Principal Component Analysis for 3D-Manipulator

Robot Control System

Moussa Hamadache*, Jaehoon Kim, Dongik Lee.
School of Electronics Engineering, Kyungpook National University
1370 Sankyug-dong, Buk-gu, Daegu 702-701, Korea
*E-mail: [email protected]

Abstract—In this paper, a shape recognition technique for a 3-

DOF manipulator robot control system is proposed. This work is
devised into three parts. First, the shapes were recognized under
the Principal Component Analysis (PCA) algorithm and the
shape’s characteristics (position and orientation) were extracted
and stored in the PCA-bases. Second, based on the Graphical
User Interface (GUI) tool, a simulation on 3D-space of the 3-
DOF manipulator robot was performed. Finally, the PCA-bases
were applied as a command –signal to the 3-DOF manipulator
robot control system. Several shapes were considered and tested
to verify the method.
Figure 1. The global system structure.
INTRODUCTION
One of the main disciplines in the industry is the robotics. SHAPE RECOGNITION TECHNIQUE
Which can be used in the development of new technologies, A. Principal Component Analysis (PCA)
the combination and/or the cohesion of robotics with different By seeking the projection that best represent data in a least
applications as artificial intelligent, fuzzy-logic, image square sense, the PCA algorithm permit to reduce the
analysis, and computer vision system allow the innovation and dimension and data compression. In image analysis a 2-D
reduces the manufacture costs. In this purpose, it is important facial image is represented as 1-D vector by touch each
to achieve several knowledge and techniques. One of this row/column into a long vector, if the set of sampled image
knowledge is the computer vision area that proved its utility in given by M vectors of size N, and pj’s represent the pixel.
robot control system. One of the most popular techniques in
vision system is the shape recognition, which applied in = ,…, ; = 1, … , (1)
several research domains, not only in computer science, but
neuroscientists and psychologists also. It is the general opinion
The image’s mean centered is calculated by subtracting the
that advances in computer vision research will provide useful
mean image m from each image vector.
insights to neuroscientists and psychologists into how human
brain works, and vice versa. In this purpose, this paper was
proposed, where the shape recognition task was performed = ∑ (2)
successfully under the PCA algorithm.
PCA was developed to recognize a shape from an image and = − is define as mean centered image.
token through a camera under the extraction of the shape’s Our interest is finding a set of M orthogonal vectors vi’s
characteristics (coordinates & orientations). And then these which have the largest possible projection onto each of the ci’s
characteristics were stored on the PCA-bases. These bases for which the quantity.
were used as a command-signal to the robot control system.
So the recognized shape will be drawn by matching image
= ∑ ( ) (3)
feature to a geometrical model of the shape and compute its
orientation angles relative to the robot coordinate system. In
addition, the direct kinematics equation of the 3-DOF robot is maximized with the orthonormality constraint = .
was established. Where vi’s and i’s are given by the eigenvectors and
The system structure used in this study is illustrated in the eigenvalues of the covariance matrix.
Fig. 1
= . (4)

