LOGICAL
R 54006,
r prehistorical
ional way for
he data. Con-
> acquired be-
increases due
th less experi-
er virtual im-
that the accu-
tortion model
mation has to
tual images 2
ent overlap to
extracted and
ited to reduce
tomatic DSM
es features as
uality control
w images and
which can be
frames com-
erna in Crete
it the archeo-
ocessing (4m
eurs with less
fore the geo-
d the camera
se bundle ad-
extraction of
lighted)
Figure 2. Excavation site of Eleftherna
2. ACQUISITION - FIELD WORK
The images have been acquired in the early morning avoiding in
this way the generation of saturated spots that can appear at
strong noon sunlight. A digital camera Mavica FD81, with pixel
size 4.7 um, focal length 4.708 mm and dimensions 1024 x 768
pixels (Georgiadis et al., 2000). Colour images were acquired
but B/W were finally used. As mentioned previously, the ge-
ometry of the block of images had some deficiencies. The ratio
base-to-height and the rotations were not optimally configured
(bases between the camera stations were rather small, close to 4
cm, and differences in omega angle, were up to 36 grad). In to-
tal, 2 strips were used, each consisting of four images with 70-
80 % overlap. Additionally, control points have been estab-
lished by topographic methods with an estimated accuracy of 1-
2 cm. Only 10 out of the 17 measured GCP’s were finally used
due to unsatisfying visibility and radiometric quality. An exam-
ple of an acquired image with GCP’s overlaid is shown in Fig-
ure 3.
Figure 3. Example of image with control points marked
The estimated height error for the given imaging configuration
was 077 0.04. Even though for archaeological documentation it
can be accepted.
3. BUNDLE ADJUSTMENT
The epipolar geometry has been checked using the coordinates
of the GCP's, the exterior and interior orientation of the images
given by an adjustment performed on the field. A deviation of
approximately 15 - 20 pixels from the true position of the
epipolar line was found. Therefore, a bundle adjustment with si-
multaneous self calibration had to be performed in order to im-
prove the accuracy of the orientation and model the lens
distortion parameters. This is necessary since radial symmetric
distortion is the largest systematic error source when using
solid-state cameras with low cost CCTV — type lenses and short
focal lengths (5-20 mm) (Beyer,1992). In the bundle adjustment
7 images were used (3 from the first strip and 4 from the sec-
ond), 14 tie points and 10 manually measured GCP’s. The in-
house developed program SGAP was used and an iterative ad-
justment was performed. In a first step orientations and tie point
coordinates were computed, excluding points with residuals
greater than 3 cm. In a second step, camera constant, principal
point coordinates and the Brown distortion parameters (Brown,
1971) were computed. Finally epipolar lines have been checked
for the existing GCP’s. In Figure 4 the epipolar lines before
and after the adjustment for 2 GCP’s are shown. The largest
standard deviations computed in the bundle adjustment were (in
m): oy- 0.0244, oy- 0.0193 and 6,-0.0301. The overall theo-
retical precision calculated from the 24 points was (in m): o, =
0.0171, o, = 0.0142, o, — 0.0210. The camera constant was
computed as 5.164 mm (initially it was 4.708). The precision
achieved was within the given requirements and therefore the
extraction of the 3D information more accurate and reliable.
Figure 4. Deviation of epipolar lines before and after bundle ad-
justment with self calibration. The epipolar lines are
shown for 2 GCP's. On the left column the epipolar
lines before bundle adjustment and on the right col-
umn after bundle adjustment. Point number 2 is close
to the centre of the image, while point 11 near the
border where the influence of the additional parame-
ters, modelling lens distortion, is larger.
4. VIRTUAL IMAGE
The virtual image can be defined as an image that substitutes a
sequence of images acquired from different stations, the sub-
images. In case of sub-images acquired with the same camera,
the virtual image has the same camera constant but a new vir-
tual size, field of view and camera exterior orientation. Each
virtual image substituted 3-4 images, without change in pixel
size dimensions (4.7 pum). A reference plane is defined, the
camera exterior orientation is set to the average exterior orienta-
tion of all 3 or 4 images, the focal length is kept the same and
the principal point is set to 0,0. By projecting the 4 corners of
each image on to the predefined plane (see Fig.5) and back-
projecting through the orientation of the virtual image, the
-441-
