been carried out with 460 eye
images created using normal and NIR illumination. Some
training images are shown in Fig. 2. The testing is carried
out using another 1,700 images taken from the same
database. The block LBP histogram feature transform is
carried out on the training images. Then PCA is carried out
on the LBP feature vectors. Now the weight vectors cor-
responding to each sample is found by projecting the
samples to PCA subspace. These weights along with the
ground truth are used for SVM training. The training is
carried out with different kernels and accuracy levels and
the classification results are shown in Table 2. The results
reveal that the performance of the SVM is consistent under
normal and NIR illuminated images. The algorithm was
cross validated using EEG signals and reported in an earlier
work [3]. The real-time algorithm is coded into a single
board computer having Intel Atom processor, with
1.6 GHz clock speed and 1 GB RAM. The overall pro-
cessing speed is found to be 12 fps, which is acceptable for
accurate estimation of PERCLOS. The algorithm is tested
under laboratory conditions. However, on-board conditions
are quite different from the laboratory conditions and
therefore testing in practical driving scenarios is necessary.
Hence, two types of on-board tests were performed sepa-
rately for day and night to test the robustness of our
algorithm in actual driving conditions. The only difference
in the night driving tests was the use of NIR lighting.
Once the eye have been classified as open or closed, the
PERCLOS value is computed using (1). A 66.67 % over-
lapping time window of 3 min has been used to compute
the PERCLOS value for every minute and in the event it is
found to be [20 %, an alarm is sounded.
2.4 Overall schematic
The overall schematic is shown in Fig. 3. The PERCLOS
value is computed after an interval of 1 min after the eye is
localized and eye state is classified in real-time. The face
hit rate was also taken as an input, and in the event of low
face hit rate, the driver was alarmed to pay attention to the
road. Figure 4 shows the overall block diagram of the
system while Fig. 5 shows the process flow for estimating
PERCLOS.
3 On board testing
3.1 System description
The alertness monitoring system consists of a single USB
camera having maximum frame rate of 30 fps at a reso-
lution of 640 9 480 pixels. The camera is placed directly
on the steering column just behind the steering wheel to
obtain the best view of the driver’s face. The camera
placement was finalized after a lot of testing to obtain the
best result.
Table 1 Algorithm: block LBP
Training Phase
1. Eye images are resized to 50 9 40 resolution and each image is
divided into sub-blocks of size 5 9 4
2. The LBP feature values of the sub-blocks are found and 16 bin
histogram of each block is calculated
3. The histogram of each block is arranged to form a global feature
descriptor
4. PCA is carried out on the feature vectors and 40 Eigen vectors
having largest Eigen values were selected
Detection phase
1. ROI from face detection stage is obtained
2. For each sub window the block LBP histogram is found and is
projected to Eigen space
3. The sub window with minimum reconstruction error is found
4. If the reconstruction error is less than a threshold it is considered
as a positive detection
Int J Adv Eng Sci Appl Math (April–September 2013) 5(2–3):94–103 97
123
The SBC is powered at 12 V DC from the car supply.
Power supply from the car battery has the advantage of
nullifying the requirement of a separate power source which
adds to the cost of the system as well as space requirement.
The typical current drawn from the input source is
*1,200 mA. The approximate typical power drawn from
the supply is 15 W. A voltage regulator unit, comprising of
IC LM317 along with some resistors, a capacitor and an
inductor, is used before the input to remove high voltage
spikes from the car supply. The voltage regulation unit
along with a 1.5 A fuse ensured protection to all other units.
The regulator unit specifications are obtained after mea-
suring the input current and voltages under actual operating
conditions. The processing unit comprises of Intel Atom
N450 processor having a clock speed of 1.6 GHz. The
motherboard of the embedded platform has x86 architec-
ture. The system has been made very rugged for automotive
environment specially to protect it from vibration and jerks
of the vehicle. An NIR lighting arrangement, consisting of a
matrix of 3 9 8 Gallium Arsenide LEDs, is also connected
across the same supply in parallel with the Embedded
Platform. After thorough testing it was found out that the
3 9 8 matrix was the minimum sufficient size to illuminate
the face correctly. The NIR module is operated at 10 V DC
and draws a typical current of 250 mA. The lighting system
is connected through a light dependent resistor (LDR), to
automatically switch on the NIR module in the absence of
sufficient illumination. A seven inch LED touch screen is
used to display the results. Table 3 gives the power ratings
of the components.
