eration of a 
his purpose 
on of gross 
rocess. The 
weight for 
y that the 
individual 
points are 
( from the 
to its filter 
eached (all 
of iterations 
nodel and a 
oints. 
ım are the 
3). For the 
sidering an 
easurement 
gross errors 
1 12). The 
its “return 
n is a bell 
etection in 
h value (h) 
nally it can 
in order to 
n respect to 
ft and right 
the weight 
1elp of two 
in distance 
the DTM 
d that the 
nd without 
off-terrain 
id. rejected 
or the filter 
e found in 
on a “good 
erefore this 
hich occur 
we use the 
| similar to 
the help of 
rm that we 
irse-to-fine 
the coarse 
finer level 
ists of all 
sed instead 
ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision", Graz, 2002 
  
  
  
  
  
  
  
i 
pe | 
| | 
| 
| 
  
Figure 1: Sequence of the hierarchic robust interpolation 
  
d) 
a) Creation of a data pyramid, small points: original data, thick points: data pyramid (lowest point in a regular 5m interval). 
b) DTM generation in the coarse level by robust interpolation, the remaining point on the house is eliminated with an 
asymmetric and shifted weight function. The surface in the first and last iteration is shown. 
c) Coarse DTM with a tolerance band, all original points within the tolerance band are accepted. 
d) DTM generation in the fine level by robust interpolation using an asymmetric and shifted weight function. Again, the first 
and the last iteration is shown. 
The hierarchic robust interpolation proceeds like the following: 
1. Create the data pyramids with the lower resolutions. 
2. Apply robust interpolation to generate a DTM, starting at 
the coarsest level. 
3. Compare the DTM to the data of the higher resolution and 
accept points within a certain tolerance band. 
The steps 2 and 3 are repeated for each finer level of detail. The 
sequence of the hierarchic robust interpolation on a synthetic 
laser scanner profile in a build-up area is presented in fig. 1. 
Further details about hierarchical robust interpolation, its 
implementation in the software package SCOP++ and the 
results for an OEEPE dataset can be found in (Pfeifer et al., 
2001). 
4 EXAMPLES 
In the following the results of these algorithms applied to 
different datasets are presented. As it will be seen the procedure 
is adapted to the characteristics of each dataset. 
4.1 Robust Interpolation of Tacheometric and Photo- 
grammetric Data 
The Municipality of Vienna ordered a test project for the 
determination of a DTM from available digital map data. In the 
streets the data was captured with total stations Photo- 
grammetric measurements were used for eaves, park areas and a 
few other features. The initial classification of terrain points was 
performed by point attributes stored in a database. The resulting 
dataset with all classified terrain points was very 
inhomogeneous due to the fact that there were a lot of points 
along the streets and only a few points in backyards and park 
regions. Therefore we decided to densify the data to get a 
complete DTM over the whole test area. The densification was 
performed by exporting a regular 5m raster after a triangulation 
of the data. The DTM was computed by linear prediction 
considering different point class accuracy. Therefore we were 
able to give the originally measured points a higher accuracy in 
contrast to the densified data (a small measurement variance of 
25cm? vs. 1m? for the densification points). 
   
      
       
  
   
  
  
U 
mn 
W 
SZ S 
ZZ 
Y 
T 
M 
MR 
M N 
3 
x) 
Sees 
m 
NN 
i 
0 
KX 
Kr 
A 
" 
j 
X 
jn 
j 
X 
A 
y 
a 
i 
Y^ 
ses 
           
S 
o OS Sao OO UT 
SS 25252 SD OO OS ZZ 
SSR SS ZZ 
< > 
SS SS = = = 233% 
= = IT > ZZ ZZ ZZ 
— = SSCS 2 OO Eee OS D DUO SS = | 
sn EE ZZ 
pe DOC OQ > 2 > 22 > 
p pp ppp ee ETS m ZZ 
Figure 2: Perspective view of the DTM (5m grid with) from the 
selected terrain points (from the database) after densification 
SO 
This leads to a DTM, which is mainly influenced by the 
measured data. The interpolated grid is only used to get a 
complete surface model. A perspective view of a part of this 
model can be seen in fig. 2. The terrain in this area is rather flat, 
but a few spots indicate data errors caused by false point 
attributes. A closer look at the data in combination with a geo- 
referenced vector map of the building blocks showed, that there 
are some misclassified points on the houses due to a false 
attribute and there where also a few points with obviously false 
z-values. Therefore it was necessary to eliminate these gross 
errors.