Forward control of robotic arm using the information from stereo-vision tracking system

Abstract

In this paper we present the feed-forward neural network controller of robotic arm, which makes use of tracking method applied to stereo-vision cameras mounted on the head of the humanoid robot Nao, in order to touch the tracked object. The Tracking-Learning-Detection (TLD) method, which we use to detect and track the object, is known for its state-of-art performance and high robustness. This method was adjusted to be usable with a stereo-vision camera system, in order to provide 3D spatial coordinates of the object. These coordinates are used as the input for the feed-forward controller, which controls the arm of a humanoid robot. The goal of the controller is to move the hand of the robot to the object by setting arm joints into position corresponding to the object location. The controller is implemented as an artificial neural network and trained using the error back-propagation algorithm. The experiment, which demonstrates the proof of the concept, is also denoted in this paper.

Full text

Forward Control of Robotic Arm Using the
Information from Stereo-vision Tracking System
Michal Puheim1, Marek Bundzel2, Ladislav Madarász3
Department of Cybernetics and Artificial Intelligence
Technical University of Kosice
Letná 9, 042 00 Košice, Slovak Republic
1michal.puheim@tuke.sk, 2marek.bundzel@tuke.sk, 3ladislav.madarasz@tuke.sk
Abstract—In this paper we present the feed-forward neural
network controller of robotic arm, which makes use of tracking
method applied to stereo-vision cameras mounted on the head of
the humanoid robot Nao, in order to touch the tracked object.
The Tracking-Learning-Detection (TLD) method, which we use
to detect and track the object, is known for its state-of-art
performance and high robustness. This method was adjusted to
be usable with a stereo-vision camera system, in order to provide
3D spatial coordinates of the object. These coordinates are used
as the input for the feed-forward controller, which controls the
arm of a humanoid robot. The goal of the controller is to move
the hand of the robot to the object by setting arm joints into
position corresponding to the object location. The controller is
implemented as an artificial neural network and trained using
the error back-propagation algorithm. The experiment, which
demonstrates the proof of the concept, is also denoted in this
paper.
Keywords—arm controller, neural network, TLD, tracking,
learning, detection, humanoid robot, Nao, stereo-vision
I. INTRODUCTION
This paper proposes a possible application of Tracking-
Learning-Detection (TLD) [1]–[3] algorithm in stereo-vision
on the Nao robot [4]. TLD is able to detect and track an object
in a continuous series of video frames. Position of the object is
defined as coordinates of its bounding box (x, y, width, height).
This gives us the idea about the position of the object in the
frame, but no further information about its location in the
environment. This problem occurs due to the fact that TLD is
only processing two-dimensional picture and the information
about the third dimension is not available.
In order to obtain the additional information, it is possible
to employ another detector using the camera with different
viewpoint. With the information about the object position in
the frames from both cameras and the information about
mutual position of these cameras, we are able to determine the
position of the object in three-dimensional space quite
precisely.
Then we can use this information as the input for the
forward arm controller, based on the inverse control model
with goal to enable the robot’s interaction with the tracked
object in three-dimensional environment.
II. TRACKING-LEARNING-DETECTION
Tracking-Learning-Detection (TLD) [1]-[3] is an object
tracking method usable for long-term tracking of arbitrary
objects in unconstrained environments [2]. TLD system is
made of three independent components:
• Tracker – a short term tracker based on Lucas-Kanade
method [5], which is used to generate examples for
training of the detector.
• Detector – has a randomized forest form [6], enables
incremental update of its decision boundary and real-
time sequential evaluation during run-time [2]. Runs
independently from the tracker.
• Learning algorithm – so called “P-N Learning” [1] uses
the tracker to generate positive (P) and also negative
(N) examples that are further used in order to improve
the model of the detector.
Before the tracking begins, the bounding box of the object
is manually selected by the supervisor. After that, during TLD
runtime, the object is tracked by the tracker component. During
the tracking, the model of the object is built for use with the
detector component. It is built upon the information from the
first frame, as well as the information provided by the tracker.
If the detector has successfully finished the training process, it
enables re-detection of the object, whenever the tracker fails.
Both detector and tracker make errors, the stability of the
whole system is achieved by mutual cancellation of these
errors [2].
A. Tracking
The tracker used with TLD, as proposed by [3], is Median-
shift (or Median-flow) tracker. It is a rectangular shaped
tracker based on the Lucas-Kanade tracker [5], which is robust
to partial occlusions and also estimates translation and scale. It
is designed to track number of points between consecutive
frames and estimates a rectangle displacement and scale
change using median of tracked points. Tracked points are
selected as 50 % of the most reliable points within tracked
rectangle, using forward-backward error, as proposed by [3]. In
each frame a completely new set of feature-points is tracked,
which makes the tracker very adaptive [2].
