Information Technology Reference
In-Depth Information
Commodity cluster computing
is a paradigm for parallel/distributed com-
puting that proposes using a reasonable large number of available computing
resources to perform parallel computation at low cost, since those resources
are suppose to be already available for non-high performance computing tasks
(i.e., oce equipment, educational computers, personal notebooks) [11]. Com-
modity cluster infrastructures, usually built by integrating low-cost personal
computers and other devices using a local-area network (LAN), have been used
since the mid-1990s for solving a wide range of problems in different application
domains, including the scientific and industrial ones [12]. They are character-
ized as
Beowulf clusters
, from the pioneering work by Sterling and Becker at
NASA [13].
In this work a parallel implementation of yaw angle determination is pre-
sented. This system runs in a ground station which receive images from the
camera mounted in the quadrotor, performs the yaw angle estimation and sends
the result again to the quadrotor. The implemented parallel application was
evaluated over a Beowulf cluster using two benchmark sets: the specific public
data-set by Lee et al. [14] and a benchmark set built with own images. In order
to mitigate the latency of the LAN a multilevel data decomposition is proposed
using a pyramidal hierarchy called Master/Taskmaster/Slaves. The experimen-
tal evaluation indicates that significant reductions in the execution time are
achieved when using this hierarchy instead of the classical Master/Slave. The
speedup analysis demonstrates that using a multilevel data decomposition an
improved of 2
can be achieved.
The paper is organized as follows. Section 2 presents the basis for camera
orientation estimation and particularly for the case of yaw angle estimation, be-
sides a brief explanation of the spectral features. Section 3 gives an overview
of the proposed approach to implement a parallel orientation estimation algo-
rithm. The experimental analysis of the proposed parallel algorithm is reported
in Sect. 4 using a precise public data set of images and orientation, as well as us-
ing own images with an ad hoc orientation method for reference. Finally, Sect. 5
presents the conclusions and formulates the main lines for future work.
×
2 Homography Based Yaw Angle Determination
Given two images of the same plane taken with a camera from different points of
view, each characteristic point
m
a
belonging to image
i
a
and their corresponding
m
b
belonging to image
i
b
are related by a plane induced homography
H
ba
[8]
such that
m
a
=
H
ba
m
b
.
The homography
H
ba
represents the spatial translation and the rotation be-
tween the different camera positions. Particularly, if camera movement is limited
to a plane parallel to the plane containing the characteristics and the rotation
is around a normal vector to this plane, the homography
H
ba
is defined by
H
ba
=
R
z
t
0
1
(1)