Figure 6 shows the different modules used in the sys-
tem. The speed of the algorithm was found to be 12 fps,
which is very acceptable for estimation of PERCLOS.
3.2 Testing
The main objective of the testing was to evaluate the
performance of the system under on-board conditions.
Using the sequence of frames captured from the camera,
the processor processes the frames to compute PERCLOS.
Based on a threshold value of 15 %, a voice alarm is
generated to warn the driver. The NIR system is switched
on automatically using a LDR based on external lighting
conditions. The tests were carried out on a rough road
followed by a smooth road (highway). Twenty subjects
were used for the tests under day as well as night driving
conditions. The face and eye detection was found to be
accurate for both smooth as well as jerky road conditions.
In first test, during the midday, when the light fell directly
on the camera lens, the camera sensor got saturated and
image features were lost. The problem was rectified in
Fig. 1 Eye localization using
LBP
Fig. 2 Sample training images for eye state classification using SVM
Table 2 Detection results with SVM
Kernel function tp tn fp fn tpr (%) fpr (%)
Linear SVM 838 808 42 12 98.58 4.94
Quadratic 827 765 85 23 97.29 10
Polynomial 848 807 43 2 99.76 5.32
98 Int J Adv Eng Sci Appl Math (April–September 2013) 5(2–3):94–103
123
future tests by using a camera with a superior image sensor
to remove the saturation effect.
The tests were carried out on a rough road followed by a
high way road. The face and eye detection was found to be
fairly accurate. The face detection rate, when the driver was
looking straight, is found to be 98.5 % whereas the eye
detection rate was 97.5 % on an average for both day as well
as night. With head rotations of the driver, the face detection
rate dropped down to 95 % whereas that for eyes was 94 %.
The eye state classification was found to be 97 % accurate.
The overall false alarming of the algorithm was found to be
around 5 %. The PERCLOS values are shown in the display
Fig. 3 The overall vision based algorithm
Fig. 4 Schematic diagram of
the overall system
Int J Adv Eng Sci Appl Math (April–September 2013) 5(2–3):94–103 99
123
for every minute, on a running average window basis of
3 min. The main issues faced during the day driving test are
the image saturation, noise due to vehicle vibration and lights
coming from other vehicles during night.
4 Results
The system was found to be robust in detecting the face and
eyes. The face detection rate, when the driver was looking
straight, is found to be 98.2 % whereas the eye detection
rate was 97.1 %. With head rotations of the driver, the face
detection rate dropped down to 94.5 % whereas that for
eyes was 94.2 %. The eye state classification was found to
be 98 % accurate. The overall false alarming of the algo-
rithm was found to be around 5.5 %. The error in PER-
CLOS estimation was calculated by the difference of true
PERCLOS computed manually with the PERCLOS value
obtained from the algorithm. The error is found out to be
4.9 % approximately. The NIR lighting system works well
for the test duration. From the above on-board tests, it is
evident that this system can be used as a safety precaution
system which might prevent a number of road accidents
due to drowsiness.
Figure 7 shows the robustness of the system against
invariance of gender of the subject as it is found to detect
female faces and eyes. The figure also reveals that the
system is invariant to the height of the driver owing to the
positioning of the camera.
From the above on-board tests, it is evident that this sys-
tem can be used as a safety precaution system which might
prevent a number of road accidents due to drowsiness.
5 Validation
The validation of PERCLOS as an effective measure of
drowsiness level has been carried out using EEG signals
which is considered to be an authentic indicator for the
same. In this work, the EEG analysis primarily comprises
of the parameter based approach described in [35–37]. An
experiment was conducted on 13 subjects where profes-
sional drivers were used in the age-group of 20–40 years.
They were asked to drive on a simulated driving environ-
ment and video of facial images as well as raw EEG data
were captured simultaneously. The details of the experi-
ment is reported in an earlier publication [38].