57
CINTI 2013 • 14th IEEE International Symposium on Computational Intelligence and Informatics • 19–21 November, 2013 • Budapest, Hungary
978-1-4799-0197-5/13/$31.00 ©2013 IEEE

The forward-backward error is defined as a difference
between two feature-point trajectories, where the first one is
defined by forward motion of tracked point and the second one
by its backward motion. The next paragraph defines this error
according to [3].
Let S = (It, It+1, …, It+k) be an image sequence and xt be a
point location in time t. Using an arbitrary tracker, the point xt
is tracked forward for k steps. The resulting trajectory is
Tk
f = (xt, xt+1, …, xt+k), where f stands for forward and k
indicates the length. Our goal is to estimate the error of
trajectory Tk
f within the image sequence S. For this purpose, the
validation trajectory is constructed first. Point xt+k is tracked
backwards, up to the first frame and produces
Tk
b = (x't, x't+1, …, x't+k), where x't+k = xt+k. The final forward-
backward error for given point is then given as the Euclidean
distance between xt and x't.
Fig. 1. Tracking of the point trajectory during forward-backward error
calculation. Main idea is that some points cannot be re-tracked to their initial
position [3].
This kind of recursive tracking is possible as long as the
tracked object is visible in the image. The tracker will
eventually fail in case of occlusion or other dynamic change in
the appearance of the object. In this case the median shift is
greater than the accepted threshold and the recursive tracking is
forced to stop. When this happens, tracker is unable to
reinitialize itself, because it lacks mechanism for object
detection [7]. In this situation the TLD relies on the detector
component.
B. Detection
A detector of the TLD system is based on scanning win-
dow [8] and the randomized fern forest classifier [6]. The
training and testing is on-line [1]. Using the scanning window
approach, the input image is scanned across possible positions
and scales and at each sub-window a binary classifier decides
about presence of the object [1]. The following paragraphs
describe detector design as explained in [2].
The model of the given object is defined by a set of image
patches, which represent possible appearances of the object.
Each patch is described by a number of local 2bit Binary
Patterns (described in Fig. 2), which position, scale and aspect
ratio were generated at random. These features are randomly
partitioned into several groups of the same size. Each group
represents a different view of the patch appearance. The
response of each group is represented by a discrete vector that
is called a “branch” [2].
The classifier used for detection has the randomized forest
form [6]. It enables online update and sequential evaluation.
The forest consists of several trees, each of them is build from
one group of features. Every feature in the group represents a
measurement taken at a certain level of the tree [2].
Fig. 2. Local 2bit Binary Patterns used in object detector. Features encode
local gradient orientation within the object bounding box. Their position
within image patch, scale and aspect ratio are generated at random [2].
At the beginning, every tree contains one single branch
defined by the selected patch. With every positive example a
new set of branches is added to the forest. Growing and
pruning of the forest is therefore incremental – one example is
processed at a time [2].
Evaluation of an unknown patch by the tree is very
efficient. The features within each tree are measured
sequentially. If the patch reaches the end of the tree, it is
considered positive. Otherwise, if it differs from the defined
branches, the measurement is terminated and the patch is
considered negative. The sequential nature of the randomized
tree enables real-time evaluation on a large number of
positions. The input image is scanned with a sliding window.
At each position, every tree makes a decision, whether the
underlying patch is in the model or not. The final decision is
obtained by the majority vote [2].
Fig. 3. Detector based on scanning window and randomised fern forest
classifier [1].
C. Learning
Learning algorithm, employed in the TLD system, is called
P-N Learning [1]. It uses positive (P) and negative (N)
examples that are used in order to improve the model of the
Mi. Puheim et al. • Forward Control of Robotic Arm Using the Information from Stereo-vision Tracking System
58

detector. Main idea of this approach is to make use of
inevitable errors of both tracker and detector. The stability of
the system is achieved by error cancellation of both
components. The rest of this section gives a brief description of
P N Learning according to [1].
Training of the initial classifier is performed in the first
frame. The posteriors from all ferns in the classification tree are
initialized to zero and are updated by positive examples
generated by affine warping of the selected patch. The
classifier is then evaluated on all patches. The detections far
from the target represent the negative examples and further
updates the posteriors [1].
