HandicappedPeopleVirtualKeyboardControlledbyHeadMotion

Detection
Ondrej Krejcar
Department of Information Technologies,
Faculty of Informatics and Management,
University of Hradec Kralove,
Hradec Kralove, Czech Republic
[email protected]
Abstract Nowadays, computers and modern technologies are
available to almost all people. However, we should not forget
about disabled or handicapped customers. Those people should
have access to a PC as well as to the Internet. This project
implements the boundary for people with movement disability.
Although, the basic assumption is that they can at least move their
heads and eyelids. These parts of the body are monitored by web
cameras and with their help the user moves around the virtual
keyboard shown on the screen. In this way the user is able to
utilize their own PC. Moreover, through winking an eye he/she can
virtually press a button and write the required symbol. The whole
software for detection system was established in a development
environment of the Visual Studio 2010 from Microsoft company.
The main advantage of our solution is a cheaper price in compare
to any existing solutions. By this fact the solution is possible to use
even in third countries like in Africa continent.
Keywords HCI; face detection; users interface; handicapped
people; Open CV libraries
I. INTRODUCTION
As mentioned in the abstract, the developed software is
aimed at operating a PC using movements of the head and
eyelids. In order to detect head motion using web camera, it
is necessary to implement face detection of a particular user
and to determine the direction of its movement. The face
detection is basic and important technology from the range
of scanning systems which are used as a boundary between a
PC and a person [1]. A large number of authors have already
dealt with such problems. They have published many books
and developed considerable numbers of algorithms and
methods for both object detection in production processes
and face detection. In recent years, face detection algorithms
gave rise to extremely precise methods for face localization
in images [2].
The main issue of each face detection system is a camera
which contains a CCD (Charge-Coupled Device) sensor.
With ample speed this sensor generates images of space in
front of the camera. These images are then saved by the
camera onto a hard disk in video file format. It is mostly the
format of AVI (Audio Video Interleave) [3]. The detection
system then searches according to the chosen algorithm for a
face in the images of the video, eventually in its segments
and identifies their position based on the number of pixels
from the left margin and the number of pixels from the upper
margin. The programmed softwares objective is to contrast
the offset of detected parts in comparison with former
images. According to the offset of parts, it enables the
movement on a virtual keyboard. The motion of eyelids is
then analyzed as the pressing of a certain button on the
keyboard. A person usually blinks with both eyes at once,
which is the reason why the program cannot respond to this
movement.
Therefore it is clear that two components have the most
significant impact on the systems quality. These are the
apparatuses for face scanning of a user and the algorithm for
face detection from a scanned image. It is very important to
know the value of resolution in a web camera. If using a
camera with small resolution, the users motion does not
have to be detected with required certainty. On the other
hand, if using a high-quality camera with high resolution and
speed of image scanning, it might take longer for older
computers to implement detection algorithm and face
location.
Perhaps the most complicated issue of detection system
concerns choosing the correct algorithm for face detection.
First of all, it is necessary to consider the uncountable
number of variations in human appearance. People differ
from one another not only through the shapes and sizes of
individual body parts, but also through skin pigmentation,
eye color and hair and eyebrow lengths. A woman has an
entirely differently shaped face than an unshaved man etc.
Another disturbing impact might be caused by a movement
behind a user, i.e. a sudden change of light intensity in a
monitored scene, or the entry of another person to the
scanned zone. All of these and many other obstacles are
supposed to be eliminated by this system in order to
accomplish the highest levels of quality and reliability. On
the other hand, there are also some advantages such as the
homogeneity of faces in comparison with other categories of
detected objects. All of these share a certain canonical
structure, including two eyes, a nose and a mouth. These are
symmetrically placed with regard to longitudinal axis of oval
contour (although not all of the components are traceable
from all visual angles) [4].
It is suitable to use face static components for detection
of head movement. These components represent the motion
of a whole face (i.e. nose, birthmarks).
II. PROBLEM DEFINITION
In recent times, modern technologies and computers
became essential parts of our lives. They are present in our
homes, schools, offices etc. Access to the internet and
working on computers became part of a routine for ordinary
people. However, how do you access the internet or send
emails to friends living on different continents, when you
cannot move your hands and fingers? Fortunately, a large
number of experts started to be concerned about this issue
and developed many alternative boundaries for the operating
of computers even without a mouse and a keyboard.
