DEN ER
ristics
e ——M——— ÓáÓ————
points in the
lage.
— —— ——
the distance
nera and the
all points in
the whole
the whole
e pixel size.
A variety of
S have been
ese tests have
utages due to
, and loss of
position and
a stream. The
is to exploit
accuracy of
el of a few
istellation is
for: a. truck
n figure (6).
vas generally
process by
'ame, i.e. by
b
R m (0- For
' system. In
SAT system,
> at the level
b tests were
Department.
In this test the Litton 90-100 was tested on a rotating
platform which gives independent attitude with an accuracy
of 3 arcsec. Figure (7) shows the difference between the INS
attitude parameters and those produced by the rotating
platform. The figure confirms that the short term INS attitude
errors are of the order of 10-20 arc seconds for roll and pitch,
and 30-50 arcsec for azimuth. This results in errors of 1 cm
or less for points 50 m away. Another test is planned to
confirm such accuracies under survey conditions.
Os
e
Azimuth
— N UU | uA
Co. «0 Oso
Angular Accuracy (arcsec)
©
-40 -20 0 20 40
Angular Velocity (deg/sec)
Figure 7: INS Attitude Accuracy as a Function
of Angular Velocity
The updated INS positions will be mainly used for
interpolating camera coordinates and detecting GPS cycle
slips. Inertial systems combine very high short-term
accuracy and high data rate, and, thus, are ideally suited for
these tasks. The accuracy of using INS positions in
interpolating the camera coordinates depends on the
position drift rate between updates. Figure (8) shows a
typical drift behavior between update intervals of 20 sec in
static mode. The figure confirms a drift of 3 cm/10 sec in
latitude. This means that for a GPS updating interval of 1
second, the INS position accuracy is at the level of few
millimeters, better than the GPS accuracy for this time
interval.
0 20 40 60 80 100
Time (sec)
Figure 8 : INS position drift in static mode
The idea of detecting cycle slips is based on comparing the
measured integer cycle number with the number predicted
from the INS derived position. If a cycle slip occurred, the
respective carrier phase can be corrected by resetting the
integer value to the integer value closest to the predicted
value. Obviously, the method will work best for short time
intervals. In environments were cycle slips are frequent, as
for instance inner cities, the success of cycle slip detection
depends mainly on the INS position accuracy. In order to
correct cycle slips at the one cycle level, INS positions must
be accurate to 0.5 cycle, or better, for the outage interval.
The accuracy of the positions derived from INS are a function
of the update rate and the vehicle dynamics. Detailed tests are
currently under way to determine the time interval in which a
stand-alone INS will give 10 cm accuracy.
Camera errors contribute to the error budget through
i E e
pointing errors in r . The accuracy of r depends on the
distance between the target and the cameras, the size of the
target, and the accuracy of pin-pointing the target. To check
c :
the accuracy of r we use the orientation parameters obtained
from the bundle adjustment to determine the object
coordinates from the image coordinates of stereo pairs. This
is equivalent to positioning points with respect to the
stereo-pairs, independent of GPS/INS data. The test was done
by establishing a test field of ground control points (GCP)
and taking some images for this test field in kinematic
mode. Figure (9) shows the difference between the distances
computed from the computed local coordinate and from the
GCP coordinates. The figure shows that these errors are
distance dependent, as expected, and reach a magnitude of
10-15 cm for objects 50 m away from the van.
20
= |
310 5
eO
er : I
£e,
.& 0 | E =
<-l a
= =
-20 1 | L 1 A 1 3 1 L
0 10 20 30 40 50
Distance from camera (m)
Figure 9 : The accuracy of measuring distances
with CCD cameras
The georeferencing equation (5) contains four unknowns (3
coordinates and one scale factor) in three equations. Having
a second camera of the same scene will add three extra
equations and one unknown for the same point (1). A least
squares solution of the space intersection between the two
rays from the two images is computed. With a one second
image data rate and a speed of 60 km/h, the van moves 16 m
between exposures, and it is expected that the same object
will , therefore, appear in four consecutive images. This adds
extra redundancy and geometry to the spatial intersection
problem. Figure (10) illustrates the error reduction resulting
from adding redundancy and geometrical constraints to the
minimum configuration. Figure (10a) shows the geometry of
the situation with the target point visible in four pairs of
images. If the four individual pairs are used for positioning,
the upper curve in Figure (10b) is obtained. It shows errors
which grow rapidly with distance from the target and are
considerably larger than expected. This is almost certainly
due to stability problems of the camera mount in this
specific run. When combining all images to determine the
target point, the positioning error is reduced to 0.25 m
which is well within the required accuracy. It should be noted
that the combination is not just a weighted mean of the four
individual determinations but indirectly introduces the
coordinate differences between exposure stations, measured
by GPS/INS, as a constraint into the equation. Instead of
247