Yawning Recognition based on

Dynamic Analysis and Simple Measure

Michał Ochocki and Dariusz Sawicki

Warsaw University of Technology, Institute of Theory of Electrical Engineering,

Measurements and Information Systems, Warsaw, Poland

Keywords: Worker Fatigue, Yawning Detection, Facial Landmarks.

Abstract: Nowadays drivers fatigue is amongst significant causes of traffic accidents. There exist many academic and

industrial publications, where fatigue detection is presented. Yawning is one of the most detectable and

indicative symptoms in such situation. However, yawning identification approaches which have been

developed to date are limited by the fact that they detect a wide open mouth. And the detection of open

mouth can also mean talking, singing and smiling, what is not always a sign of fatigue. The research aims

was to investigate the different situations when the mouth is open and distinguish situation when really

yawning occurred. In this paper we use an algorithm for localization of the facial landmarks and we propose

a simple and effective system for yawning detection which is based on changes of mouth geometric

features. The accuracy of presented method was verified using 80 videos collected from three databases: we

have used 20 films of yawning expression, 30 films of smiling and 30 films with singing examples. The

experimental results show high accuracy of proposed method on the level of 93%. The obtained results have

been compared with the methods described in the literature – the achieved accuracy puts proposed method

among the best solutions of recent years.

1 INTRODUCTION

1.1 Motivation

Fatigue and sleep symptoms are today very

important factors for psychology of work and safety

determination. The National Sleep Foundation

(National Sleep Foundation, 2017) has been

conducting research in this field for many years.

Fatigue and falling asleep are not only a simple

work-related problem. In many situations; for many

professions they lead to accidents causing deaths

(Osh in figures, 2011). Driver is one of such

profession, where fatigue consequences could be

dangerous.

Today many types of cars are equipped with

fatigue sensors (Bosch at CES, 2017). Fatigue tests

are carried out using many different fields of driver

activity (Friedrichs and Bin Yang, 2010). The most

popular are eye movements, eye blinking (especially

frequency of closing state) and yawning. In some

systems physical activity is also analyzed. In many

systems the yawning analysis is associated

effectively with the eye tracking (Weiwei, et al.,

2010) or with eye blinking analysis (Rodzik and

Sawicki, 2015). It is physiologically natural. There

exist many publications, where the surveys of

techniques for fatigue recognition we can find

(Friedrichs and Bin Yang, 2010, Coetzerm and

Hanckem, 2009, Sigari, et al., 2014). The industrial

publications (Bosch at CES, 2017, Fatigue Risk

Assessment, 2017) can be also treated as interesting

reviews of discussed problem. These documents

show how quickly the technology of employee

security is being developed and implemented.

Many systems for fatigue recognition are based

on yawning detection. It is commonly believed that

identification of yawning is very simple. After all, it

is just detection of a wide open mouth. But people

open their mouths also while talking, laughing, and

especially while singing. We are not sure whether

considering a driver singing at work is an

exaggeration. But the question is important: does

every mouth opening for a long time means

yawning? How to distinguish yawning from singing

or laughing in simple way?

Ochocki M. and Sawicki D.

Yawning Recognition based on Dynamic Analysis and Simple Measure.

DOI: 10.5220/0006497701110117

In Proceedings of the International Conference on Computer-Human Interaction Research and Applications (CHIRA 2017), pages 111-117

ISBN: 978-989-758-267-7

1.2 The Aim of the Article

The aim of this study was to develop a simple and

effective system for yawning detection. We have

tried to analyze different situation when the mouth is

open (for different purposes) and distinguish

situation when yawning is really occurred. Many

advanced methods used to recognize yawning are

known from the literature of the subject. Our goal

was to try to find a simple geometric measure that

would give comparable results.

2 YAWNING DETECTION IN THE

PREVIOUS WORKS

In almost all systems, where the yawning is detected

as the fatigue symptom, the recognition methods are

built in similar way. The camera (one or sometimes

two) registers head/face image. Using image

analysis the mouth is recognized and its state (close

or open) is detected. In case of open state of mouth

we have to decide if it is yawning.

Alioua, Amine and Rziza in paper (Alioua, et al.,

2014) have applied support vector machine (SVM)

for face extraction and circular Hough transform

(CHT) for mouth detection. Then, using CHT they

identified yawning. The advantage of this solution is

that such system does not require any training data.