The first very interesting system was proposed and
designed in the Czech Republic. It is called the i4control.
The system i4control is the new type of computer
periphery which enables disabled immobile users to operate
personal computers by eye motion (eventually by head
movement). This system substitutes an ordinary mouse and
offers to its user a simple way to communicate with installed
applications by eye movement [Fig. 1] [7].
Figure 1. i4control User Interface.
Another product which uses a similar system is called
LifeTool IntegraMouse. Basically it is a mouse controlled by
lips motion and the pressing of a button is accomplished by
sipping or blowing. This system is ideal for users with
various physical problems such as whole body paralysis,
after amputation, muscular dystrophy or disseminated
sclerosis [Fig. 2] [8].
Figure 2. IntegraMouse User Interface.
The system My Voice is another very interesting option.
This program enables users to operate a computer and install
programs exclusively by vocal commands. Using these
commands, users are able to carry out the same actions as
would be otherwise realized by mouse and keyboard. The
program is capable of immediately recognizing vocal
commands from random person and therefore, it is not
necessary for the user to narrate anything before usage.
Figure 3. SmartNav User Interface.
Finally, there is a product SmartNav from NaturalPoint
company. SmartNav uses infrared (IR) camera for
monitoring users head motions [Fig. 3] [10]. The system is
based on scanning the users head using infrared LED.
Special reflex facets are placed on the users head (glasses)
and then the radiation is reflected back to a camera which
analyses its motion.
All described solutions are very expensive [Table 1].
This fact was one of motivation to develop new solution as a
low cost.
TABLE I
EXISTING USER INTERFACE SOLUTIONS FOR HANDICAPPED USERS
Solution name Price [USD] Principle
i4control 2200 Eyes movement
IntegraMouse 3400 Lips motion
My Voice 650 Voice recognition
SmartNav 400-500 Head motion
Our Solution free Head and eyes motion
As mentioned in the introduction, in this project a virtual
camera that scans users head motion was used in order to
enable the user to move over a virtual keyboard. However,
this solution method gives rise to certain problems mainly
during head detection and its individual parts. Another
person might approach the visual field of the camera which
might lead to a fault in a detection algorithm. Moreover, it is
necessary to consider the distance in which the users face
will be scanned. In order for the system to be cheap and
available to everyone, it is also necessary that the system is
not composed of parts whose purchase price is equal to
hundreds of USD. This is an example of an HD camera.
Therefore it is necessary to use ordinary web cameras that
are found in most of the notebooks.
These cameras usually have a resolution of around 1.3
megapixels. This means that the scanned image has
proportions of 1280x1024 pixels. The image is often
programmatically adjusted into a resolution of 640x480
pixels in the case of video scanning in order to fasten the
transfer of information on the internet. In this case it is
necessary to consider another restricting factor the distance
of a user from camera. An ordinary person is capable to
move his/her head in the range of 20 cm. If the person is in
distance of 60 cm from the camera, the head movements are
going to be scanned in a segment of 150x150 pixels. These
values are just approximate, because there are no two web
cameras from various producers that would be the same.
Furthermore, every one of them has a different defined angle
in which its lens is able to scan an image and carry it on
CCD sensor. The recognition of head movements in a range
of 140x140 pixels is very impractical for mouse movement
on a working surface with a resolution of 1280x1024 pixels.
A mouse pointer would be likely to move by jumps over the
screen which would in turn be caused by slight unwanted
head movements of a user.
The mouse movement would be then possible to realize
using virtual touchpad with the function which is offered by
every notebook. An especially interesting solution might be
designed by using accelerometer to scan head motion.
However, that is not the objective of this project.
The alternative ways of computer operating which are
described in the first part of this chapter are very
sophisticating. However, every one of these solutions brings
certain disadvantages. Firstly, for using the i4control system
a user needs to wear glasses with a camera that is placed a
couple of centimeters away from the users eyes. This might
be a very disturbing factor for some people. Moreover, a data
cable is needed for connection of the glasses to the computer.
This cable might also get attached to some object causing the
glasses to fall. A similar situation may happen with
IntegraMouse. Another system, My Voice, can be difficult
for usage outside of the home because of disturbing factors
such as television or radio in an office or other public place.
The last system, SmartNav, which is similar to the solution
proposed in this project, might be slow for the user because
of the necessity of special head movement in order to press a
button. Therefore a great emphasis in this project was put on
the utilization of all systems described advantages and their
interconnection. This leads to highest possible level of
comfort for the user.