Fig. 5 Process flow for estimating PERCLOS
Fig. 6 a Surge protector
circuit. b Camera on dashboard.
c NIR lighting system. d LCD
screen on dashboard
Table 3 Components with power consumption
Component Power (W)
Embedded processing unit 9
Webcam 1
NIR lighting system 2
LCD screen 1
Speaker 2
100 Int J Adv Eng Sci Appl Math (April–September 2013) 5(2–3):94–103
123
5.1 Preprocessing: artefact removal
Raw EEG data obtained in the experiment is contaminated
with artifacts. Such artifacts are characterized by ampli-
tudes in the milli-volt range (whereas the actual EEG is in
microvolt range) in the frequency band of 0–16 Hz [39].
The EEG signals were recorded at a frequency of 256 Hz
and the analysis was concentrated in the range of
0.4–30 Hz. Thus, the raw EEG was first processed using a
band pass filter with cutoff frequencies of 0.4 and 30 Hz.
After this, normalization is carried out to center the signal
at zero and force a unit standard deviation. Normalization
ensures removal of any unwanted bias that may have crept
into experimental recordings. The in-band artifacts were
then removed using wavelet thresholding.
5.2 Frequency decomposition
After artifact removal, the data was decomposed using a
4-level Daubechies-4 wavelet transformation. For wavelet
decomposition, suppose the coefficients are named corre-
sponding to D1–D4 and A4 as cD1, cD2, cD3, cD4 and
cA4 respectively. The coefficients cD1, cD2, cD3 and cD4
(clubbed with cA4) corresponds to the frequency bands b,
a, h and d (1 and 2) respectively.
It has already been discussed that sometime the basic
energy indices do not show a substantial change under mild
fatigued condition and suitable ratio parameters may be
better in differentiating such fatigue levels.
An increasing trend has been observed in entropy values
with fatigue at successive stages. A closer look at the graph
elucidates that the extensive entropies (SE and RERE) have
better performance than the non-extensive ones in classi-
fying the fatigue stages.
The figure shows the spatial variation of relative energy
in different bands of EEG signals as a comparison with the
PERCLOS value of a subject with increasing level of
fatigue (Fig. 8). The details are available in [38]. A stage
wise analysis of the entropy values brings out the physical
significance of this feature. As mentioned in energy and
ratio analysis, the first stage EEG signal is basically
Fig. 7 Face and eye detection
in female subjects
Fig. 8 Correlation of PERCLOS with EEG signals for 5 stages of
increasing fatigue
Int J Adv Eng Sci Appl Math (April–September 2013) 5(2–3):94–103 101
123
composed of low frequency d waves. Hence the entire
energy distribution is skewed towards the lower frequen-
cies. This represents a signal with lower disorder. There-
fore, entropy values are lowest in case of first stage for all
the entropies. As the fatigue level increases the energy in aand b bands starts to increase. This leads to the flattening of
the energy spectrum. Flattening of energy distribution
means greater disorder, which ensures higher entropy val-
ues at later stages.
The PERCLOS value was found to be coherent with the
EEG signal analysis as obtained from the results.
6 Limitations
The system will show poor results in case of drivers
wearing spectacles. The reason for this is the glare caused
in the spectacles resulting in loss of image information
about the state of the eye. There has been an attempt to
detect the presence of sunglasses in [40]. The case of
drivers wearing sunglasses will fail to estimate PERCLOS
because of lack on information regarding the eye state.
7 Conclusion
In this paper, a robust real-time system for monitoring the
loss of attention in automotive drivers has been presented.
In this approach, a face detection framework has been
presented based on a Haar classifier. The in-plane and off-
plane rotations of the driver’s face have been compensated
using affine and perspective transformations of the input
frame respectively. Matching of concatenated block LBP
feature is used to localize the eyes. Finally SVM is used to
classify the eye states into open and closed to compute the
PERCLOS values over a window of 3 min interval. The
overall speed of the algorithm is found to be 12 fps, which
is good enough for correctly estimating the state of eye.
The threshold of 15 % PERCLOS has been used for
declaring the driver as drowsy or not. A voice alarm is set
to warn the driver in the event of loss of attention. The
overall algorithms have been validated off-board using
EEG based method. On-board testing has been carried out
for both day as well as night driving. The system was found
to be quite robust both in terms of speed and accuracy.
There is further scope of research in the detection of eyes
occluded by spectacles.
Acknowledgments The funds received from the Department of
Electronics and Information Technology, Government of India, for
this study is gratefully acknowledged. The authors would like to thank
the subjects for voluntary participation in the experiment for creation
of the database.
