ordinates in x-
icipants. Again
h edge-based
‘acy in image
een confirmed
worse results
rates gives an
uracy for real
and lighting
mbrunn 1995)
of accuracy of
s has to be
istment results
s (which are
e of operator.
coordinates or
should not be
of a complete
on investigated
rocess it is not
> values. An
be performed
isured object
vailable at the
on of object
adjusted object
| space. This
).02 pixel if the
account. With
1m) a relative
Joints
3-D vector
ct points
its have been
and that they
pplications the
imum residuals
5 SUMMARY
In the first run the presented test on digital point
measurement accuracy was restricted to members of the
Working Group Industrial and Engineering
Photogrammetry of the German Society for
Photogrammery and Remote Sensing (DGPF). Although
there were only 10 different results returned, the test has
clearly shown that there is a significant need for users to
investigate their measuring operators and systems. This
test can therefore be regarded as one step to a
standardized acceptance test for digital photogrammetric
systems tha has already been required (Dold 1995,
Wendt 1995). In order to move forward a joint working
group of DGPF and VDI (German Association of
Engineers) has meanwhile been established which will
work out criteria and rules for the testing of
photogrammetric systems (based on similar guidelines for
CMMs).
The generated test images were almost suitable for the
determination of sub-pixel accuracy. A second generation
of test images should include very small points which are
often used in practical applications.
Theoretical and practical investigations have
demonstrated that the limit of sub-pixel resolution is about
11100 of a pixel. In order to achieve even higher
accuracies the number of processed bits per pixels must
be increased to 10 or 12 which is already the
performance of state-of-the-art CCD imagers.
This very high accuracy can only be obtained if perfectly
shaped and imaged targets can be used and if the whole
imaging process (illumination, optics, sensor, analog-to-
digital conversion, data transfer) could be handled with
equivalent precision. Noise, disturbed targets or changing
backgrounds will decrease accuracy significantly.
The methods used in this test include: edge-based ellipse
operators, center-of-gravity operators, template matching
and least-squares matching. Ellipse operators seem to
produce best results (1/100 pixel) but, due to the
algorithm, are restricted to points larger than 4-5 pixels in
diameter. Gross errors in point shape can be handled
very well by ellipse operators and robust adjustment
algorithms.
Adaptive center-of-gravity methods can operate on small
targets and they offer short processing times. The
achieved accuracy is slightly poorer than for ellipse
operators (2/100-3/100 pixel). However, they can hardly
recognize point disturbances nor background variations
due to the non-structural approach of the method.
Least-squares matching is an acceptable tool for point
measurement if suitable templates can be generated.
Like center-of-gravity it cannot handle point artifacts
unless they are combined with geometric constraints such
as epipolar geometry. However, this is only possible if
orientation values are available. Geometric accuracy
seems to be limited to 2/100 of a pixel.
329
The testfield image series has shown that there is a
decrease of accuracy of about factor 2-3 compared to
ideal synthetic imagery. The images have been acquired
under laboratory conditions with a brand-new Rollei digital
camera where several control parameters still have to be
optimized. The mean accuracy of 0.04 pixels confirms the
potential of other modern digital cameras (Peipe 1995).
Under practical environmental circumstances in industry it
is likely that further accuracy decreases may occur.
The main goal of this test has been achieved
successfully. The theoretical and practical limit of point
measurement accuracy has been investigated. Image
measuring accuracy can be interpreted as the probing
uncertainty known from CMM technology. It is not a
measure of object or system performance because
additional criterias like object targeting, network design or
processing time have to be taken into account.
6 REFERENCES
Broers, H., Hemken, H. (1996): Entwicklung eines
Verfahrens zur Erzeugung und Messung punktférmiger
Muster in digitalen Bildem. Diploma thesis, unpublished,
Fachhochschule Oldenburg.
Dold, J. (1995): Überprüfungsverfahren für photogram-
metrische Systeme. DGPF Annual Convention,
Hannover.
Förstner, W. (1986): Prinzip und Leistungsfähigkeit der
Korrelation und Zuordnung digitaler Bilder. Schriftenreihe
des Instituts für Photogrammetrie, Nr. 11, Universität
Stuttgart.
Godding, R., Woytowicz, D. (1995): A New Digital High-
Resolution Recording System. ISPRS Intercommission
Workshop "From Pixels to Sequences", Zürich.
Luhmann, T. (1995): Punktmessung in digitalen Bilder
mit Subpixel-Genauigkeit. In Ahlers (ed.):
Bildverarbeitung '95, Technische Akademie Esslingen.
Peipe, J. (1995): Photogrammetric Investigation of a 3000
x 2000 Pixel High-Resolution Still-Video Camera. ISPRS
Intercommission Workshop "From Pixels to Sequences”,
Zürich.
Wendt, K. (1995): Überwachung und Kalibrierung von
optischen 3D-Koordinatenmeßsystemen. DGPF Annual
Convention, Hannover.
Wester-Ebbinghaus, W. (1986): Analytische
Kammerkalibrierung. Int. Arch. Photogrammetry and
Remote Sensing, 26/5, Ottawa,pp. 77-84.
Zumbrunn, R. (1995): Systematic Errors with Retro-
reflective Targets. ISPRS Intercommission Workshop
"From Pixels to Sequences", Zurich.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996