III. NEW SOLUTION
This chapter might be imaginarily divided into two parts.
In the first part a necessary theory about programmed
detection of face and other parts is given, including
information about basic functions of developed programs.
The latter part of the chapter introduces the structure of the
program itself with description of its individual parts and
linkage.
A. Face Detection
A detection of a face and both eyes is used for keyboard
operating. The face movement represents the movement
across the keyboard. Moreover, a reasonably long wink
using both eyes gives the system order for pressing a button.
Open CV library from Intel company is used in this project
in order to launch the camera and detect individual parts.
Open CV is a free and open multiplatform library designed
for manipulation with image. It is aimed mainly at computer
vision and real-time image processing [11]. The CV is an
abbreviation which stands for Computer Vision. However, it
only works in C/C++ setting which constitutes a
disadvantage. Therefore, this project also uses Emgu CV
system [12] which is a cross platform for .NET that works as
a cover for Open CV libraries indented for image processing.
Moreover, it enables CV libraries to use a function for
languages compatible with .NET such as C#, VB, VC++,
IronPython etc. [12]. Emgu CV then allows in just few
commands image scan by web camera and detection of
individual objects on created image. In order to obtain the
highest possible speed of the whole application the scanned
image is converted into black and white color. The detection
itself is then undertaken by DetectHaarCascade command
which is based on a high-speed detection algorithm called
Viola & Jones (it is the first boundary for detection of
objects that carries out real-time detection; it was created in
2001 by Paul Viola and Michael Jones [13]) and another
algorithm called Adaboost. Adaboost constitutes an
abbreviation for Adaptive Boosting. It is an algorithm of
machine learning established by Yoav Freundem a Robert
Schapirem. Moreover, it is a meta-algorithm which might be
used in conjunction with other learning algorithms in order
to improve their performance [14].
The very first argument of a command
DetectHaarCascade is so-called Haarcascade (cascade haar
style) which is a file of .XML type saved in the same folder
as the used application. The detected object is parameterized
in this file.
The detection algorithm returns a face variable of VAR
data type. Then it is necessary to search only for the wanted
face using a command foreach and transfer it to a variable of
rectangle type in this way a rectangle which is represented
by coordinates of upper right corner and the size of its sides
is acquired. The testing of detection algorithm uncovered
that any insignificant zooming in or out of the users face
from the camera would cause unstable changes of the
rectangles size (or rather square, because both width and
height are consistent). For this reason, the upper left corner
coordinates of the square are inapplicable and the middle of
this square was used for the movement across a keyboard.
The middle coordinates do not considerably change
regarding to different sizes of human faces. These
coordinates might be calculated using the following formula:
; where W is the width of a square
According to a similar formula it is even possible to
calculate coordinates Y. In this way those coordinates which
are necessary for movement across the virtual keyboard are
obtained. It was discovered during another testing of the
detection algorithm that the best ratio between distance and
sensitivity of a control is within the 211x211 or 295x295
pixels of detected square. (The size of square changes by
jumps.) The distance of users face from the camera in these
circumstances was maximally 65cm. If this distance
becomes greater, it would not be possible to operate the
keyboard just by slight turn of a head, because the face
detection itself would fail due to the face being turned too
much. Moreover, it would lead to a failure of detection of
eyes. In the case when the detected face had size of 211x211
pixels, users head could be turned by the range of 140x140
pixels.
2
W
X X
Corner Middle

B. Application Structure
The proposed virtual keyboard [Fig. 4] has 7 rows and
maximally 14 buttons in one single line. Therefore, the
boundary was invented in a way to move to the next button if
there is a change by 10 pixels upward or in width. In this
way slight head motions and failures of detection algorithm
are eliminated.
Figure 4. User interface of developed application.
The pressing of a button might be represented in two
ways - eye winking or special mouth movement (i.e. opening
of a mouth). The first way was chosen for this project,
because mouth movement was discovered to cause frequent
failures of the system. The program analyzed a beard as a
part of a mouth of an experimental subject. In order to
remove this problem it would be necessary to modify or
create absolutely new cascade haar style.
Moreover, the detection would have to be done in a
colorful image which would significantly increase the time
necessary for accomplishment of detected algorithm and
reduce the speed of developed application. It was much more
convenient to use the scanning of eye motion and, in order to
press a button, to close both eyes for a reasonable period of
time.