References
1. Arun, S., Sundaraj, K., Murugappan, M.: Driver inattention
detection methods: a review. In 2012 IEEE Conference on Sus-
tainable Utilization and Development in Engineering and Tech-
nology (2012)
2. Matousek, M., Petersen, A.: A method for assessing alertness
fluctuations from EEG spectra. Electroencephalogr Clin. Neuro-
physiol. 55(1), 108–113 (1983)
3. Kar, S., Bhagat, M., Routray, A.: EEG signal analysis for the
assessment and quantification of driver’s fatigue. Transp. Res. 13,
297–306 (2010)
4. Routray, A., Kar, S., Bhagat, M.: EEG analysis for the assessment
and quantification of driver’s fatigue. Transp. Res. Part F Traffic
Psychol. Behav. 13(5), 297–306 (2010)
5. Cajochen, C., Zeitzer, J.M., Czeisler, C., Dijk, D.J.: Dose–
response relationship for light intensity and ocular and electro-
encephalographic correlates of human alertness. Behav. Brain
Res. 115, 75–83 (2000)
6. Knipling, R. R., Wierwille, W. W.: Vehicle-based drowsy driver
detection: current status and future prospects. IVHS America
fourth annual meeting, pp. 2–24, April 1994
7. Dhupati, L., Kar, S., Rajaguru, A., Routray, A.: A novel drows-
iness detection scheme based on speech analysis with validation
using simultaneous EEG recordings. 2010 IEEE Conference on
Automation Science and Engineering (CASE) (2010)
8. Bundele, M. M., Banerjee, R.: Detection of fatigue of vehicular
driver using skin conductance and oximetry pulse: a neural net-
work approach. In Proceedings of the 11th International Con-
ference on Information Integration and Web-based Applications
& Services (2009)
9. Onken, R.: DAISY, an adaptive, knowledge-based driver moni-
toring and warning system, In Proceedings of the Intelligent
Vehicles ‘94 Symposium (1994)
10. Singh, S., Papanikolopoulos, N.: Monitoring driver fatigue using
facial analysis techniques. In Proceedings of IEEE/IEEJ/JSAI
International Conference on Intelligent Transportation Systems
(1999)
11. Ayoob, E. M., Steinfeld, A., Grace, R.: Identification of an
‘‘appropriate’’ drowsy driver detection interface for commercial
vehicle operations. In Proceedings of The Human Factors and
Ergonomics Society Annual Meeting (2003)
12. Fletcher, L., Petersson, L., Zelinsky, A.: Driver assistance sys-
tems based on vision in and out of vehicles. In Proceedings IEEE
on Intelligent Vehicles Symposium (2003)
13. Smith, P., Lobo, N.V., Shah, M.: Determining driver visual
attention with one camera. IEEE Trans. Intell. Transp. Syst. 4(4),
205–218 (2003)
14. Zhu, Z., Ji, Q.: Real time and non-intrusive driver fatigue mon-
itoring. In IEEE Intelligent Transportation Systems Conference,
Washington, DC (2004)
15. Bergasa, L.M., Nuevo, J., Sotelo, M.A., Barea, R., Lopez, M.E.:
Real-time system for monitoring driver vigilance. IEEE Trans.
Intell. Transp. Syst. 7(1), 63–77 (2006)
16. Eriksson, M., Papanikolopoulos, N.P.: Driver fatigue: a vision-
based approach to automatic diagnosis. Transp. Res. Part C 9(6),
399–413 (2001)
17. Sacco, M., Farrugia, R.: Driver fatigue monitoring system using
support vector machines. In 5th International Symposium on
Communications Control and Signal Processing (ISCCSP)
(2012)
18. Dinges, D., Mallis, M. M., Maislin, G., Powell, J. W., and
National Highway Traffic Safety Administration: Evaluation of
techniques for ocular measurement as an index of fatigue and the
basis for alertness management. (1998)
102 Int J Adv Eng Sci Appl Math (April–September 2013) 5(2–3):94–103
123
19. Dingus, T.A., Hardee, H., Wierwile, W.W.: Development of
models for on-board detection of driver impairment. Accid. Anal.