Fig. 4. A trajectory of an object represents a structure in video..Patches close
to the trajectory are positive (YELLOW), patches far from the trajectory are
negative, only the difficult negative examples are shown (RED). In training,
the real trajectory is unknown,but parts of it are discovered by a validated
adaptive tracker [1].
When the P-N learning is initialized in the first frame by
training the Initial Detector, the initial position of the Lucas-
Kanade tracker is set using the selected patch. Afterwards, for
each of the consecutive frames, the detector and the tracker
find the location(s) of the object. As depicted in Fig. 4 and
Fig. 5, the patches close to the trajectory, given by the tracker
and detections far away from this trajectory are used as positive
and negative examples respectively. If the trajectory is
validated (i.e. forward-backward consistent), these examples
are used to update the detector. However, if there is a strong
detection far away from the track, the tracker is reinitialized
and the collected examples are discarded [1].
Fig. 5. Representation of the P-N Learning. Object is tracked by the tracker
and simultaneously detected by the detector. Patches close to the trajectory
update the detector with positive label, other patches update the detector with
negative label [9].
The trajectory of the tracker is then denoted as P-N
Tracker, since it is reinitialized by a detector trained by online
P-N learning. The detector that is obtained after processing all
frames of the sequence is called Final Detector [1].
III. STEREO-VISION TRACKING SYSTEM
An advantage of stereo-vision systems is their ability to
extract more information from an image than typical single
camera systems [10]. Goal of the proposed stereo-vision
tracking system [15] is to determine the spatial coordinates of
the tracked object in the three-dimensional environment and
provide it to the arm controller. The arm controller is then
supposed to use this information in order to move hand of the
robot towards the object.
A. Calculation of Depth in Stereo-vision Systems
Let us have two identical cameras with parallel optical axes
(i.e. in rectified configuration), as shown in Fig. 6. If f is the
focal distance of the cameras, c is the distance between the
cameras, a = c/2 is the distance of cameras from the central
optical axis, XR is the x-coordinate of the object in the image
from the right camera and XL is the x-coordinate of the object in
the image from the left camera, then by using the similarity of
triangles we get the following equations:
z
ax
f
XL−
=, z
ax
f
XR+
= (1)
By merging these equations and elimination of x we get the
formula, which can be used to calculate the z coordinate:
RL XX
af
z−
−
=2 (2)
z
x
Camera 1
Camera 2
X
L
X
R
a
a
c
f
[x, 0, z]
Object
Fig. 6. Estimation of object z coordinate using the stereo cameras with
parallel optical axis (i.e. in rectified configuration). Distance between the
cameras is c = 2a, where a is distance of the camera from the optical axis.
XR and XL are coordinates of the object as seen by right and left camera
respectively. If f is focal length then z coordinate can be calculated using XR
and XL.
59
CINTI 2013 • 14th IEEE International Symposium on Computational Intelligence and Informatics • 19–21 November, 2013 • Budapest, Hungary

The equation (2) basically expresses that with known
camera parameters f and d, we can use the difference in vision
of the same object by both cameras in order to determine the
distance of the object from the image plane z (and also the
distance of the object from the cameras). The difference in
vision between two cameras is called disparity [10].
B. Integration of TLD in the Stereo-vision Tracking System
In spite of the simplicity of the formula used to calculate
depth, stereo-vision is not a trivial task, particularly because of
the correspondence problem [11]. The root of this problem is
that search of the corresponding points (and objects) in the
images from both cameras is difficult and usually requires
sophisticated algorithms based on the edge detection (e.g.
RANSAC [12]). In order to solve this problem for the purpose
of this work, we used the combination of two TLD systems
tracking the same object simultaneously, one on each camera.
TLD 1
TLD 2
Camera 1
Camera 2
3D World
Object
image
image
sync
Triangulation
(x
1
, y
1
)
(x
2
, y
2
)
Object
coordinates
(x, y, z)
Fig. 7. Simplified block diagram of the stereo-vision tracking system which
employs the TLD method for object tracking. Tracked object is captured by
two cameras at the same time. Images from the cameras are used as the input
for two synced TLD systems. Each of them produces the 2D coordinates of
the target object in respective image. The couple of 2D coordinates is then
further used to calculate 3D coordinates of the target by means of the
triangulation method.
