d to metric
esigned for
ric quality.
, and there
the above
leo images
ile changes
ximate data
size of the
3 distortion
zontal and
uced from
number of
n Fig. 4(A),
horizontal,
radial and
stimate the
other lens
ns by using
the pixel
vas 3.5um,
@8-Blackt
@8-BlackH
Os-wntel
OS-whiteH
@F-Biockt
BiF-Biack-H
QF-Lon-L
OF -Light-H
r Speed
@ Vertical
Directioion
B Horizontal
Direction
A Diagonal
Direction
® Vertical
Directioion
8 Horizontal
Direction
A Diagonal
Direction
3)
| no friction
stand, but
The target
above the
be 2.8m/s,
image blur
was improved with the shutter speed. The relative motion
during exposure was 30 rows In Fig. S(A), 12 in (B), and
3in(C) Image (D) is one later field(1/60s) of Image
(C). Vicek(1988) suggested that the relative motion
during the exposure should not exceed 0.5 pixel for a
clear image. Although we have not reached 0.5 pixel, we
expect that relative motion lower than 0.5 pixel will
provide a clean image without blur. To lower the relative
motion, a short exposure is mandatory.
(A) Shutter 1/100s
(B) Shutter 1/250s
Fig. 5. Falling Target Images at Various Shutter Speeds.
5-2 Dynamic Monitoring
Source Sequential Imaging and Pre-Processing: The
left and right video images were recorded on each S-
VHS VCR. Then the VCR was connected to the PC
installed frame grabber. We planned to use the
stopwatch’s 1/100s digits as a code for synchronization.
However its letter size was too small to identify it when it
was attached to the control frame located close to the
rear of the car. There was no other alternative than using
the frame advancing function(1/60s) of the VCR. Before
capturing the real sequential image, significant
scenes(e.g. lights on and off, door open and close) were
designated as milestones. We counted the number of
fields between each milestone several times. It does not
exceed 2 field's(1/30s) difference during 10s. In this
context, we captured every 4" field(1/15s), comprising
Synchronized sequential images for monitoring the
movement of a car. The PC works on a S-VGA display
board, and the resolution of the images captured using
the frame grabber is 788(H)*468(V). Each captured
image was saved in TiFF format, its file size is 1.1Mb.
This image is a 24-bit Band-Interleaved-Pixel(BIP)
format, and was converted into Band Sequential(BSQ)
images in the "IDRISI for Windows" format. There were
no significant image quality differences between the S-
VHS and VHS video cameras when the images were
recorded by S-VHS VCRs using the S-Video format. 77
pairs of sequential images were digitized from S-VHS
VCR. Image coordinates of three points in each image
Were processed to determine their object coordinates.
Sub-Pixel Target Coordinates: The image
coordinates of the targets were generated from the
preprocessed video image in four-step process:
* Noise removal using median filter.
e Searching window centered a target.
e Reclass based on a threshold.
e Determination of the target center coordinates using
equation(1).
These processes for the sequential images were
conducted by batch job using the macro command of
IDRISI.
The enhanced images using median filter showed more
condensed gray value dispersion than the original images
in the histogram, and more ideal circle appearance.
Target windows were detected manually, while the
threshold was determined from the histogram. The
change from background to target was clear in the
histogram, and generally it was very close to the
threshold [(maximum + mean of gray value)/2]
suggested by Wong & Ho's(1986). In order to see the
significance of the threshold, the target coordinates were
calculated using another threshold (original threshold -
4), the differences of target center coordinates obtained
by each threshold were within 0.1 pixel. In Fig. 6., (A) is
the original target image, (B) its filtered image, and (C) its
reclassed image. It was known that image enhancement
causes target shift to some extent. In this test, the mean
difference between two was nearly zero, and the
maximum difference was 0.02 pixel in 20 targets.
(A) (B)
Fig. 6. Images of Circular Target
3-D Object Coordinates: For 3-dimensional tracking of
the targets in each image sequence, we prepared image
coordinates of all control points and targets in the first
image. Under the assumption that the control frame was
fixed, the target coordinates were only prepared in the
remaining image sequences. This tracking scheme
reduces the work from obtaining image coordinates to
running the program. As described in section 3-2, DLT
and UNBASC2 could be used to derive 3-dimensional
object coordinates using non-metric image coordinates.
While DLT affords pixel unit as a input data, UNBASC2
requires metric units. Pixel units may be considered as
comparator coordinates having different scales for each
axis, but UNBASC2 failed to obtain the object
coordinates. Therefore, the pixel sizes of the two cameras
were derived by trial and error. UNBASC2 then provided
more accurate results than DLT in the test image.
However, DLT was used for this experiment because it
can cope with sequential data that may recourse to the
same position in another sequence when above tracking
scheme was adopted.
It is well known that the imaging geometry plays a great
role in determining the accuracy. When the imaging
geometry is ill conditioned, the accuracy of DLT
becomes worse when increasing the number of
parameters. Fig. 7 shows the average root mean square
errors of DLT(12 unknown parameters) for different B/D
ratios. Just 20 control points were used as image
coordinates. In the adopted coordinate system, the base
157
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996