Wang and Shi in (Wang and Shi, 2005) have used

Kalman filter at the state of face region tracking. In

this algorithm mouth is extracted from proper

region. It is simple and effective solution.

Saradadevi and Bajaj in paper (Saradadevi and

Bajaj, 2008) used Support Vector Machine in

analysis of face/mouth image. After training SVM

allowed distinguishing between “Normal Mouth”

and “Yawning Mouth”. Ibrahim at al. in paper

(Ibrahim, et al., 2015) implemented local binary

pattern based on color differences. Algorithm

measures the mouth opening and identifies yawning.

This paper is the first one where Authors also

analyzed situation, when the yawning mouth is

covered by hand for a period of time. In the solution

presented in paper (Li, et al., 2009) two different

cameras have been used: Camera A records the

driver's head according to face position. Image from

Camera B is used for detection of driver’s mouth

using Haar-like features.

The Viola-Jones algorithm (Viola and Jones,

2001) is very popular in extraction of the face

region. It detects face correctly, when the face is

tilted up/down to about ±15 degrees and left/right to

about ±45 degrees (Azim et al., 2009). Authors of

paper (Abtahi, at al., 2013) have implemented the

Viola-Jones algorithm for face and mouth detection,

they have used the back projection theory for

measuring the rate of mouth changes in yawning

detection. This idea have been developed and

presented, by practically the same team of Authors,

in paper (Omidyeganeh et al., 2016). The region of

mouth is found by Viola-Jones algorithm. Then the

mouth region is segmented by k-means clustering

method. Manu in paper (Manu, 2016) presented

system where the Feature Vector containing states of

left eye, right eye and mouth was used in training

and classification. Papers (Omidyeganeh et al.,

2016) and (Manu, 2016) are the latest published

conference documents concerning the yawning

recognition, which we found during our work.

Yawning detection is not so simple task.

Especially if we try to compare yawning to singing,

smiling and other states when the mouth is open.

The simplest solution used in algorithms for fatigue

recognition is to analyze the height of opened

mouth. In the work (Kumar and Barwar, 2014) the

contour of open mouth is analyzed, Y coordinate

values greater than a certain threshold, informs

about yawning. The Authors of the paper (Alioua, et

al., 2014) propose a distinction between slightly

open mouth and widely open mouth using circular

Hough transform (CHT). Although very elegant, it is

a time consuming solution. Area analysis of the open

mouth can give good result. It is so convenient that

after detection of opening, we can mark the visible

interior of the mouth on the basis of the difference in

color. Such solution is proposed by the Authors of

the work (Omidyeganeh et al., 2016). After

highlighting the area of mouth, the Authors specify

the number of pixels in the image of mouth opening.

If it is greater than proper threshold, the yawning

state is detected. Similar method is used in (Li, et al.,

2009) but additionally Authors of this paper

analyzed the ratio between height and width of the

area recognized as the mouth opening.

3 PROPOSED ALGORITHM FOR

YAWNING RECOGNITION

In frequently used methods to identify yawning (and

blinking as well), this phenomenon is analyzed

"holistically" – taking into account the features of

the entire face. It is important in this situation to

accurately distinguish the relevant elements (mouth

or eyes and mouth) in the properly identified face.

Hence, in such algorithms there is an independent

condition that the face is recognized in a

"reasonable" way (Li, et al., 2009).

A typical problem in this task is to analyze the

image of face that is tilted and / or rotated. The

Author of paper (Manu, 2016) pointed out that his

system does not always work properly, when head

was tilted left or right. Authors often pay attention to

this problem by proposing additional algorithms for

the normalization of the position associated with the

"compensation" of deviation from the horizontal line

(Wang and Shi, 2005, Hasan et al., 2014). The

Authors of the paper (Wang and Pendeli, 2005)

examine the degree of mouth openness by

considering the trigonometric function (cos)

dependent on the rotation of the mouth. Similarly,

the height of the mouth opening and the angle of

deviation from the horizontal line are taken into

account in (Azim et al., 2009).

In order to separate the problems of analysis as

well as to improve efficiency and accuracy of

yawning identification, we propose a two-stage

recognition algorithm.

 Facial Identification and Separation of the

Mouth Shape. The aim of this stage is to

obtain the shape of the mouth from an image

of head. In addition, the resulting shape of the

mouth is normalized regardless of lighting

conditions, and regardless of the head position

(tilting up - down and left - right). But of

course, in a certain range of correct work, i.e.

when the entire mouth is recorded despite the

tilt.