The proposed stereo-vision tracking system uses two TLD
subsystems (see Fig. 7), which are based on the implement-
ation by George Nebehay [7]. Tracked object is selected
manually by the human operator from the image of single
camera. At this point, the initial object model for the first TLD
subsystem is created. This model is then copied to the second
TLD subsystem. Both models are regularly synchronized in
order to avoid divergence in object detection. Outputs of each
subsystem are the coordinates of the object bounding box
(x, y, w, h) in respective camera image (see Fig. 8). Coordi-
nates of the object itself within the image are defined as the
center (xc, xc) of the respective bounding box.
Let us assume that wi is the image width (in pixels), hi is the
image height (in pixels), f is the focal distance of the cameras
(in pixels) and c is the distance between cameras (in meters). If
(xL, yL) are the center coordinates of the object on the left image
and (xR, yR) are the center coordinates of the object on the right
image (given in pixels), then the 3D coordinates of the target
object (x, y, z) can be calculated as:
x
y
Image
Object Bounding box
h
x
y
w
Fig. 8. TLD tracks the kernel of the object which is delimited by bounding
box. It is defined by horizontal and vertical coordinate of the upper left corner,
width and height. These four values (x, y, w, h) represent the output of the
TLD system.
RL XX
cf
z−
−
=, (3)
,
2
2f
w
X
z
c
x
i
L⎟
⎠
⎞
⎜
⎝
⎛−
+= .
2
f
h
X
zy
i
R⎟
⎠
⎞
⎜
⎝
⎛−
−= (4)
In application to humanoid robots, it is more convenient to
use polar coordinates of the object instead of the Cartesian
coordinates. These are given as (d, α, β), where d is the
distance of the object and α, β are directional angles. The
distance d of the target object from the cameras is the length of
the vector (x, y, z):
222 zyxd ++= . (5)
If λv is the vertical field of view of the camera and λh is the
horizontal field of view of the camera, then the directional
angles α, β can be approximated as:
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛−≈ 5,0
i
L
vh
y
λα
, ⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛−
−
+
≈5,0
2i
RL
hw
xx
λβ
. (6)
IV. ROBOTIC ARM CONTROLLER
Goal of the controller is to move the hand of the robot to
the position of the tracked object using the information about
the object, provided by the tracking system. In order to achieve
this goal, we used feed-forward neural network with three
inputs and three outputs.
In order to train the neural network, it is necessary to
generate a sufficient set of training data. It can be created by
exploiting the inverse model of the system.
Mi. Puheim et al. • Forward Control of Robotic Arm Using the Information from Stereo-vision Tracking System
60

A. Topology of the Artificial Neural Network
Inputs of the artificial neural network are polar coordinates
of the object position related to the robot body (which means
that head pitch and yaw are also taken into account when we
determine the object directional angles). Outputs are the values
of joint positions of the robotic arm corresponding to the
current object position in 3D space. There are two hidden
layers within the neural network, each including four neurons.
Object horizontal
direction angle
Object vertical
direction angle
Object
distance
1st hand joint
position
(elbow roll)
2nd hand joint
position
(shoulder roll)
3rd hand joint
position
(shoulder pitch)
Fig. 9. Arm controller of the humanoid robot is based on the feed-forward
neural network. Inputs are the information about the object position related to
the robot's body. Outputs are the positions of the arm joints. Setting arm joints
to these positions will effectively move robot's hand to the position of the
tracked object.
B. Training of the Neural Network
We set the robotic arm to use three degrees of freedom, i.e.
the movement is enabled on three joints only. Now let us
assume, that these joints are “shoulder roll”, “shoulder pitch”
and “elbow roll” or (s1, s2, s3). Let us assume, that the tracked
object is some random position around the robot. The task is to
determine the joint position (s1, s2, s3) corresponding to the
polar spatial coordinates of the object (d, α, β). In other words,
the task is to perform the transformation:
T: (d, α, β) → (s1, s2, s3), (7)
which is not possible without an adequate mathematical
model or feedback control. Other option to carry out the
transformation T is to use the trained neural network.
In order to generate the training data, we put the tracked
object into the robot's hand. If we put the arm joints in one
certain position (s1, s2, s3), we can use the stereo-vision
tracking system in order to get the polar spatial coordinates
(d, α, β) of the object in robot's hand. In other words, it is
possible to carry out the inverse transformation:
T’: (s1, s2, s3) → (d, α, β). (8)
We can use it to generate corresponding pairs of input and
output data [(d, α, β); (s1, s2, s3)]. This data can be used in
training of the neural network arm controller using the error
back-propagation algorithm.