This solution is also well-founded because during the
turning of the head, the eyes, especially in limit value, might
be covered by the nose. The system is then capable of
detecting only one of the eyes. However, at least one of the
eyes should be almost always recognizable by algorithm on
the acquired image from a web camera. If the algorithm is
unable to find either of the eyes for 0.5 1.5 s and then
detects at least one of them, it will evaluate it as the pressing
of a button. In this way it is possible to eliminate shorter
autonomous eye winking and also the cases in which eyes
would not be detected for longer periods of time (i.e. if a user
moves too far from the camera).
Figure 5. UML activity diagram of developed user interface for handicapped users.
The fundamental part of the application is formed by a
main class a keyboard - and by a supplementary class a
camera (next viz. Image 1). The class camera was rather used
during the studies of parameter transfer by methods among
different classes. Two functions for working with virtual
keyboard are placed within the class keyboard. These are the
methods detect in which the detecting algorithm is
implemented and timer1_Tick. The latter method is
excessively important, because it calls other methods and
assures that the whole application is running smoothly in a
certain loop. Firstly, it calls the method take_picture,
placed in the camera class, which returns images taken by a
web camera. The method forms a mirror image of the
previous one alongside an imaginary vertical axis. The
turnover is necessary, because the image on the screen would
not otherwise move in the same direction as the user.
Furthermore, the method timer1_Tick calls the method
detect whose entry parameter is an image taken by web
camera. This method is able to find users face and eyes by
using already mentioned libraries Emgu CV. The outcome is
again the given image, which is then modified through
having colorful rectangles drawn into it that border the
already found objects (the face and eyes). Then the image is
simply adjusted by timer1_Tick method to the size that
corresponds with the actual size of the picture_box into
which the image is subsequently drawn as a bit map.
However, as a result the detect method carries out more
tasks. At the beginning it is necessary to calibrate the head
position and the middle of the keyboard. Therefore the
detect method is used right at the start to draw a black
square in the middle of the image. Consequently, the user has
to place his face into this square. If this is done correctly, the
detect function overwrites the variable value ini_pom to 1
and the condition in the detect function which draws this
black square stops working. Conversely, the condition which
initiates the find_button method takes place. Moreover, it
changes the particular button on a keyboard to a red color
and draws a red square around the users head. The initiated
method find_button has the middle of a users face as its
input parameter and a button which corresponds to the
particular face positioning is its output. An algorithm for
searching in a two dimensional field for keys called Button
was implemented into this method. Those keys are situated
due to the position of the middle of a face in accordance to
the middle of a scanned image and also to the number of
keys in a particular keyboard row. Furthermore, certain
insensitivity for position changes of less than 10 pixels was
implemented in order to eliminate minor head motions and
detection algorithms faults (as mentioned before). The
function of algorithm and of the whole method is very hard
to describe. However, it is perfectly understandable from a
written code.
The matrix of Button keys which was mentioned before
was implemented at the beginning using the
initialization_key method that is called in key_load
method. Only two keys from the matrix, Enter and Space,
occupy more spaces. This is due to the fact that they take in
more columns or rows and the transfer to them must be
possible even from those locations.
If the timer interval of 1 ms is used, the speed of the
whole application should be dependent only on the speed of
a camera and Viola & Jones detection algorithm.
IV. IMPLEMENTATION
This chapter is concerned about the practical
implementation of the new solution.
The application was developed using the C# language of
.NET Framework in Visual Studio 2010 environment from
Microsoft Company. The whole graphic page of the
application was invented in GUI which is a library of
graphical users boundary. Apart from the usual elements as
Button, Picture box or Label, another two components were
used to run in the background. Those were Timer and
FolderBrowserDialog.
FolderBrowserDialog is a graphic component which
serves as an instrument for browsing folders in a computer
and after choosing one of the folders it reveals the route
towards it. This function was used in the initialization part of
the program. The whole initialization runs in Form1_Load
method. It is necessary within the program because the
application works with two external XML files, in which the
two above mentioned cascade styles for detection of face and
eyes are saved. Those files are called haarcascade_eye.XML
and haarcascade_frontalface_alt2.XML.
After the first initiation of the application, the program
firstly determines whether the ini_data.txt file is located in
the applications directory. If this file is missing, the
application creates it. Consequently, it initiates a system
message for a user in which it passes on the information
about necessary initialization of the program and the user is
asked to insert the path towards files *.XML with cascade
styles. The component FolderBrowserDialog is used for
inserting this path.
