n pixel and
ined during
parameters
film images
ase in aerial
Np must be
]ually.
ige is expli-
. The image
y given via
crosses are
the trans-
es can only
points. The
ientation is
one has to
he fiducials
ducial num-
icted in fig-
fiducials in
distribution
nal, autono-
being fast,
|) it should
eras:
- imagery with different pixel size,
- black and white as well as colour imagery,
- positive and negative imagery,
- a varying number of fiducials,
- imagery with unknown pose.
The last point refers to the position of the film image
during scanning. Assuming that the images have not been
placed at an arbitrary angle on the scanner, eight diffe-
rent poses are possible, and the correct pose can only be
found by locating an asymmetrically placed pattern such
as the image number in the digital image.
The input for interior orientation comprises the digital
image, the camera calibration report detailing the came-
ra type, the image coordinates of the fiducials, the lens
distortion and information about the asymmetry pattern,
and the scanner report including the pixel size of the
digital image. The output consists in the pixel coordinates
of each fiducial, the parameters of the chosen trans-
formation from pixel to image space, and computed ac-
curacy figures. Since there are only a limited number of
fiducials in each image, the achievable degree of redun-
dancy is limited in interior orientation. Thus, particular
care must be taken in order to recognize and eliminate
blunders.
In order to solve the given task essentially a number of
steps have to be carried out:
- approximate positioning of fiducials,
- subpixel positioning of fiducial centres,
- positioning of the asymmetry pattern,
- computation of transformation parameters.
In concert with the strategy considerations mentioned in
chapter 2 the approximate location of the fiducials should
be determined in an image with reduced resolution. In a
higher pyramid level, the fiducial centre is not visible, and
the whole fiducial should be used as a primitive. Due to
the regular shape and size of the fiducial and its homoge-
neous brightness distribution, binary template matching
lends itself for this task. As an alternative at this low
resolution the fiducial can also be described by its outline
and subsequently located by means of polygon matching.
Subpixel positioning of targets has been a research sub-
ject for a considerable length of time (see e.g. Mikhail
1984; Luhmann 1988; Trinder 1989; Tichem, Cohen
1994). Due to the small size of the fiducial centre area
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
approaches are more appropriate than edge detection
algorithms trying to identify the perimeter of the fiducial
centre. Three possibilities come to mind:
- centre of gravity methods,
- least squares matching, and
- cross correlation.
Centre of gravity measures attribute a weight to each grey
value and do not sufficiently use the fiducial model. Least
squares matching relies on grey value gradients. How-
ever, most gradients in the vicinity of the fiducials are
rather small, and thus noise effects can have a significant
influence on the results. Thus, cross correlation is the best
choice. Subpixel accuracy is reached by fitting a two-di-
mensional function to the correlation coefficients and
determining the position of its maximum.
Methods for the determination of the image pose depend
on the available pattern. Therefore, no general recom-
mendations can be given here. Note, that if the imagery
is scanned directly from the film role, the pose is known
a priori and does not need to be determined within
interior orientation.
Examples for autonomous interior orientation were pub-
lished by Kersten, Háring (1995); Lue (1995); Schickler
(1995b) and Strackbein, Henze (1995). Features of these
algorithms are presented in table 1. At the time of writing
(April 1996) no external evaluation of these methods is
known to the author. This might have to do with practical
problems of automatic interior orientation. When inve-
stigating real fiducials in detail one finds that surprisingly
often the contrast between the image content around the
fiducial and the fiducial itself is rather poor. Another
major problem can occur if older cameras have been used
for the acquisition of the analogue film images: the fidu-
cial centre is not as bright as expected. In newer cameras
the fiducials are often illuminated by light emitting diodes
(LED). However, a problem remains for colour images:
due to the spectral characteristics of a LED the fiducials
are only visible in the red channel. Dust and scratches in
the images around the fiducials and scanning without
proper parameter settings can further decrease the qua-
lity of the fiducials in the digital image. For instance, the
area to be scanned is sometimes selected too small, and
thus the fiducials are only partly present in the digital
image. Examples of some of these cases are depicted in
figure 3.
301