In order to achieve satisfactory results of the trained neural
network it is necessary to normalize the input and output data
on the interval between 0 and 1. This fact may become a
problem, for example if the domain of the given input (or
output) is infinitely large. In this case, it is necessary to limit its
range to appropriate values. For example maximum distance
range should be limited to the length of robot's arm. Also the
minimum distance range should be limited according to the
limits of the stereo-vision tracking system. Similar
considerations should be made for other input/output domains
[16].
V. EXPERIMENT
In order to test the applicability of the proposed arm
controller, we performed an experiment, which tested the
ability to touch the tracked object. We put the object 20 times
in several different positions within reach of the robot's hand.
In each attempt to touch the object, we determined its success.
Results are shown in Table I.
The experiment has proven that the proposed combination
of stereo-vision tracking system and neural network arm
controller is feasible. Precision of the hand movement is
sufficient for the feed-forward controller. It can be further
improved by implementation of feed-back control.
In the current stage of development, the proposed object
tracking system is not able to provide more information about
the target object except for its location. This can be a problem
if it is necessary to manipulate with the objects of different
sizes or shapes. This problem can be solved by employing a
different object tracking method (see [13]) or by tracking
various parts of the tracked objects separately in order to
determine its orientation additionally.
VI. CONCLUSION
In this work we made a brief overview of TLD system
architecture, as well as its individual components. Illustration
of its possible application in stereo-vision tracking system has
also been mentioned. We applied this system to the Nao
humanoid robot and used its output information about the
tracked object as the input for the arm controller in order to
enable the robot to interact with the tracked object. We carried
out the experiment in which we demonstrated the viability of
the proposed system.
At present the proposed system is running on the PC and
connecting to the Nao robot via proxy connection. For future
work it would be interesting to employ the proposed system on
the cloud, which would enable multiple robots to take
advantage of it [17][18].
ACKNOWLEDGEMENT
The work presented in this paper was supported by VEGA,
Grant Agency of Ministry of Education and Academy Science
of Slovak Republic under Grant No. 1/0298/12 – “Digital
control of complex systems with two degrees of freedom.” The
work presented in this paper was also supported by KEGA
under Grant No. 018TUKE-4/2012 – “Progressive methods of
education in the area of control and modeling of complex
systems object oriented on aircraft turbo-compressor engines.”
61
CINTI 2013 • 14th IEEE International Symposium on Computational Intelligence and Informatics • 19–21 November, 2013 • Budapest, Hungary

TABLE I. RESULTS OF THE HAND CONTROLLER EXPERIMENT
NN Inputs (object coordinates) NN Outputs (hand joint positions)
OK?
d (m) v (rad) h (rad) sp (rad) s
r (rad) e
r (rad)
1. 0.134146 0.027178 0.290937 0.057461 0.357181 -0.788319 Y
2. 0.190570 -0.068388 0.589369 -0.172792 0.477867 -0.169463 Y
3. 0.198999 -0.073607 0.765424 -0.184782 0.740034 -0.160568 Y
4. 0.164649 0.103553 0.424750 0.061034 0.431430 -0.588861 Y
5. 0.221017 -0.173743 0.923930 -0.495389 0.778589 -0.039320 Y
6. 0.223769 0.021826 0.735878 -0.166607 0.515333 -0.049318 Y
7. 0.162134 -0.435731 0.842525 -0.614742 1.071152 -0.477502 N
8. 0.194244 -0.387002 1.168614 -0.553836 1.345710 -0.174646 N
9. 0.191741 -0.055399 0.831764 -0.122697 0.975315 -0.333835 Y
10. 0.185708 -0.190491 0.733462 -0.329463 0.733202 -0.207705 Y
11. 0.153799 -0.011105 0.150542 -0.038457 0.190594 -0.393123 Y
12. 0.163303 0.320239 0.386022 0.225971 0.462984 -0.762712 Y
13. 0.185293 0.017649 0.357128 -0.072228 0.261859 -0.185558 Y
14. 0.217610 -0.034607 0.627409 -0.253171 0.383026 -0.042686 Y
15. 0.201207 -0.011139 0.518828 -0.137416 0.347757 -0.096682 Y
16. 0.209144 -0.268592 1.175758 -0.458409 1.302464 -0.116085 Y
17. 0.129244 -0.330636 0.627409 -0.385313 0.940373 -1.098963 N
18. 0.141164 -0.175151 0.122507 -0.166845 0.178159 -0.397670 N
19. 0.193073 -0.401333 0.510929 -0.773710 0.276209 -0.030392 Y
20. 0.193977 0.237270 0.403652 0.181841 0.320561 -0.257261 Y
Successful touch: 16 Failed touch: 4 Percentage: 80 %
Each row represents one measurement, distance values d are given in meters, all other values are given in radians. Inputs of the neura l netwo rk hand controller: d – object distance, v – vertica l direct ional angle of the
object, h – horizo ntal directiona l angle o f the obj ect. Out puts of t he neural networ k hand cont roller: sp – shou lder pitc h joint, sr – shou lder roll joint, er – elbow r oll joint. Last co lumn determines if the touc h was
successful or not .
