NT AND ANALYSIS
out how soil responds to
be important in order to
ometric qualities of three
n S-VHS video recorder;
analysis procedure based
VHS video recorder and
and deformation analysis
cal experiments.
images in a centrifuge
ge data storage capacity
> of sufficient quality for
tion analysis. This paper
oO assess the geometric
target location algorithm
> video recording, and J-
r the on-line tracking of
rmation analysis without
cussed.
QUENCES WITH A S-
ER
archive image sequences
er can then be used off-
interest. The use of a S-
pe has the advantage of
] the ability to capture
e. The accuracy of target
yaluated by Hoflinger and
). Under ideal conditions,
r than one of tenth of a
black targets are inserted
cation errors are typically
1 imaged target is small,
two consecutive directly
nna 1996
typically occupying 3x3 to 5x5 pixels. The target intensity level
is also low, normally 50 - 150 grey scale levels above the
background. Figures 1 and 2 illustrate an example of the
differences between simultaneously frame grabbing and tape-
grabbing images of a stable soil sample. Figure 1 shows the
discrepancy vectors between target locations measured on two
consecutive images recorded by an S-VHS recorder and
subsequently captured from tape by an EPIX frame grabber.
Figure 2 shows target location discrepancy vectors between two
consecutive images grabbed directly from the monochrome
camera with the EPIX board. Since the target array and camera
were stable during these two image sequences, the difference in
magnitude between the vectors can be attributed to degradation
caused by the S-VHS recorder. Comparing the two figures, the
geometric degradation of target locations from the S-VHS
recorder is up to 0.5 pixels, about four times larger than direct
grabbing.
To rigorously verify the observed metric performance of the S-
VHS system, a number of laboratory experiments using
different types of target and illumination have been carried out.
Both retro-reflective and black conventional targets of different
sizes were used. Each target type was attached to a board, the
board being white for the simple black targets and matt black
for the retro-reflective targets. The boards were fixed in turn to
an optical bench. A Pulnix TM-6CN camera was fixed on the
bench and an adjustable source of light placed behind the
camera. A JVC, SR-S368E S-VHS recorder was used both to
record the sequential images and to play back the tape for image
grabbing. A comparison between target location errors obtained
using the S-VHS recorder and direct image grabbing based on
different target type, size, and illumination has been made. In all
cases, a centre weighted algorithm (Chen & Clarke, 1992) has
been used to measure target image co-ordinates. Results are
illustrated in figures 3, 4, 5, & 6.
o
8
°
2
8
m Large target m Large target
0.25 + © Small target © Small target
o e o
Sa S »
ü 3 2
RMS location errors (pixels)
°
3
RMS location errors (pixels)
E
a
1
=
—-—
T
e
8
VAT
40 ©0 80 100 120 140 160 180 200 220
Target intensity (grey level)
o
50 100 — 150 200 250 300
Target intensity (grey level)
Figure 3 RMS location error of retro Figure 4 RMS location error of retro
targets using S-VHS targets by direct grabbing
Figures 3 and 4 show the RMS x co-ordinate location error with
different target image intensities and sizes for retro-reflective
targets captured using each method. Two target sizes were used:
small (imaged at about 3 pixels in diameter) and large (about 10
pixels in diameter). Each image acquisition and location
procedure was repeated 100 times under stable conditions and a
RMS target location discrepancy computed for each set.
The results shown are dependent on the experimental set-up
used, but the differences between results provide some clear
indications. It can been seen that in the case of directly grabbing
retro-reflective target images differences in target size have only
a very small influence. Target location repeatability is better
than 1/20th pixel. In the S-VHS case, if the imaged targets have
good contrast and are of reasonable size, RMS locations of the
order of 1/20th a pixel can be obtained. However if target size
or contrast is lowered the RMS location error increases rapidly.
0.30 0.30
m Large target m Large target
© Small target
© Small target
0.25 À
o
s
o
o
8
o
8
1
o
a
o
$
RMS location errors (pixels)
o
a
o
5
RMS location errors (pixels)
o
8
&
1
o
8
&
o
8
T T T T T
40 60 80 100 120 140 160 40 80 80 100 120 140 160
Target intensity (grey level) Target intensity (grey level)
Figure 5 RMS location error of Figure 6 RMS location error of
conventional targets through tape directly grabbed conventional targets
In all S-VHS cases errors in the y direction were found to be
lower. This suggests that line-jitter is a significant effect.
Figures 5 and 6 show the RMS x co-ordinate location
discrepancy given different intensities and sizes of black
conventional target using both direct grabbing and S-VHS
recording. It is difficult to set up a high contrast target image
using black targets against a white background. Consequently
the performance of the conventional targets is much poorer than
that of retro targets. RMS location error can be up to 0.3 pixels
with a maximum target location error of 1 pixel. Results from
these laboratory experiments were found to be in agreement
with practical geotechnical applications.
The degradation in the S-VHS data are caused by increased
image noise and timing errors introduced during the analogue
recording process. The analogue recording process also
introduced a reduction in contrast, especially if the targets were
small and of low initial contrast. Tape recording is also
undesirable because a large quantity of data storage is still
required when selecting and grabbing images on playback.
Repeated use of the tape in this way will give rise to some loss
of target location accuracy. To summarise, for small targets of
low image intensity, occupying less than 3x3 pixels and less
than 100 intensity values, image measurements from the S-VHS
recorder gave rise to target location errors of up to one pixel.
Direct image grabbing, under the same circumstances, resulted
in a maximum measurement noise of less than 1/10th of a pixel.
3. EVALUATION OF THE EFFECT OF JPEG IMAGE
COMPRESSION ON TARGET LOCATION ACCURACY
To overcome the restrictions of using a S-VHS recorder in
centrifuge tests, on-line image compression schemes have been
considered. Two generally accepted standards are MPEG
(LeGall, 1991) and JPEG (Wallace, 1991). These methods have
become popular in recent years because of their high
compression ratio, optimisation for visual quality, and potential
for hardware implementation. MPEG is a video compression
algorithm, which relies on two basic techniques: block-based
motion compensation for the reduction of temporal redundancy
and transform domain-based compression for the reduction of
spatial redundancy. The MPEG standard is designed for the
compression of sequential images which have high redundancy
between successive images. However because of hardware
availability and the typical usage of individual images in
photogrammetric applications, only the JPEG compression
method has been evaluated at the time of writing.
At first sight, the JPEG image compression algorithm proposed
by the Joint Photographic Experts Group (JPEG) offers a viable
way of accomplishing image compression tasks. Framegrabbers
with JPEG hardware are widely available commercially
providing imaging rates ranging from 2 to 25 frames per
second. Compression ratios are typically user selected according
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996