Prev. 19(4), 271–283 (1987)
20. Viola, P., Jones, M.M.: Robust real-time face detection. Int.
J. Comput. Vis. 57(2), 137–154 (2004)
21. Gupta, S., Dasgupta, A., Routray, A.: Analysis of training of
parameters in Haar. In IEEE International Conference on Image
Information Processing, Shimla (2011)
22. Mohanty, P.K., Sarkar, S., Kasturi, R.: Designing affine trans-
formations based face recognition algorithms. In Proceedings of
the IEEE Computer Society Conference on Computer Vision and
Pattern Recognition (2005)
23. Lien, J.J., Kanade, T., Cohn, J.F., Li, C.: Detection, tracking, and
classification of action units in facial expression. Robotics Auton.
Syst. 31(3), 131–146 (2000)
24. Asteriadis, S., Nikolaidis, N., Hajdu, A., Pitas, I.: An eye
detection algorithm using pixel to edge information. In ICSSP
(2006)
25. Huang, J., Wechsler, H.: Eye detection using optimal wavelet
packets and radial basis functions (RBFs). Int. J. Pattern Rec-
ognit. Artif. Intell. 13(7), 1–18 (1999)
26. Song, J., Chia, Z., Liub, J.: A robust eye detection method using
combined binary edge and intensity information. Patt. Recognit.
39, 1110–1125 (2006)
27. Zhou, Z., Geng, X.: Projection functions for eye detection. Pat-
tern Recognit. 37, 1049–1056 (2003)
28. Wang, P., Green, M. B, Ji, Q., Wayman, J.: Automatic eye
detection and its validation. In Proceedings of the 2005 IEEE
Computer Society Conference on Computer Vision and Pattern
Recognition (2005)
29. Hadid, A., Heikkila, J., Silven, O., Pietikainen, M.: Face and eye
detection for person authentication in mobile phones. In First
ACM/IEEE International Conference on Distributed Smart
Cameras (ICDSC ‘07) (2007)
30. Liu, X., Xu, F., Fujimura, K.: Real-time eye detection and
tracking for driver observation under various light conditions. In
IEEE Intelligent Vehicle Symposium (2002)
31. Bhowmick, B., Kumar, K. S. C.: Detection and classification of
eye state in IR camera for driver drowsiness identification. In
IEEE International Conference on Signal and Image Processing
Applications (2009)
32. Eriksson, M., Papanikolopoulos, N. P.: Eye-tracking for detection
of driver fatigue. In Proceedings of the International Conference
on Intelligent Transportation Systems, Boston (1998)
33. Sirohey, S.A., Rosenfeld, A.: Eye detection in a face image using
linear and nonlinear filters. Pattern Recognit. 34(7), 1367–1391
(2001)
34. Ahonen, T., Hadid, A., Pietikainen, M.: Face description with
local binary patterns: application to face recognition. IEEE Trans.
Pattern Anal. Mach. Intell. 28(12), 2037–2041 (2006)
35. Jap, B.T., Lal, S., Fischer, P., Bekiaris, E.: Using EEG spectral
components to assess algorithms for detecting fatigue. Expert
Syst. Appl. 36(2), 2352–2359 (2009)
36. Papadelis, C., Kourtidou-Papadelis, C., Panagiotis, D., Bamidis,
D., Chouvarda, I., Koufogiannis, D.: Indicators of sleepiness in an
ambulatory EEG study of night driving. In 28th IEEE EMBS
Annual International Conference, New York, Aug 30–Sep 3,
(2006)
37. Papadelis, C., Chen, Z., Kourtidou-Papadelis, C., Bamidis, D.,
Chouvarda, I., Koufogiannis, D.: Monitoring sleepiness with on-
board electrophysiological recordings for preventing sleep
deprived traffic accidents. Clin. Neurophysiol. 118(9), 1906–1922
(2007)
38. Kar, S., Routray, A.: Analysis of electroencephalograph signals
for detecting fatigue in human drivers, PhD Dissertation (2012)
39. Nunej, P., Srinivasan, R.: Electric fields of the brain: the neuro-
physics of EEG, 2nd edn. Oxford University Press, New York
(2006)
40. Singvi, M., Dasgupta, A., Routray, A.: A real time algorithm for
detection of spectacles leading to eye detection. In 4th Interna-
tional Conference on Intelligent Human Computer Interaction
(IHCI) (2012)
Int J Adv Eng Sci Appl Math (April–September 2013) 5(2–3):94–103 103
123