The simulation can also be used for real world situations. Let us
say that we have a target area we would like to measure. The
distance to the closest relevant detail of the area is 350 meters
away and the distance to the background of our target area is 400
meters. This kind of situation can occur for example in a case of
a cityscape where we would like to measure some specific area
within a city. Now, if we only have a normal photographic
camera stand and would still like to capture a panoramic image
with a long focal length lens to achieve high resolution from such
a long distance, we need to consider the error caused by the offset
projection centre. If we mount the camera on the stand from the
camera base, the projection centre is, usually, located inside the
lens. In the case of a long focal length this distance can be many
centimeters.
For example in the case of a 600mm lens, it is safe to assume we
could have approximately 200 to 300mm offset in the Z-
direction. We can also assume 10 to 20mm offset in the X-
direction if we do not calibrate our rig. Using a 35mm format
camera the diagonal FoV is 4? 10' while the required FoV to
cover our target is 8? 10'. If we use 50% overlap, we need to
capture five photos. The resulting errors are presented in Table 1.
The Z-direction offset is 250mm and 15mm in the X-direction.
Offset in the Y-direction is set to zero and the camera is only
rotated horizontally.
Angle -4.085° -2.005? 0.075? 2.155? 4.235?
Point 1 0.4mm 0.9 mm 2.2 mm 3.5 mm 4.8 mm
Point? 4.8mm 6.1 mm 7.3mm 8.6 mm 9.8mm
Table 1. Errors at the second target plane with different bearing
angles. The distance to the first target plane is 350 meters and
400 meters to the second target plane. Z-direction offset causes
Points 1 and 2 to produce different errors with different angles.
4. DISCUSSION
According to simulation, perspective errors due the eccentric
panoramic frame-image acquisition are relatively small if the
distance between camera system and objects is long enough.
Even if we have examined only errors in the object space, this
error is proportional to errors at the image space. Therefore, after
a threshold distance, the maximum error at the image space can
be tolerable, for example less than one pixel. For example, Wei
et al. (1998) found similar conclusions in their research. For
seamless stitching of sub-images, such small errors at the image
space are advantageous. However, with some limitations we are
able to have relative small error also for targets that are closer
than previously mentioned threshold distance.
Overall, the perspective error caused by the projection centre
offset obviously weakens the geometrical quality of the
panoramic imaging system. However, if we only consider a
shallow area of interest, the error caused by the perspective error
is smaller than if we consider the whole image area as a whole as
shown in Figure 10. This gives us various alternatives if we are
only able to acquire images of sub-optimal quality regarding the
panoramic geometry.
The first possibility is to use the images as they are, in a
photogrammetric image block. Unfortunately, in the case of a
long focal length this might cause problems due to correlation in
the exterior orientation of the camera station. If we stitch the
photos into a panoramic image, the geometrical exterior
orientation of the image is more stable provided that the internal
geometrical quality of the panorama is robust enough.
If we have a large projection centre offset between the sub.
images, the perspective error of the resulting image can be too
large to permit the use of such a photo in a metric
photogrammetric application. In that case, one possible Way to
use such an image material is to stitch the photos, as shown in
the Figure 10, using only features located within a limited depth
of range. Stitching becomes seamless only within that depth of
range, but we can create more stitched images for different
depths. Examples of such stitched photos can be seen in the
middle and rightmost photos of the Figure 10. The closer
stitching distances result in poorer accuracy or smaller target
depth. On the contrary, a usable depth of range increases if the
offset of a perspective center becomes smaller. The leftmost
image in Figure 10 illustrates how a small offset allows seamless
stitching to a large depth of range.
Even if we have an eccentric sub-image acquisition, we can use
images for photogrammetric scene reconstruction if we use only
those parts of images that are within depth of range in focus
(Figure 10, the middle and rightmost photos). On the other hand,
similar constraints can be used also for finding recommended
areas from the image planes for automatic feature matching and
thus improve stitching of sub-images into full panoramic
mosaics.
Figure 10. Left image shows two photos stitched with nearly
concentric panoramic camera rig showing negligible perspective
error. Middle and rightmost images show two photos with offset
projection centre stitched together using only the points in close
vicinity to the red line shown in the photo.
Our simulation program can be used to estimate the proper usage
of the photos at hand to help to decide in which way to best use
the image material. We use the simulation program in our own
projects, mainly in cases where we have to measure a specific
target located far away. Another application is oblique aerial
imaging. Using the simulation, we can estimate the perspective
error
captu
betwe
can al
educa
One «
we ar
persp
accur
seque
The
intro
imag
came
and |
softw
and :
calcu
parti
This
only
of ra
main
The
imag
calib
some
adva
simu
Brig
navi
Bro
gran
Den
Proc
Syst
Fan.
Rec
Con
200
Fert
Sph
Pan
IEE
Buc
Gas
Cat
Inte
139