School of Electronics Engineering,, Kyungpook National University

978-1-4673-0784-0/12/$31.00 ©2012 IEEE 395

 With C is the matrix composed by the column vectors ci B. Shape Recognition
placed side by side. W is N×N matrix. Finding an eigenvectors Shape recognition has several tasks as: identification
of W is not practical solution. In linear algebra a common where the labels of individuals must be obtained; recognition
theorem show that the vectors vi and scalars i can be obtained of a person, where it must be decided if the individual has
by computing the eigenvectors and eigenvalues of the M×M already been seen; categorization, where the face or the shape
matrix CTC. Let ei and µi be the eigenvectors and eigenvalues must be assigned to a certain class. Once the eigenfaces have
of CTC, respectively. been computed, several types of decision can be made
depending on the application.
= . (5)
Training vectors that represent the basis of the space are
by multiplying the left of the both sides by C: ( )= computes with PCA, and these basis vectors or eigenvectors
( ). are directed from the largest variance of the training vectors,
and it named by eigenfaces as we mention previously. Each
This means that the first M-1 eigenvectors vi and eigenface can be viewed a feature. In our case when a
eigenvalues i of CCT are given by Cei and µi, respectively. particular shape is projected onto the face space, its vector into
Cei need to be normalized in order to be equal to vi, and the the face space describes by its eigenface coefficients called the
rank of the covariance matrix cannot exceed M-1. The nonzero weights. By taking the smallest weight vector, a largest input
eigenvalues of the covariance matrix give the eigenvectors vector, facial image can be handled, and the original shape can
that produce an orthonormal basis for the subspace within be reconstructed with an acceptable error.
which most image data can be represented with a small
amount of error. This eigenvectors are ranged from the largest In our application we consider just the identification. Each
to the smallest according to their corresponding eigenvalues, shape in the training set is transformed into the face space and
and the largest eigenvalue give the eigenvector that reflect the its parameters (angles, directions) are stored in memory as
greatest variance in the image and vice versa. Usually 90% of PCA-bases. Ones, a new shape is given as an input to our
the total variance is contained in the first 5% to 10% of the system, it will projected onto the face space. Then our
dimensions. We can project the facial image onto M’(<< M) algorithm calculates its Euclidean distance εk from all stored
dimensions by computing. shapes. The smallest distance give the desired shape, and the
reconstruction give the similar shape to the original one. This
information’s are used to control the 3-DOF manipulator robot
Ω= … = . (6)
to draw the same shape given by the new input.
Where yi is the ith coordinate of the facial image in the 3-DOF MANIPULATOR ROBOT
new space, which come to be the principal component. The How to control the end-efector of the robot manipulator to
vector vi is known as eigenfaces. For shape recognition we track a pre-define trajectory is one of the most important task
will minimize the Euclidean distance. in robotic. In several cases, the inverse kinematics algorithm is
used to determine the desired rotational angles of each joint,
= ‖Ω − Ω ‖. (7) given the desired position of the end-effector. Another
algorithm, the Direct Kinematics is used in this work.
Where Ωk is the vector describing the kth shape class. If the
euclidean distance εk is less than a predefined threshold, a The Direct kinematics algorithm permit to determine
shape is classified as belonging to the class k. the accumulative effect that come from the set of variables of
each link, in goal to find the position and/or orientation of the
The PCA algorithm can be given by Fig. 2. end-effector. The analysis of the direct kinematics was made
using the Denavit Hartembrerg convention as follows.
Matrix X N dimensional input space
Input data
= , , , ( , )
(8)
Matrix W W=cov(X) =E [(X-m) (X-m)]
Covariance of X =

cos( ) − sin 0
Valuable k 1≥ 2≥ 3≥…≥ N
kth eigenvalue of W cos( ) sin( ) cos( ) cos( ) − sin( ) − sin( )

sin( ) sin( ) sin( ) cos( ) cos( ) cos( )
0 0 0 1
Vector ek ek corresponding to eigenvalue k
kth eigenvector to k


Projection: Ce Lower than input dimension The skeleton structure of the 3-DOF manipulator robot is
New feature vector
given by Fig. 3
Figure 2. Flowchart of the Principal Component Analysis PCA.

396
 The test of the PCA algorithm with the training-data has
been implemented by Matlab 7.6.0, a code source example is
shown

Figure 3. The Skeleton structure of the 3-DOF manipulator robot.

Then the denavit table is obtained as follow.

TABLE I. DENAVIT HARTEMBRERG PARAMETERS.

Joint j
1 90 0
To ensure how well the PCA algorithm can identify each
2 0 0 individual’s shape, we proposed a test-database set that
contain 7 images of same size as training-database 320×240,
3 0 0 the results are given in Fig. 5

After calculation and arrangement of the equations, the

end-effector position vector is given by.

= − + −
+ +
+ (5 + )

= − + + +
+ (5 + ) (9)

= − − + − +
+ − (5 + )+ −
2

SIMULATION AND RESULTS

In the simulation test of the algorithm proposed to control
the 3-DOF manipulator robot by one of the most used shape
recognition techniques, the PCA. A 42 images of 7 different
shapes (square, triangular, rectangular, …ect) as shown in Fig.
4 are considered as a training-database. Each image has a size
of 320×240, that create a covariance matrix W of size
76800×76800.

