rrect loca-
the match-
ion of tra-
ried out by
, obtained
ribed type.
ig an LC-
ies of the
ample, by
igth and a
onstructed
es.
\ages, hu-
(as could
ng a DTP
(this also
ex. as ob-
n collima-
| Still nec-
but corre-
hing algo-
jrid model
, for each
n the his-
VV [-.
model
rs (AZ/H)
9: 104H
Considering that similar tests on analytical plotters give
0, = +05 10^ H, it seems that the last value can be ac-
cepted as being very satisfactory. On the contrary, case
a) gives unacceptable results, and case b) is also critical
for the most standard requirements.
4.2.2 Points of a "real model". A stereoscopic model
formed by two colour images at the average scale of
1:4,000 over the town of Ferrara has been used. About
50 points have been surveyed using a DIGICART40
analytical plotter and measuring their analytical co-ordi-
nates. The images have been digitised at 600 dpi and
the same points have been measured using the mixed
procedure as explained before: a rough human collima-
tion at pixel resolution (with an on-line geometric calibra-
tion) and an automatic refinement of the collimation at
sub-pixel resolution, finally give their digital co-ordinates.
The average r.m.s.e. of the sub-pixel correlation is 6um (
i.e. 0.19 pixel). Discrepancies between the analytical and
digital co-ordinates, expressed in cm, are shown in fig.7.
The standard deviation is ox,=+9 cm and c,=+5 cm.
f
/
/ / {
lli
Frequencies
Discrepancies [cm] s 20
Figure 7 - Discrepancies in single point collimation
4.3 Automatic DEM production
A DEM of 250 points covering rural and urban areas,
regularly spaced in 10 m in X and Y, has been surveyed
on the model of Ferrara described above, with the
DIGICART 40. The same DEM has then been automati-
cally measured by the VLL algorithm as described in 3.2.
Each iteration used 9 horizontal planes centred onto the
approximate Z-value. The first distance between the
planes was AZ = 2m and the last AZ = Ah = 2cm (8 it-
erations). 78,6% of the points have been automatically
and successfully measured without any human interven-
tion (green points). The remaining 21.4%, where the r
coefficient is too small, have been marked in red or yel-
low and required a further operator intervention: 4%
resulted to be already correct and 17.4% needed a new
collimation. The computation speed is approx. 28 pts/s.
The discrepancies between the analytical and digital co-
ordinates of the green points, expressed in cm, are
shown in fig. 8. The standard deviation is 0,=+8 cm.
Frequencies
BR 8838858 8% 8
A ne
o o
o wo
-20 -15 -10 -5 0 5 10 15 20
Discrepancies [cm]
Figure 8 - Z discrepancies in DEM automatic data
capture
5. CONCLUSIONS
Performances of an LC-DPS, where DTP scanned im-
ages (600 dpi) and human collimations at 1 pixel resolu-
tion are frequently used, can be significantly improved if
a rigorous geometric calibration of the digital images is
carried out and if matching sub-pixel algorithms are
applied during the restitution phases.
By comparing the Z measurements on an artificial "grid
model", it results that c, changes from + 5 - 10^ H (non
calibrated images) to = + 2 - 10^ H (calibrated images),
up to = + 0.6 - 10“ H in the case of sub-pixel automatic
correlation. This last accuracy is comparable with that
obtainable by an analytical plotter.
Practical tests on a real model (colour images at scale
1/4,000) confirm such results when observing both single
points manually and a DEM grid by an automatic proce-
dure.
6. REFERENCES
[1] Kraus, K., 1993. Photogrammetry Vol. 1. Dümmler,
Bonn
[2] Kólbl, O., Bach, 1994: Tone reproduction of photo-
grammetric scanners, OEEPE test report
[3] Gonzales, R.C.Woods, R.E. 1992. Digital Image
Processing - Addison Wesley
[4] Pratt, W.K., 1991. Digital Image Processing. Wiley,
New York
[5] Gruen, A., Baltsavias, E., 1988: PE&RS 54, pp. 633-
641
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B1. Vienna 1996
