ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision, Graz, 2002
the others at the same time. Those four images can be
integrated into a composite image or viewed individually. The
CAMIS sensor has been used in multispectral imaging and
mapping purposes by mounting it in an airplane with GPS and
INS systems. These auxiliary sensors provide very good
position and attitude data for stabilizing the subsequent bundle
block adjustment. The calibration procedure required a number
of steps and they are summarized below.
2.1 Site preparation
The idea of the calibration was to layout some targets in the
object space, locate them accurately, and acquire an image of
those targets by the sensor. Then, we relate the coordinates of
the targets in both systems, image and object space, in order to
obtain the camera parameters. So, the procedure starts by
setting up the calibration site. First, we designed the targets to
be cross shapes so their center positions will be obtained very
easily. Then they have been laid out in an “X” pattern that
allows us to recover the needed geometric parameters and
systematic errors as shown in figure 2. Those targets were
placed on an almost flat service wall and the sensor position
was located around 8 meters away from that wall (at that
distance we could use the “infinity” focus position. In order to
register the objects in the scene, an object space coordinate
system was established at the site and two other instrument
stations were marked to use in the measurements.
w* Lu
a “nr
Figure 2. Targets layout
2.2 Measurements and adjustment in object space
Two three arc-second theodolites were mounted at the
referenced stations and were used to measure directions to all
relevant objects of the network: target centers, theodolite
locations, and camera case monuments. The origin of the object
space was chosen to be theodolite one at station one at the right
of the sensor position. Many manual measurements were made
of the camera physical layout, using machinist calipers. The
lenses were also placed on an optical bench for determining the
locations of the nodal points. A cross section of one camera is
shown in figure 3. This was needed to locate the camera front
nodal point with respect to the camera body, which would be
located in the network by theodolite observations. The spacings
between the wall targets were measured with a steel tape.
Having all these observations, we end up with an
overdetermined system of equations. We developed a bundle
program to simultaneously adjust the theodolite and distance
observations. As a result of that, we determined all of our
targets and camera stations in the referenced coordinate
system. The next step was capturing images by the sensor(s).
Front nodal
point |
ES nmi k<-> Separation = 2.5
| Camera | LI
; Body
I ;
I |
|. CCD
: Arrav
I
! |
b ass An | Note: front & rear
| = 16 mml| nodal points are
| | “reversed” from
Rearnodal their usual position
point
Figure 3. Cross section of one camera showing the
lenses, rear and front nodal points
2.3 Capturing images and obtaining image space
coordinates
Before the measurements, the sensor was mounted on a leveled
plate fixed on a survey tripod. In this sense, the exposure
stations were fixed and predetermined to an accuracy of a few
millimeters. In this step we tried to simulate the real working
conditions by setting the lenses to the “working” infinity focus
position. Images were viewed after captured to verify
acceptable radiometry. With the band pass filters our
illumination setup was just sufficient to produce acceptable
image definition for the targets. In the future, stronger light
sources would be used to allow more flexibility. After this
step, the laboratory work ended and the processing procedure
started.
2.4 Image space calculations
Once the images were captured, we ran a cross correlation
matching program to get rough approximation of target
positions in the image space to within a pixel. The cross
correlation matching function works by computing the
similarity between two same sized widows (Mikhail, Bethel
and McGlone, 2001; Mikhail and Ackerman, 1976). One
window patch contains the ideal target and the other contains a
window from the image. In general, a matching problem is a
key algorithm for other applications and image analysis.
Despite the fact that, the cross correlation matching results
showed that we are only away from the exact position by a
pixel or less, we needed more accurate and precise methods to
guarantee the sub-pixel precision. This level of precision is
necessary for a camera calibration problem. Least squares
matching (LSQM) is very adequate technique for this purpose.
LSOM utilizes the first derivative of the intensity in both x and
y directions to obtain the best correspondence and the exact
matching can be reached by moving one window with respect
to the other one (Atkinson, 1996). Some obstacles, such as
radiometric effects, were faced and solved by modifying the
algorithm. As a result of this step, the image space coordinates
for
techi
Figu
mat
2.5
First
calib
mod
factc
expli
in «
para
an o
initié
unce
algoi
para
dece
curv
para:
tabu
In o1
appr
obta:
refin
3.1
Cros
corré
appr
man
noise
geon
Atki
appr
pixe
The
matc
pixe
corre
max
Usu:
matc
as sl