The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B5. Beijing 2008
The model was extended by additional parameters for the
compensation of remaining systematic errors, allowing the
standard deviation of unit weight obtained from a spatial
resection to be reduced to 0.1 pixels for a 14-Megapixel Kodak
camera equipped with a Nikkor fisheye lens. Translated to
object space, this corresponds to lateral accuracy of 1 mm at 10
m distance.
There are different fisheye projection geometries: Equi-distant,
equi-solid-angle and orthographic projection (Ray, 1994;
Backstein & Pajdla, 2002; Abraham & Forstner, 2005). The
geometric concept is based on the dependence of the image
radius r’ and the angle of incidence a (Figure 3).
Experiments confirmed that the used fisheye lens complies best
with the equi-solid-angle model. Therefore only this model will
be described in the following. The image radius r ’ depends on
the angle of incidence a:
r'= 2c ■ sin
“ where Xma= EH
(5)
In order to express the image observations x' and y ’ as function
of the exterior and interior orientation as well as on the object
point coordinates, equation (5) has to be introduced in the
following equation, whose derivation is explained in detail in
[Schwalbe & Schneider, 2005]:
x'= ;c 0 ' +
y=T 0 ' +
\yj
+ Ay'
+ 1
(6)
The correction terms Ax ’ and Ay ’ contain additional parameters
for the compensation of radial-symmetric (A h A 2 , A 3 ) and
decentering (Bi, B 2 ) lens distortion (Brown, 1971) as well as
affinity and shear of the image coordinate system (C/, C 2 ) (El-
Hakim, 1986):
Ax' = x' (A i r ,2 +A 2 r' 4 +A 3 r' 6 ) + Bj (r ,2 + 2x a ) + 2B 2 x' y' + C,x' + C 2 y' (7)
Ay' = y'(A i r' 2 +A 2 r' 4 +A 3 r' 6 ) + 25,x'y' + B 2 (r a +2y a )
3. COMBINED BUNDLE BLOCK ADJUSTMENT
The geometric model of fisheye lens cameras was integrated
into a bundle adjustment software package, which was
originally developed and implemented for a combined analysis
of laser scanner data and central-perspective or panoramic
image data (Schneider & Maas, 2007).
The bundle adjustment software package supports the
calculation as free network adjustment and handles outlier
detection. Since different types of observations have to be
adjusted simultaneously, it is necessary to assign adequate
weights to the laser scanner and fisheye image observations.
For this purpose a variance component estimation procedure
was implemented in the adjustment. Thus, the precision
characteristics of laser scanner and fisheye lens camera will be
optimally utilised, and an improvement of the adjustment
results can be achieved. Furthermore, this allows for a
qualification of the measurements in terms of realistic accuracy
values for fisheye lens data and laser scanner data.
4. PRACTICAL EXPERIMENTS
4.1 Calibration room
In order to practically assess the presented method, multiple
laser scans and images with a fisheye lens camera were
acquired in a test field, which is designed for the calibration of
fisheye lenses (Schwalbe, 2005).
The test field is a room (4><5><3 m 3 ), where 100 signalised
object points are distributed at the surrounding walls and at the
ceiling in a way that they form concentric circles on the fisheye
image (Figure 4). Object point targets are designed as black
circles (0 10 mm) on white background with a ring code, to
allow for an automatic target detection and identification. The
targets are orientated in a way that they face perpendicular to
the centre of the room. Reference coordinates were determined
using a common photogrammetric measurement system.
However, the coordinates are only used as approximate values
in the bundle adjustment and to define the superordinated
coordinate system. The XY-plane of this coordinate system is
orientated horizontal and the Z-axis is a vertical axis.
Figure 4. Fisheye image from the centre of the calibration room
orthogonal to the ceiling with highlighted object point numbers
4.2 Laser scanner data
A terrestrial laser scanner Riegl LMS-Z420i was situated in
each comer of the calibration room and tilted 45° vertically, in
order to allow for the recording of points on the ceiling.
Additionally, two scans from the centre of the room were
recorded with a tilt angle of 90°. The laser scanner was also
rotated 90° horizontally between both scans.
The angular resolution of the laser scans was 0.035°, which
corresponds to a scan point distance of 2.5 mm in 4 m distance.