ity to use
sequence.
means to
1 object-
-Oriented
|. But the
phasized.
lates well
formats.
[12], the
that has
to DXF,
e-data as
easily be
me very
rammetry
chitectural
1. Remote
pp. 771 -
c stations,
n Digital
AD Meets
Forms" for
er at the
gration of
aint-based
surement.
‘utures'95
ation in
3D Feature
netry and
omputers,
ynference,
e, NJ:LEA
udien zur
lett-Cotta,
iction in
insparenz,
e. Design,
MA: MIT
1al Reality
cification.
1dex. html
REAL TIME TRACKING OF A DYNAMIC OBJECT
Amir Saeed Homainejad
Department of Geomatics
The University of Melbourne
Parkville 3052 AUSTRALIA
Telephone : +61 3 344 6806
Facsimile : +61 3 347 2916
Email : homain@sunburn.sli.unimelb.edu.au
KEY WORDS: Template, Real Time, Tracking, Automatic, Stereo Matching
ABSTRACT
This paper outlines a method for the automatic tracking of a dynamic object in real time. The method is able to
position the visual system concurrent with the processing of the tracking. The visual system moves approximately
along the Z axis whilst the object moves throughout the scene from the left to the right. A program in C language
has been developed to position the visual system and to provide stereo matching in 30 ms. In addition, the
program is able to track the object in less than 30 ms. The program is robust, reliable, and able to locate the target
with sub-pixel precision. This paper will explain and describe the method and present some experimental results
that achieve real time tracking.
1. INTRODUCTION
Computer vision is a technique that includes a wide range of
image processing and statistical pattern classifications, and is
used for vision perception. Computer vision has had a vast
amount of applications; eg, pattern recognition, pattern
detection, and pattern extraction. This paper will focus on
object tracking which is a process that involves the above
mentioned applications as well. Fundamentally, visual sensors
are the main sources for this type of tracking because they
enable a visual feedback into the processing. In addition, they
provide information about the object and the background, and,
hence, are a facility for a robot and/or an autonomous vehicle to
calibrate and position itself with the background.
Visual systems based on the kind of the sensor are classified to
two major groups: passive and active systems. Active systems
supply their own source of energy to illuminate features of
interest. Two examples are sonar and radar sensors. In contrast,
passive systems require external light sources, such as sunlight.
Such systems as tracking a manipulator robot, navigating a
mobile robot, self controlling manufacturing require different
approaches in finding reliable methods for tracking. The
methods vary, based on the use of mono cameras or multi
cameras, the use of stable visual systems or dynamic visual
systems, and the use of active/passive systems. The major
issues that affect the mono camera method are:
* singularity,
* the exiting noise on the images that degrades the
computation of the depth of a point,
* a small error in the measured coordinates of a point in the
image results in a large error in the depth.
The first problem can be compensated, if at least three targets
on the object are detected (Papanikolopoulos et al. 1992). The
two latter problems can be overcome by imposing certain
constraints. For example, imposing a constraint whereby the
243
camera or object is moving along a defined axis will reduce
depth error. In contrast, the main disadvantage of multi camera
methods is slow processing. This problem can be overcome by
reducing precision. Such the methods are used by Tistarelli et
al. (1991), Elfes (1987), and Waxman (1987).
This paper will present a robust and reliable method of tracking
a dynamic object that satisfies both aspects of fast processing
and precision. In a previous experiment (Homainejad and
Shortis 1995c), a stable multi camera setup was used to track
automatically a dynamic object. This paper describes a new
experiment, in which it is sought to answer to some basic
questions. The questions are:
e At what speed can the system track a dynamic object?
e Is the program able to reposition the vision system, when
the system is relocated?
e To what extent is the method robust and reliable?
To address the questions, the visual system was set up at a point
in front of the test field, and a stereo image was acquired. The
test field consisted the templates as explained by Homainejad
and Shortis (1995a). The program initialised the processing as
described by Homainejad and Shortis (1995c) whereafter; the
visual system was moved to a new position. The visual system
was moved approximately along the Z axis. The base line was
not perpendicular to the Z axis nor coincides with the X axis.
Therefore, the program should define the position of each
camera separately. In the new position the visual system
acquired some stereo images from the test field whilst a
dynamic object was moving from the left to the right. Figure 1
illustrates the chart of the visual system and the test field.
The organisation of this paper is based on the description of the
system and the explanation of its advantages. Furthermore, the
paper will give the ways whereby real time was achieved. The
next section will evaluate and explain the system. In addition,
different approaches of programming to achieving the goals of
the experiment will be assessed in Section 2, and Section 3
presents a conclusion about the test.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996
