Content uploaded by Nazli Mohajeri
Author content
All content in this area was uploaded by Nazli Mohajeri on Jun 23, 2014
Content may be subject to copyright.
Abstract— Blind people face several problems in their life,
one of these problems that is the most important one is
detection the obstacles when they are walking. In this
research, we suggested a system with two cameras placed on
blind person's glasses that their duty is taking images from
different sides. By comparing these two images, we will be
able to find the obstacles. In this method, first we investigate
the probability of existence an object by use of special points
that then we will call them "Equivalent points", then we
utilize binary method, standardize and normalized cross-
correlation for verifying this probability. This system was
tested under three different conditions and the estimated
error is in acceptable range.
Index Terms— Obstacle detection, Blind people,
Equivalent points.
I. INTRODUCTION
YES play a vital role in our life. All of us have seen the
blind people and know the problems that they face in
their life. In order to detection the obstacles blind people
use stick when they are walking but this instrument just can
help them find objects on the ground. Obstacle detection is
a field of effort that has led to vast progress in primary
safety systems and in primary–secondary safety systems
interaction. To detect obstacles at a medium to long
distance, either static or mobile, different technologies have
been used, like laser scanners Solutions based on the
sensory fusion of lasers canner, radar and computer vision
have been used with the purpose of obtaining additional
information for a better interpretation of the environment,
as well as for mitigating the deficiencies of each sensor [1-
5].
Our suggested system uses two cameras placed on blind
person's glasses that take online images and within image
processing detection the obstacles, and finally notifies blind
person. In this system there is no need to recognize all the
obstacles but we just require those that are in a specific
distance from the cameras. As shown in figure (1) we have
considered a virtual plane with specific dimensions, any
kind of objects that enter this plane must be recognized by
Nazli Mohajeri is with the Electrical Engineering Faculty, Sahand
University of Technology, Tabriz, Iran (e-mail: nazi_mhjr@yahoo.com).
Roozbeh Raste is with the Electrical Engineering Faculty, Sahand
University of Technology, Tabriz, Iran (e-mail: roozbeh.raste@gmail.com).
Sabalan Daneshvar is with the Electrical Engineering Faculty, Sahand
University of Technology, Tabriz, Iran (corresponding author to provide
phone: +98912-2258523; fax: +98412-3459342; e-mail:
daneshvar@sut.ac.ir).
system. Figure (2) shows the thorough algorithm of the
system.
Fig.1 Position of the camera 1(right), 2 (left) and the virtual plane
Fig.2 The thorough algorithm
II. MATERIALS AND METHODS
A. Equivalent Points
In this research, there is an object 70 cm far from the
cameras, this object will specify a particular place for itself
in the pictures taken by cameras (Fig.3). We call these
places (two points in pictures) "Equivalent Points". These
points are unique (i.e. those pixels that are equivalent for
70 Cm distance are not equivalent for 75Cm), additionally
computing these points are completely parametric and you
can change the initial conditions (initializing depends on
field of vision of cameras).
As initializing, 50 points on x axis and 40 points on y
axis (one point per cm) have been considered (i.e. 2000 pair
of equivalent points), all of these points will be probed on
virtual plane (Fig.4). Exploring the points is from the top of
the plane to bottom and from right to left (Fig.5 right). The
place of our points is expressed in centimeter and we need
to transform them into pixels. As shown in figure (5 left)
we use simple calculations for transforming. For an
example, if we multiple the division of dby (the height of
recent point) and day (the whole height of taken picture), to
An Obstacle Detection System for Blind People
Nazli Mohajeri, Roozbeh Raste, Sabalan Daneshvar*, member, IAENG
E
yr/2 (half of the height of picture) and subtract the result of
that from yr/2, we will obtain the height of the expressed
point in pixel.
Fig.3 Simulation of a point in vision of camera 1(right) and 2(left)
Fig.4 All equivalent points in picture of camera 1(right) and 2(left)
Where, day= h.tg(za) and Za is vertical field of vision of
cameras. After calculations of height, all 50 points in this
height must be explored; in this stage we define three
regions (Fig.6 right) and behave in same way (Fig.6 left).
By using this method the equivalent pixels will be detected.
If gray scale of 2 equivalent pixels has differences less than
5 in value, being an object in that place will be probable.
There must be no difference between two pixels if they are
from same object but because of existing shadows, we apply
this tolerance, if this probability is verified by the system,
more difficult decisions will be made for later sections, if
not the system will ensure that the taken pictures are from
different objects and there is no object in that distance and
will calculate other points.
B. Obstacle detection
This section is the step of making more difficult
decisions for detection obstacle. For excel describing, we
show processing on one of the pictures step by step (Fig.7).
First we convert our pictures to binary mode by changing
the value of pixels to 0 and 1, if they are in range of
equivalent pixels will get 1 if not will be 0, (Fig.8). Then
we separate the regions that involve equivalent points and
blacken the rest of the image (Fig.9).
Fig.5 Height of points in Cm and pixel mode (left) and method of exploring
the points (right)
Fig .6 Widths of points in cm and pixel mode (left) and regions on virtual
plane (right)
Fig.7 Pictures taken by camera 1(right) and 2(left)
Fig.8 Binary pictures
Fig.9 Separated regions
Fig.10 Regions that are not in vision of each camera
In order to find out whether or not these obtained regions
refer to one object, we draw a bonding box around them.
