periment is
ng is carried
are in TIFF
processing.
| Were tested
ple, reading
s 150 times
°d to use the
at increases
acters in the
nd the right
| characters,
ch, a goto
e delay was
iced by the
time delay
mmand was
| the fourth
ion, 20000
llt, the time
and stereo
proach. The
seudocoded
roach.
roach.
yaches. The
and robust
because the program without any mistakes and errors
recognises a control point in the right image that matches with
the control point in the left image. The processing is precise
because the program measures the coordinates of the control
points with sub-pixel accuracy. In addition, by using the
intersection method the coordinates of the control points in the
object system are measured with accuracy less than 1 mm. The
RMS error of the computing coordinates on the images is sub-
pixel. Table 3 shows the differences between the measured and
computed control points in the left and right images.
TIME DELAY
APPROACH
I(GOTO) — sl
ps
6 (-2.1E-06)
ES
(-2E-07)
C14 (-581E-05) -1.50E-06 0
Table 3: The differences between the measured and computed
coordinates of control points for both the left and the right
images. The unit is pixel.
Once the processing of the camera position and the stereo
matching are completed, the program continue to track the
dynamic object. The object is considered a rigid body.
According to Hunt (1987), a rigid body does not change its
shape or its size; it does not stretch, compress, twist, bend, or
deform. The object in this experiment is a plastic bottle with a
waist. The waist depth is about 2 cm. Figure 3 shows the object
in the test field.
Figure 3: Illustration of the object in the test field.
The strategy of Homainejad and Shortis (1995c) is used in this
experiment. A very important issue for the program is to define
a point upon the object that is recognisable in both the left and
the right images. The colour of the object that is used in this
experiment is white and light grey and is very similar to the
background colour. The program for recognising a common
point in the two images should apply a sharpness mask to the
images and this processing increases the time delay. For
extracting the pattern from the right image that matches the
pattern in the left image, an empirical threshold is used by the
program. The program defines a weight based on the histogram
of a common area in the left and the right images. According to
the weight and the area that already is extracted from the left
image, the area of interest in the right image is recognised and
extracted. Figure 4 demonstrates two areas of interest in the left
image and its matched area in the right image. The processing is
very precise and reliable; however, recognising the area of
interest is very difficult. The ability of the program for
recognising different patterns is tested by selecting different
points in the waist and other area of the object. Program
recognises and extracts two different points in the waist area
and other area upon the bottle. Table 4 presents the coordinates
of the two points in the waist and a pattern upper than waist
area. These points are in a very difficult areas that even an
expert operator can not simply extract and recognise those
points. But the weight enables the program to recognise and
extract even that difficult points without applying any mask and
human supervision. The weight is not constant for a sequences
stereo images because lighting is not constant for all stereo
images. In other words, each stereo image has a weight. Table 5
demonstrates the weights for three stereo images. When the
program defines a weight for each stereo image, the time delay
is increased for each stereo image by 30 ms. The time delay for
extracting a common area from a stereo image is between 30 ms
to 60 ms. This delay of time could be reduced if the weight is
constant for all stereo images. When considering table 5, the
point in the waist area has 10.6 mm difference in depth with the
other point.
POINT X Y Z
The waistarea 6.128 3.968 4.576
245
The second area 6.123 3.972 4.566
Table 4: The coordinates of two points upon the object.
CL ki Sh EME
weight 1.2 o 1.33
Table 5: A list of three weights for three stereo images.
The first The first
and and
the second the second
ponts points
IEHTIMAGE EIGHIIM AGE
Figure 4: The left and the right images of a stereo image with
areas of interest for matching.
3. CONCLUSION
This paper presents a method for tracking of a dynamic object
when the stereo vision is relocated. The method is reliable,
robust and precise. A program in C language is developed for
automatic stereo matching and positioning. In addition,
program can automatically define and track a dynamic object.
The lighting is the main issue in the processing, but the applied
weight reduces significantly the lighting issue. The time delay
for each processing is, at best, in the order of 30 ms. The
majority of time delay is related to the image registration in an
array. It seems that a special image library is necessary for
digital photogrammetry that it reduces the time delay. The
image library should satisfy the application of real time
processing for digital photogrammetry. The TIFF library is very
reliable and trustable. Table 6 presents the time delay for
different processes.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996