 Yawning Detection. The aim of this stage is

to identify yawning with particular regard to

the differences for other states of mouth

opening.

3.1 Facial Identification and

Separation of the Mouth Shape

In the face recognition and analysis an algorithm for

localization of the facial landmarks was applied

(Ochocki and Sawicki, 2016). This algorithm was

originally prepared for emotion recognition,

however can be used for any analysis of the face

geometry. On the other hand, the original aim

(landmarks detection) guarantees a high recognition

accuracy of face features. In the algorithm

preliminary analysis has been used. In this part the

set of proper regions of interest (ROI) is identified.

It is based, mainly, on analysis of features in color

space which corresponds to each region.

In order to separate the mouth area the properties

of CIE-Lab space is used. CIE-Lab is a standardized

color space Lab by the International Commission on

Illumination. Lab color model is designed to

illustrate the difference between the colors seen

through the human eye. The color is represented by

three coordinates L, a and b. Component L

represents the luminance, a determines the color

from green to magenta, b color from blue to yellow.

In the algorithm the a and b components of Lab

color model is used for mouth segmentation.

The whole algorithm can locate basic 21

landmarks of the face (Figure 1) with a very good

average accuracy of 95.62% (Ochocki and Sawicki,

2016). This is comparable to the results of the best

algorithms of this type, which were published

between 2005 and 2015.

Figure 1: The facial landmarks analyzed in the algorithm

for localization. The free face chart from (Want a free mua

face chart? 2010).

In the used algorithm, landmarks identification is

divided into four independent groups - the facial

region (oval), the eye region, the mouth region and

the region of the nose. For the purposes of yawning

analysis, we modified the algorithm so that it only

identifies the landmarks of the mouth. The modified

algorithm achieved an average accuracy of 95.2%.

The result of the algorithm's operation in the form of

a normalized image of the mouth is shown in

Figure 2.

Figure 2: Landmarks of the mouth identified by the

modified algorithm for face analysis.

3.2 Yawning Detection

To develop an effective detector, the set of analyzed

features should be extracted. For this purpose, the

selection of features was performed, i.e. the process

of identifying the appropriate set of geometric

changes in the shape of mouth.

In the yawning detection algorithm, we use the

coordinates of the landmarks obtained from facial

analysis (Figure 2.). We have introduced D

(1) as a

simple measure of the mouth opening. The value of

the D

measure is the ratio of the height of mouth to

their width in a given frame of video, determined

using coordinates of landmarks. The width of mouth

is calculated as a simple distance between proper X

coordinates (points 3 and 4). Height is calculated as

a simple distance between proper Y coordinates

(points 1 and 2).













(1)

where:

– the measure of opening, Y

– the height of

mouth (distance between points 1 and 2 – in

Figure 2.), X

– the width of mouth (distance

between points 3 and 4 – in Figure 2.).

The graph of the dynamic changes of the D

value in normalized time values by the min-max

method is presented in Figure 3. This is a graph for

the three cases of opening the mouth. The X axis

indicates the normalized elapsed time from the

beginning (0) to the end (1) of action (facial event).

This approach allows comparing actions of different

(real) duration of times.

From the analysis of the graph shown in

Figure 3, it can be concluded that the cases of

opening mouths can be easily distinguished on the

basis of the maximum D

value. Therefore, at the

stage of classification, feature d expressed by

formula (2) was extracted from the sets of measure's

changes D

=

max(



)

min(



)

(2)

where:

d – the dynamic measure of opening, D

– the

measure of opening.

Figure 3: An example of a graph which represents change

of D

measure for the three different discussed types of

opening the mouth.

4 TESTS OF THE PROPOSED

ALGORITHM

We have conducted experiments using three

independent databases:

 Smiling. In the study of smile expressions we

have used UvA-NEMO Smile Database

(Dibeklioglu et al., 2012). This database was

created to analyze the dynamics of genuine

(sincere) and false (posed) smile as part of the

Live Science program. UvA-NEMO Smile

Database consists of 1240 color videos at

resolution of 1920x1080px in controlled,

unified lighting environment. It contains 597

videos of genuine (sincere) smile and 643

videos of false (posed) smile, recorded from

400 people (185 women and 215 men).

