positional
itions that
ı typical D-
tion and
ce between
it will not
ing models
ablish strict
mation and
em that we
1g time was
ing System
our current
the shutter
time. As a
ng time has
ON BY
IN
to building-
ystem first
ed image by
| within the
the images
; and tracks
-ulates their
ts. Next, the
ated three-
Finally, the
e building's
) reconstruct
detail.
Omnidirectional image
: Estimate FOE
Calibrated image
| Feature point tracking
GPS data
Offset parameter
(b)
Calculate 3D points
| Reshape surface model GM Re-project surface model] :
X Extract texture :
(c)
Reconstruct 3D model
Figure 5. Process flow of reconstruct 3D model from
omnidirectional Stereo
3.1 Calculating offset between camera position and GPS
system position
To obtain the shooting position of an omnidirectional camera,
the offset between the camera's position and GPS system
position as determined by the GPS antenna must be determined
beforehand. This offset is calculated by using points whose
positions in space are known (SCP') as reference points. Here,
the position of a SCP projected onto a captured image is
manually plotted, and the system position is then optimized so
that the difference between it and the projected position is
minimum. The difference between this optimal GPS system
position and its initial value is taken to be the offset. If the
installation positions of the camera and GPS antenna on the roof
of the vehicle are not changed, an offset parameter determined
in the above way can be applied to all images. In actuality,
nearly the same value for this offset was obtained on calculating
it from multiple images. This demonstrates that, once an offset
has been calculated, it can be applied to all omnidirectional
images as long as the camera's position is not altered.
3.2 Estimating camera attitude
Figure 5(a) shows that part of the process flow that determines
the camera's attitude parameter at the time of shooting.
Depending on the system, it is not uncommon for such attitude
information to be determined by a gyro sensor. Nevertheless,
determining attitude at the correct camera position is
troublesome even with the use of a gyro. For this reason,
determining camera attitude at the time of shooting from an
actual image is the most ideal approach. In this regard, a
property of an omnidirectional image is that camera attitude can
! SCP (Spatial Control Point). An extension of the Ground
Control Point (GCP) concept, a SCP means a point whose
3D position in space has been surveyed. Specifically, it
refers to 3D values of points on a building's corner, side, etc.
(Ishikawa, 2001)
be approximated by the difference between FOE in the image
and the center of the image. Accordingly, FOE can be estimated
by using, for example, a vertical segment of a building in the
image. Estimating FOE in this way makes it easy to calculate
camera attitude at the time of shooting from the captured image
itself. This, in turn, makes it possible to carry out high-accuracy
3D measurements that take camera attitude into account.
3.3 3D measurements by omnidirectional stereo vision
Figure 5(b) shows that part of the process flow that performs
3D measurements by omnidirectional stereo vision. We here
describe the optical projection formula of an omnidirectional
camera and the equations for calculating depth and height
through omnidirectional stereo. Figure 6 and 7 shows the
relationships between the omnidirectional cameras and points
on target buildings. The plane which consists of an x-axis-y-axis
is made into the ground of height 0 among a figure, and z-axis is
taken in the height direction.
|
Building Projection point
Target polN
Projection point
Figure 6. Horizontal position of a building and omnidirectional
cameras
Target point
Building
Projection point
Position 1
dee
Figure 7. Vertical position of a building and an omnidirectional
camera
Now, from the projection formula of a fish eye lens, the
distance v from image center to the position where the target
point is projected on the image can be given by Equation (1).
—201—