After the next running of the application the program
investigates again the presence of initializing file
ini_data.txt. If this file is found its content is uploaded and if
there are any files .XML with cascade styles detected in the
content, a form with virtual keyboard is produced. If the
initializing file ini_data.txt is detected, but it does not contain
valid data, it is deleted and the user is informed using
MessageBox that the application will be closed and it is
necessary to reopen it (after a new start-up the whole
situation will be running in the same way as during the first
initiation of the program).
Figure 6. Part of the application user interface. User face did not
recognized the black rectangle is showed.
When the main window of the whole application is
showed, user need to sit down in front of the web camera and
position his face into the black square [Fig. 6]. When
positioning user face into the middle of the image, the black
square should become red and begin to move together with
the detected face.
Figure 7. Part of the application user interface. User face is detected and
recognized the red rectangle is showed around green rectangles at eyes
positions.
After successful detection of user face, the position and
size of user face and also user eyes is being scanned. A red
colored active key moves around the keyboard together with
the face movement [Fig. 4].
The pressing of this key is done by closing of user eyes
for the period of 0.5 1.5s. This time depends mainly on the
speed of a used computer (processing of an image from the
web camera, implementation of detecting algorithm) and
also on the speed of the used web camera. The required letter
is written into the text field under the keyboard.
The buttons ENTER, BACKSPACE and SPACE have
special functions that were programmed in order to adjust the
text as it is standard with any other keyboard. The button
stop ends the application.
V. TESTING OF DEVELOPED APPLICATION
This chapter is concerned with testing of the developed
application. The results from this testing should be compared
to the results gained from the testing of other authors.
However, these other authors results were not founded and
their solutions are built upon completely different principles.
Therefore these projects could not be compared.
In order to somehow compare gained results there is an
application which was developed in order to operate a
personal computer using movement of face or eyes. This
application was run and tested on two individual computers.
The first computer works with the processor AMD
TURION X2 2GHz (Dual Core), 3GB RAM and it scans the
user by internal web camera with 1.3 MP resolution. The
second computer operates with processor AMD TURION
X2 1.6Hz (Dual Core), the main board contains 2GB RAM
and an external camera (of older manufacture date) with 0.3
MP resolution which is used for image scanning.
A. Testing Methodology
The testing was undertaken in variety of ways. The first
is based on writing a predetermined text. The sentence which
is necessary to be written is: APPLICATION TESTING.
While writing this sentence the time taken for the task to be
completed, and the number of faults committed by the user,
are recorded. In order to limit the influences of incidental
faults (processors working capacity, users fault, etc.) this
sentence is written down and its results are consequently 5x
mapped. The arithmetic average is then calculated from the
resultant values. The same procedure is applied even for the
second computer. Finally, this type of testing is undertaken
in two differently lighted environments.
The second style of testing is based on the number of
writing of the same symbol in the beforehand set time. At the
present time the number of faults which were made by the
user at the given time, and the number of correct key presses,
are recorded into tables. The whole testing is again
consequently undertaken 5x and then the average of these
values is generated in order to lower the influence of random
faults and the users faults during the testing. Similarly to the
previous case the testing is undertaken in two environments
with different light settings.
B. Testing Results
The results of testing using 1st technique are stated in the
table [Table 2].
TABLE II
TEST RESULTS OF WRITING OF DEFINED TEXT
The abbreviation HP identifies a notebook from Hewlett
Packard Company. The abbreviation FS describes a
notebook from Fujitsu-Siemens company. The results from a
testing by 2nd technique are stated in the table [Table 3].
TABLE III
TEST RESULTS OF WRITING PER DECLARED TIME
C. Results Discussion
As it is visible from the table [Table 2], the best results
were achieved by using the application of HP notebook when
there were good light conditions. This result is not
surprising. When there is a good light, the camera scans the
image significantly faster, because it does not have to adjust
to worse light conditions. The images given by the camera
are also sharper than when taking a picture with limited
lighting. During these conditions the error rate was raised
while writing a text. When considering the tables for
notebook FS it is obvious that the precise text writing was
very difficult. The main problem was that eyes were scanned
even when the eyelids were closed. It was necessary to keep
closing the eyes in order to slightly shrivel the face and
during this the head position also always changed. Therefore,
the error rate in this case was much higher and the text
writing took longer. One solution would be supplementing
the program by adjusting the sensitivity with which the eyes
are detected on the image. (This sensitivity is adjustable in
an argument of method called DetectHaarCascade).