Genuine and posed smile of each subject was

selected from approximately five minutes of

recordings There were people of different

races, genders and ages. Only the videos of

genuine smile were used in our research.

 Yawning. A colorful video from YouTube

(The Yawn-O-Meter, 2015) at resolution of

1280x720px was used to analyze the yawning

phenomenon. This videos contains

expressions of yawning recorded in unified

lighting conditions. The videos was divided

into fragments - each fragment contains

yawning expression of another person. The

database contains 20 yawning videos from

people of different races, genders and ages.

 Singing. We have used the Ryerson Audio-

Visual Database of Emotional Speech and

Song (RAVDESS) (Livingstone, 2012).

RAVDESS was created to analyze singing and

speaking in 8 different emotions – neutral,

calm, happy, sad, angry, and anxious. The

database contains more than 7,000 color

videos from 24 actors (12 women and 12

men).

Changes of the value of d parameter depending on

discussed cases are presented in Figure 4 in the form

of boxplot graph. This approach allows determining

the characteristic of the d distribution for each case.

The analysis was carried out using:

 20 films of yawning expression;

 30 films of smiling expression;

 30 films with singing examples.

The graph of d distribution contains three areas

of high separability. Therefore, this attribute has

been subjected to a final assessment of suitability by

maximization of the classification efficiency. For

this purpose, a simple classification method was

chosen. The distinguishing feature d was compared

with threshold value based on the data collected

during the training phase.

Figure 4: The graph of d distribution for three discussed

actions.

For the optimal classification threshold of 0.54

(Figure 4), all samples with expressing yawning (20

out of 20) were correctly classified. Three of 30

smile samples and two of 30 singing samples were

also classified as yawning. In this way, the overall

result of the classification was 93.75%.

We have compared the result achieved in our

solution with the results known from the publication.

The set of results is presented in Table 1. It is worth

emphasizing the fact that we used the typical,

standard video databases. Databases that are

Table 1: Results of tests – comparison to the previous works.

Author(s)

Number of

videos/

participants

Accuracy

1. Saradadevi, M. and Bajaj, P.

(Saradadevi and Bajaj, 2008) 337 81% - 86%

2. Li, L., Chen, Y., Li, Z.

(Li, et al., 2009) 4 95%

3. Abtahi, S., Shirmohammadi, S., Hariri, B., Laroche, D., Martel ,L.

(Abtahi, et al., 2013) 20 60%

4. Alioua, N., Amine, A., Rziza, M.

(Alioua, et al., 2014) 6 98%

5. Hasan, M., Hossain, F., Thakur, J.M., Podder, P.

(Hasan, et al., 2014) 150 95%

6. Omidyeganeh, M., Shirmohammadi, S., Abtahi, S., Khurshid, A.,

Farhan, M., Scharcanski, J., Hariri, B., Laroche, D., Martel, L.

(Omidyeganeh, at al., 2016) 107 65% - 75%

7. Manu, B.N.

(Manu, 2016) 4 94%

Proposed method in this paper

80 93.75%

commonly used to test the facial image analysis.

Because we considered three independent cases

(yawning, smiling, singing), so we used three similar

(numerically comparable) data groups - 20 videos of

yawning, 30 videos of smiling and 30 videos of

singing. Perhaps 80 videos are not the basis for

generalizing the results. But we can show the

published methods that are tested on 4, 6 or 10 cases

(videos or participants) only (see Table 1). In this

context, our experiments confirm the correctness of

the proposed method.

By analyzing of previous publications, it can be

concluded that high accuracy (94% or higher) was

achieved primarily for experiments in very small

groups: 4 - 6 videos / participants. If the groups were

larger (20 videos / participants or more) then the

accuracy was on the level of 60% - 86%. The paper

(Hasan, et al., 2014) is an exception: 150 tests were

conducted to give 95% as the value of accuracy.

However, in this solution two independent cameras

and advanced, complicated algorithm have been

used. In this context it can be concluded that our

method achieves very good results, which exceeded

the known solutions.

5 SUMMARY

The yawning detection algorithm has been presented

in the paper. The algorithm recognizes the yawning

and distinguishes it from singing and smiling. The

introduced method is based on a two-step

recognition procedure. In the first step there is

segmentation of the mouth area in order to obtain 6

facial landmarks which describing the shape of the

mouth. The purpose of the second stage is to detect

yawning with particular regard to differences for

other states of mouth opening.