Since taken images by each camera include some regions
that are not exist in second one, we must cut these regions
out (Fig.10), this should carry out specially when a part of
object is in vision of one of the cameras and not in vision of
other one. Now the center of two separated regions must be
calculated so we should change the pixels to Cm, for
avoiding having error we accept 3 cm tolerance and ignore
regions that are less than 10 pixels.
When an object is in front of a camera, it looks bigger in
vision of that camera than another one, for solving this
problem we have performed a method that stretches the
picture of objects and makes them standardized and
comparable. Fig.11 shows this method.
Fig.11 Width of object in picture and its real width
Fig.12 Standardized pictures of camera 1(right) and 2(left)
Where, CO is center of object and HW is half of the
width of region. f is distance between the center of object
and center of camera, y is the real width of object and za is
horizontal field of vision of camera.
By means of these calculations, we defined a coefficient
that can express the real size of object in picture versus its
distance from the center of camera. This coefficient can
help to make two images comparable and stretching them
by a "projective transformation" (Fig.12).
We employed normalized cross-correlation for
discovering the similarity of images, if the value of cross-
correlation of images is more than 0.9, we will consider
them similar (Fig.13). In ideal form the value must be 1 but
considering it 0.9 doesn’t make mistake, on the other hand
since we have employed lots of sections for finding object,
it looks impossible that this insignificant difference could
make error.
Fig.13 Result of normalized cross-correlation
After this step the existence of obstacle is confirmed and
the place of that will be determined by its position in that
three regions. Finally the system will notify the user as
these sentences "there is an object in front of you" for
region number 2," there is an object on your right" for
region number 1 and " there is an object on your left" for
region number 3.
III. RESULTS AND DISCUSSIONS
This system was tested under three conditions and for
three objects. For taking images we used CANON SX120
IS, 0.3 mega pixels, without using any automatic forms.
The program was processed by DELL Inspiron 1545
notebook with processor: Intel(R) Celeron(R) CPU @ 2.20
GHz and installed memory (RAM) 2.00 GB.
First we tested the system for a circle object under lamp
light and the camera was fixed 55cm height from ground
level (Fig.14-a). In Figure (14-b) square object was 70 cm
far from camera but circle one was nearer, we observe that
system could not detect the circle object. In next one square
object was in same place and circle object was farer from
camera, system could not detect the circle one like the
former picture (Fig.14-c). Figure (14-d) describes the result
of our experiment under sunlight for a black object and
camera was fixed 70 cm height from ground level. In figure
(14-e), object was in same place but camera was in 81cm
height with 20 degrees of depression angle. Duration of
processing was 33.7969 seconds.
In this research, our purpose is detection of those
obstacles that their distance from cameras is 70Cm.
Undoubtedly system has error, for estimating this error we
tested the system for 4 distances 70 cm, 71 cm, 72 cm and
69 cm, the results show that system detected an obstacle for
70 and 71 cm (Fig. 14-f and 14-g) but not for 72 and 69
cm, thus the error of system is: (72-70 , 69-70) => ( 2 , -1 )
cm.
IV. CONCLUSION
Our suggested system has high sensitivity for the height of
cameras and distance between them. In this research, we
designed a system with two cameras placed on blind
person's glasses that their duty is taking images from
different sides. By comparing these two images, we could
be able to find the obstacles. In the proposed method, first
we investigated the probability of existence an object by use
of special points that then we called them "Equivalent
points", then we utilized binary method, standardized and
normalized cross-correlation for verifying this probability.
This system was tested under three different conditions and
the estimated error is in acceptable range.
(a)
(b)
(c)
(d)
(e)
(f)
(g)
Fig.14 Results of tested system
REFERENCES
[1] Felipe Jiménez, José Eugenio Naranjo, "Improving the obstacle
detection and identification algorithms of a laserscanner-based collision
avoidance system," Transportation Research Part C, Article in press.
[2] Kunsoo Huh, Jaehak Park, Junyeon Hwang, Daegun Hong, "A stereo
vision-based obstacle detection system in vehicles," Optics and Lasers
in Engineering, Vol. 46, pp. 168–178, 2008.
[3] J. Jesus Garcia, Jesus Urena, Manuel Mazo, Felipe Espinosa, Alvaro
Hernandez, Cristina Losada, Ana Jimenez, Carlos De Marziani,
Fernando Alvarez, Enrique Garcia, "Sensory system for obstacle
detection on high-speed lines," Transportation Research Part C ,
Vol.18, pp. 536–553, 2010.
[4] Iwan Ulrich and Illah Nourbakhsh, "Appearance-Based Obstacle
Detection with Monocular Color Vision," Proceedings of the
Seventeenth National Conference on Artificial Intelligence and
Twelfth Conference on Innovative Applications of Artificial
Intelligence, 2000.
[5] Joshua Redding, Jayesh N. Amin, Jovan D. Boskovic, Yeonsik Kang,
Karl Hedrick, Jason Howlett, Scott Poll, "A Real-time Obstacle
Detection and Reactive Path Planning System for Autonomous Small-
Scale Helicopters", AIAA Guidance, Navigation and Control
Conference, Carolina, USA, Aug. 2007.