If we focus on the results recorded in the table [Table 3],
it is obvious that the best results were again achieved by
using notebook HP which is situated in a well lighted setting.
During the testing it was also discovered that the sensitivity
of the application is still too high. On the other hand it has to
be noticed that attempts numbers 4 and 5 in tables are always
slightly better than the ones before. This means that practice
plays a considerable role when using the application and it is
necessary for the user to become accustomed to the operation
of the keyboard through face motion.
VI. CONCLUSIONS
Within this project we were able to successfully practise
working with a web camera through the detection of
different parts of the human body on a scanned image. After
certain changes the developed application could be largely
beneficial to handicapped people and it could simplify the
way in which they use their personal computer. It would be
necessary to solve two essential problems. The first problem
is the communication of the virtual keyboard with other
computer processes (Word, Explorer, etc.). The second
problem is getting all the function keys to work (CAPS,
SHIFT, etc.) and also the keys for operating the cursor. A
virtual keyboard, which is possible to be operated through
face motion, was established. The main advantage of this
project is that the user does not have to buy an expensive
hardware (if he has a camera integrated in a notebook). The
only thing necessary is to install Emgu.cv software (which is
free) and initiate the invented application.
ACKNOWLEDGEMENTS
This work was supported by SMEW Smart
Environments at Workplaces, Grant Agency of the Czech
Republic, GACR P403/10/1310.
REFERENCES
[1] Zhu, HJ., Koga, T.` lu d`ou |u:d ou ^duboo:
u!_o`hm `h ou! ^uo|o!u`ou `mu_, 'u IEICE
ELECTRONICS EXPRESS, \o!um , '::u 1, . 1125-
1131, 2010, DO'` 10.1587/elex.7.1125
[2] Sznitman, R., Jedynak, B.` ^` T:`u_ o lu D`ou
uud ou!`u`ou, 'u IEEE TRANSACTIONS ON PATTERN
ANALYSIS AND MACHINE INTELLIGENCE, \o!um 3?,
'::u 10, . 1914-1920, 2010, DO'`
10.1109/TPAMI.2010.106
[3] < http://cs.wikipedia.org >
http://cs.wikipedia.org/wiki/AVI [Accessed 8.10.2010]
[4] Hershler, O., Golan, T., Bentin, S., Hochstein, S.` Th `d `udo
o u d`ou, 'u JOURNAL OF VISION, \o!um 10,
'::u 10, ^`! uum| ?1, 2010, DO'` 10.1167/10.10.21
[5] < http://www.inb.uni-luebeck.de >
http://www.inb.uni-luebeck.de/publikationen/pdfs/
BoHaRiMaBa09.pdf [Accessed 8.10.2010]
[6] Peng, ZY., Ai, HZ., Hong, W., Liang, LH., Xu, GY.` Hu!`u
|u:d u uud u`u! uu d`ou ou `do
:_mu:, 'u JOURNAL OF COMPUTER SCIENCE AND
TECHNOLOGY , \o!um 18, '::u ?, . 241-246, 2003,
ISSN: 1000-9000
[7] < http://www.i4control.cz/ >
http://www.i4control.cz/ [Accessed 19.10.2010]
[8] < http://www.lifetool.at/ >
http://www.lifetool.at/show_content.php?sid=218 [Accessed
19.10.2010]
[9] < http://www.fugasoft.cz >
http://www.fugasoft.cz/index.php?cont=myvoice#o-programu
[Accessed 19.10.2010]
[10] < http://www.naturalpoint.com >
http://www.naturalpoint.com/smartnav/products/about.html
[Accessed 19.10.2010]
[11] < http://cs.wikipedia.org >
http://cs.wikipedia.org/wiki/OpenCV [Accessed 5.11.2010]
[12] < http://www.emgu.com >
http://www.emgu.com/wiki/index.php/Main_Page [Accessed
5.11.2010]
[13] < http://en.wikipedia.org >
http://en.wikipedia.org/wiki/Viola-Jones_object_detection_framework
[Accessed 5.11.2010]
[14] < http://en.wikipedia.org >
http://en.wikipedia.org/wiki/AdaBoost [Accessed 5.11.2010]