The proposed solution simplifies the detection

procedure of yawning considerably. The effect of

the first stage is the normalized image of the mouth.

So there is neither problem with geometric

distortions nor a deviation from the horizontal line,

which have always been a problem in solutions

known from publications. On the other hand, the use

of ROI and facial landmarks gives the possibility of

proposing a simple measure of the mouth opening.

The classification of mouth stage (and thus yawning

detection) by analyzing the value of proposed

measure is more effective than calculating the sum

of pixels of the mouth opening - often used in

known solutions. On the other hand, the size of the

mouth is a strictly individual feature and additionally

depends on videos resolution. Thus, there may be

problems in yawning identification based on the size

of the opened mouth. This problem does not occur at

all if we use facial landmarks and our measure.

The aim of the work has been achieved. We have

managed to propose a geometric measure that allows

distinguishing yawning from other mouth openings

(smiles and singing). This measure takes into

account the dynamics of changes in the shape of the

mouth and in addition is simple and effective.

The introduced method has given very good

results. This has been confirmed experimentally.

Effectiveness of the presented algorithm was

verified using 80 videos collected from three

independent databases. The developed method

achieved value of accuracy on the level of 93.75%.

Comparison of the obtained results with the methods

described in the literature has been also carried out.

The achieved accuracy of 93.75% for 80 video

samples puts proposed method among the best

solutions of recent years.

ACKNOWLEDGEMENTS

This paper has been based on the results of a

research task carried out within the scope of the

fourth stage of the National Programme

"Improvement of safety and working conditions"

partly supported in 2017–2019 --- within the scope

of research and development --- by the Ministry of

Labour and Social Policy. The Central Institute for

Labour Protection -- National Research Institute is

the Programme's main coordinator.

REFERENCES

Abtahi, S., Shirmohammadi, S., Hariri, B., Laroche, D.,

Martel, L. 2013. A yawning measurement method

using embedded smart cameras. In: Proc. of 2013

IEEE International Instrumentation and Measurement

Technology Conference. 1605-1608. doi: 10.1109/

I2MTC.2013.6555685

Alioua, N., Amine, A., Rziza, M. 2014. Driver's Fatigue

Detection Based on Yawning Extraction. International

Journal of Vehicular Technology. Vol. 2014, Article

ID 678786, 7 pages. doi: 10.1155/2014/678786

Azim, T., Jaffar, M.A., Ramzan, M., Mirza, A.M. 2009.

Automatic Fatigue Detection of Drivers through

Yawning Analysis. In: Ślęzak D., Pal S.K., Kang BH.,

Gu J., Kuroda H., Kim T. (eds) Signal Processing,

Image Processing and Pattern Recognition.

Communications in Computer and Information

Science. Vol 61. Springer. doi: 10.1007/978-3-642-

10546-3_16

Bosch at CES 2017: New concept car with Driver

Monitor. http://www.bosch-presse.de/pressportal/de/

en/bosch-at-ces-2017-new-concept-car-with-driver-

monitor-84050.html (retrieved March 31, 2017)

Coetzerm R.C., Hanckem G.P. 2009. Driver Fatigue

Detection: A Survey. In: Proc. of AFRICON '09

Nairobi, Kenya, 1-6, doi: 10.1109/AFRCON.2009.

5308101

Dibeklioglu H., Salah A.A., Gevers T. 2012. Are You

Really Smiling at Me? Spontaneous versus Posed

Enjoyment Smiles, In: Proc. European Conference on

Computer Vision (ECCV 2012). 525-538..

Fatigue Risk Assessment. Gain visibility to hidden risk

factors to see the full impact of fatigue and distraction

on your operations. CATERPILAR (2017)

http://www.cat.com/en_US/support/operations/frms/fr

a.html (retrieved April 10, 2017)

Friedrichs, F., Bin Yang 2010. Camera-based Drowsiness

Reference for Driver State Classification under Real

Driving Conditions. In: Proc. of 2010 IEEE Intelligent

Vehicles Symposium, San Diego USA, 101-106. doi:

10.1109/IVS.2010.5548039

Hasan, M., Hossain, F., Thakur, J.M., Podder, P. 2014.

Driver Fatigue Recognition using Skin Color

Modeling. International Journal of Computer

Applications, 97 (16), 34-40.