REFERENCES
[1] Z. Kalal, J. Matas, K. Mikolajczyk: “P-N Learning: Bootstrapping
Binary Classifiers by Structural Constraints.” In Conference on
Computer Vision and Pattern Recognition. 2010.
[2] Z. Kalal, J. Matas, K. Mikolajczyk: “Online learning of robust object
detectors during unstable tracking.” In 3rd On-line Learning for
Computer Vision Workshop 2009, Kyoto, Japan, IEEE CS, 2009.
[3] Z. Kalal, J. Matas, K. Mikolajczyk: “Forward-Backward Error:
Automatic Detection of Tracking Failures.” In International Conference
on Pattern Recognition, Istambul, Turkey, 23-26 August, 2010.
[4] Aldebaran Robotics: Nao Website, [Online]. 2013. Available:
http://www.aldebaran-robotics.com/en/
[5] B. D. Lucas and T. Kanade: “An iterative image registration technique
with an applica-tion to stereo vision.” In Proceedings of the International
Joint Conference on Artificial Intelligence, pages 674–679, 1981.
[6] L. Breiman: “Random forests.” ML, 45(1):5–32, 2001.
[7] G. Nebehay: “Robust Object Tracking Based on Tracking-Learning-
Detection.” Diplomarbeit. Fakultät für Informatik, Technische
Universität Wien. 2012.
[8] P. Viola, M. Jones: “Rapid object detection using a boosted cascade of
simple features.” CVPR, 2001
[9] Z. Kalal, J. Matas, K. Mikolajczyk: Poster – Components of TLD
[Online]. Available: http://kahlan.eps.surrey.ac.uk/featurespace/tld/
Publications/2010_cvpr_demo_poster.pdf
[10] R. Hartley and A. Zisserman: “Multiple View Geometry in Computer
Vision”, Cambridge University Press, 2004.
[11] Abhijit S. Ogale and Yiannis Aloimonos: “Shape and the stereo
correspondence problem”. International Journal of Computer Vision 65
(1). October 2005.
[12] Martin A. Fischler and Robert C. Bolles: “Random Sample Consensus:
A Paradigm for Model Fitting with Applications to Image Analysis and
Automated Cartography”. Comm. of the ACM 24 (6): 381–395. June
1981.
[13] A. Yilmaz, O. Javed, M. Shah.: "Object Tracking: A Survey." ACM
Comput. Surv. 38, 4, Article 13, 45 pages. Dec. 2006.
[14] M. Puheim.: “Application of TLD for object tracking in stereoscopic
images”. Diploma thesis (in slovak). Technical University of Košice.
Faculty of Electrical Engineering and Informatics. Košice. 2013. 68
pages.
[15] M. Puheim , M. Bundzel, P. Sinčák, L. Madarász: “Application of
Tracking-Learning-Detection for object tracking in stereoscopic
images”. Symposium on Emergent Trends in Artificial Intelligence and
Robotics. Košice. Sep. 2013.
[16] Nagy, I., Várkonyi-Kóczy, A.R., Dineva, A.: GA Optimization and
Parameter tuning at the Mobile Robot Map Building Process. IEEE 9th
International conference on Computational Cybernetics, July 8-10, 2013.
Tihany, Hungary. pp. 333-338. ISBN 978-1-4799-0060-2.
[17] Jordán, S., Haidegger, T., Kovács, L., Felde, I., Rudas, I.: “The Rising
Prospects of Cloud Robotic Applications.”, IEEE 9th International
conference on Computational Cybernetics, July 8-10, 2013. Tihany,
Hungary. pp. 327-332. ISBN 978-1-4799-0060-2.
[18] Tar, J.K., Rudas, I., Kósi, K., Csapó Á., Baranyi, P.: “Cognitive Control
Initiative”. 3rd IEEE International Conference on Cognitive
Infocommunications, December 2-5, 2012, Košice, Slovakia. pp. 579-
584. ISBN 978-1-4673-5188-1.
Mi. Puheim et al. • Forward Control of Robotic Arm Using the Information from Stereo-vision Tracking System
62