Figure 5. The test results of the shape recognition algorithm: (a) the test-
databases shapes, (b) the shapes constructed by the PCA algorithm.

As we observe the proposed algorithm well recognize the

corresponding shape (Equivalent Image) of the test shape
(Test Image).
Using the GUI tool of the Matlab 7.6.0, we simulate the 3-
DOF manipulator on 3-D space see Fig. 6.

Figure 4. The training-database of the different shapes.

397
 An example of the source code of the Robot Control
algorithm is given.

CONCLUSION
A shape recognition technique represented by the PCA
algorithm to control a 3-DOF manipulator robot was proposed.
Figure 6. 3D simulation of the 3-DOF manipulator robot. The simulation of the PCA algorithm using a training-database
and test-database of different shapes was tested and the results
To control the 3-DOF robot, a new picture was injected of confirm its accuracy to recognize and reconstruct the desired
a new shape that product a new image. This image shape. The modeling of the 3-DOF robot by the direct
transformed into the face space, and its components are kinematics equation was applied. The GUI tool in Matlab
considered as an input data to the robot controller. By using 7.6.0 was used to simulate the manipulator in 3-D space. An
the information gotten from the PCA algorithm as the image of a new shape was inserted and used to move and
orientations and directions extracted from the new inserted control the robot to ensure that the end-effector will draw a
shape, the 3-DOF manipulator has controlled to ensure our similar shape to the new inserted shape. The simulation results
task goal. Therefore, the end-effector draw a similar shape to were obtained and clearly show the good performances of the
the inserted one, the simulation of the robot control algorithm proposed algorithm.
has been tested and the results are given by Fig. 7. (a), (b), and
(c)
ACKNOWLEDGEMENT
This research was supported by the MKE(The Ministry of
Knowledge Economy), Korea, under the CITRC(Convergence
Information Technology Research Center) support program
(NIPA-2011-C6150-1102-0011) supervised by the
NIPA(National IT Industry Promotion Agency).
REFERENCES
(a) [1] J.A. Velarde-Sanchez, S.A. Rodriguez-Gutierrez, L.G. Garcia-
Valdovinos, and J.C. Pedraza-Ortega, “5-DOF Manipulator Simulation
based on MATLAB-Simulink Methodology,” in 20th IEEE Electronics,
Communication and Computer (CONIELECOMP), International
Conference on, Cholula, Mexico, 2010.
[2] A.T. Matthew, and A.P. Alex, “Eigenfaces for Recognition,” Journal of
Cognitive Neuroscience, vol. 1, MIT Press Cambridge, MA, USA,
1991.
[3] A.P. Alex, M. Baback, and S. Thad, “View-Based and Modular
Eigenspaces for Face Recognition,” IEEE Conference on Computer
Vision & Pattern Recognition, USA, 1994.
(b) [4] S. H. Murphy, J.T.Y. Wen, and G. N. saridis, “Simulation of
cooperating robot manipulators on a mobile platform,” IEEE
Transactions on Robotics and Automation, pp. 468—478, 1991.
[5] M. Nakadokoro, S. Komada, and T. Hori, “Stereo visual servo of robot
manipulators by estimated image features without 3D reconstruction,”
Systems, Man, and Cybernetics, IEEE SMC '99 Conference
Proceedings, IEEE International Conference on, Tokyo, Japan, 1999.
[6] C.M. Giulia, C. Christian, L.S. Emanuele, M. Giovanni, and C.C. Maria,
“Principal components analysis based control of a multi-dof
underactuated prosthetic hand,” Journal of neuroengineering and
rehabilitation, Italy, 2010.
[7] R.J.J. Shing, S.C. Tsai, and M. Eiji, “Neuro-Fuzzy and Soft Computing:
(c) A Computational Approach to Learning and Machine Intelligence,”
Figure 7. The test results of the Robot Control Algorithm: the shape drawn Prentice Hall, 1997.
by the end-effector of the 3-DOF manipulator robot corresponding to the [8] K.I. Diamantaras, and S.Y. Kung, “Principal Component Neural
desired shapes. Networks: Theory and Applications,” Wiley-Interscience, 1996.

398