Ibrahim, M.M., Soraghan, J.J., Petropoulakis, L., Di

Caterina, G. 2015. Yawn analysis with mouth

occlusion detection. Biomedical Signal Processing

and Control. 18, 360-369. doi: 10.1016/

j.bspc.2015.02.006

Kumar, N., Barwar, N.C 2014. Detection of Eye Blinking

and Yawning for Monitoring Driver’s Drowsiness in

Real Time. International Journal of Application or

Innovation in Engineering & Management (IJAIEM).

3 (11), 291-298.

Li, L., Chen, Y., Li, Z. 2009. Yawning detection for

monitoring driver fatigue based on two cameras. In:

Proc. of 12th International IEEE Conference on

Intelligent Transportation Systems, 1-6. doi:

10.1109/ITSC.2009.5309841

Livingstone, S.R., Peck, K., Russo, F.A. 2012.

RAVDESS: The Ryerson Audio-Visual Database of

Emotional Speech and Song. Paper presented at: 22nd

Annual Meeting of the Canadian Society for Brain,

Behaviour and Cognitive Science (CSBBCS),

Kingston, ON

Manu B.N. 2016. Facial features monitoring for real time

drowsiness detection. In: Proc. of 12th International

Conference on Innovations in Information Technology

(IIT), doi: 10.1109/INNOVATIONS.2016.7880030

National Sleep Foundation. http://sleepfoundation.org/

(retrieved April 11, 2017)

Ochocki, M., Sawicki, D. 2016. Facial landmarks

localization using binary pattern analysis, Przeglad

Elektrotechniczny, 92 (11), 43-46. doi:

10.15199/48.2016.11.11

Omidyeganeh, M., Shirmohammadi, S., Abtahi, S.,

Khurshid, A., Farhan, M., Scharcanski, J., Hariri, B.,

Laroche, D., Martel, L. 2016. Yawning Detection

Using Embedded Smart Cameras. IEEE Transactions

on Instrumentation and Measurement. 65 (3), 570-

582. doi: 10.1109/TIM.2015.2507378

Osh in figures 2011: Annex to Report: Occupational

Safety and Health in the Road Transport sector: An

Overview. National Report: Finland. In: Report of the

European Agency for Safety and Health at Work.

Luxembourg: Publications Office of the European

Communities.

Rodzik, K., Sawicki, D. 2015. Recognition of the human

fatigue based on the ICAAM algorithm. In: Proc. of

the 18th International Conference on Image Analysis

and Processing (ICIAP 2015). LNCS 9280 Springer.

373-382. doi: 10.1007/978-3-319-23234-8_35

Saradadevi, M., Bajaj, P. 2008. Driver Fatigue Detection

Using Mouth and Yawning Analysis. International

Journal of Computer Science and Network Security, 8

(6), 183-188.

Sigari, M-H., Pourshahabi, M-R., Soryani, M., Fathy, M.

2014. A Review on Driver Face Monitoring Systems

for Fatigue and Distraction Detection. International

Journal of Advanced Science and Technology, 64, 73-

100. doi: 10.14257/ijast.2014.64.07

The Yawn-O-Meter (How Long Can You Last?)

https://www.youtube.com/watch?v=AJXX4vF6Zh0

(retrieved February 30, 2015)

Viola, P., Jones, M. 2001. Rapid Object Detection using a

Boosted Cascade of Simple Features. Computer Vision

and Pattern Recognition (CVPR). 1, 511-518. doi:

10.1109/CVPR.2001.990517

Wang, T., Pendeli, S. 2005. Yawning detection for

determining driver drowsiness. In: Proc of IEEE Int.

Workshop VLSI Design & Video Tech. 373-376. doi:

10.1109/IWVDVT.2005.1504628

Wang, T., Shi, P. 2005. Yawning detection for

determining driver drowsiness. In: Proc. of 2005 IEEE

International Workshop on VLSI Design and Video

Technology, 373-376. doi: 10.1109/IWVDVT.2005.

1504628

Want a free mua face chart? 2010.

http://smashinbeauty.com/want-a-free-mua-face-chart/

Retrieved 25 January 2013

Weiwei Liu. Haixin Sun, Weijie Shen 2010. Driver

Fatigue Detection through Pupil Detection and

Yawning Analysis. In: Proc. of 2010 International

Conference on Bioinformatics and Biomedical

Technology (ICBBT), Chengdu China, 404-407. doi:

10.1109/ICBBT.2010.5478931