by no more than two images. Thus, such a concept of 
surface reconstruction takes full advantage of the 
multiple image overlap and uncouples the DEM 
generation from the classical stereo model. Also, the 
integral approach has the prospect to automate the aerial 
triangulation, the DEM generation and the orthophoto 
production in one batch process. 
This paper reports in general on the current status of the 
ongoing system development. It shortly reviews the 
approach to the automatic aerial triangulation with 
special attention to the DEM aspect. Practical results are 
given for the initialization part of the system based on an 
integrated DEM generation at a coarse pixel resolution of 
480 um. Main attention is paid to the controlled tests of 
the automatic aerial triangulation of two blocks with 45 
images and 21 images, resp. At the end of the paper, 
preliminary results of an integrated DEM generation are 
presented, indicating the advantage of the multiple 
image matching approach. 
2. Concept of automatic aerial triangulation 
2.1 General remarks 
Two basic key techniques play an essential role in our 
concept. Firstly, it takes into account that the GPS 
technology is well established in aerial triangulation. It is 
well-known that GPS navigation systems provide regular 
block forms and GPS positioning techniques pre- 
determine the projection centers with an absolute 
accuracy of at least 30 m. Thus, the initialization is 
considerably simplified, since a sufficient overlap of the 
homologous image patches is guaranteed, except for 
large scale photography in mountainous terrain and 
camera attitudes larger than 2 degrees. Secondly, we 
use predominantly the feature-based matching method 
as the matching strategy. This means that point clusters 
are measured and transferred by means of image 
processing techniques instead of single points. The 
implied measurement philosophy aims at a high 
redundancy which is one of the preconditions for highly 
accurate and reliable results. 
Basically, the approach is intended as a fully automatic 
process which can start from scratch by using only little 
initial block information. The key idea of the approach is 
to use an integrated block adjustment in the matching 
strategy in combination with robust statistics. Thus, the 
AT process provides as main results both orientation 
parameters and adjusted object coordinates. The entire 
procedure is characterized by two main steps. It starts 
with the initialization which determines accurately 
enough the tie point areas at the Gruber point positions. 
The kernel system then applies the matching strategy in 
the homologous image patches through the image 
pyramid (Figure 1). 
The possible input data are manifold and comprise very 
crude initial block data like the flight index map, the strip 
azimuths and a mean terrain height, as well as GPS/INS 
sensor data eventually in combination with a DEM. The 
digital images are given at appropriate resolutions (e.g. 
15 um or 30 yum), eventually with a coarse overview 
image and the associated interior orientation. Also, 
sufficient ground control is assumed. The number and 
406 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
distribution of ground control follows the same known 
rules as for a conventional aerial triangulation. Precise 
airborne GPS antenna positions can be introduced 
additionally as control information to reduce the number 
of ground control points. 
The preparation part of the AT system defines some 
program parameters. If the interior orientation has not 
been applied so far, the fiducials are semi-automatically 
measured. Also, the ground control points have to be 
measured interactively. In the exceptional case of 
signalized points semi-automatic matching tools can 
also be applied. Note that only the preparation is an 
interactive part of the system. All the other process steps 
can be invoked in batch. 
  
Preparation: 
* parameter editing 
* interior orientation 
* measurement of control points 
  
  
  
Initialization of tie point areas 
  
  
  
Visual check / Edit 
  
  
| Derivation of (sparse) image pyramid 
  
  
  
Kernel system: 
* Feature extraction 
* Preliminary matching 
* automatic point transfer in 
robust bundle adjustment 
  
  
  
  
  
  
final block adjustment 
  
  
Tnteractive 
Figure 1: General workflow in MATCH-AT 
2.2 Initialization 
The initialization has to provide the Gruber positions in 
the images accurately enough for the pull-in range of the 
subsequently invoked kernel system. Assuming a pull-in 
range of 10 pixels for the feature-based matching 
technique, which is predominantly used in the kernel 
system, the image patches to be matched should not be 
shifted against their homologous position by more than 1 
cm at a coarse pixel resolution of say 1 mm. 
The initialization can use the mentioned input data in 
various combinations. For instance, if GPS/INS data are 
given with today's accuracy of 30 m and 0.2 - 0.5 
degrees respectively, it is very easy to directly derive the 
tie point areas with an accuracy of at least 1 cm in the 
image. In case of large height undulations, a DEM is 
advantageous to compensate for the critical relief 
displacement. Although GPS has almost become a 
standard, low cost INS is going to become attractive and 
    
    
     
   
   
   
    
   
   
    
   
   
    
  
   
    
   
   
  
  
  
  
  
   
   
    
    
   
    
  
  
  
  
  
  
   
   
  
    
  
    
   
   
    
    
   
  
  
    
    
    
   
   
    
DEMs 
also hs 
with th 
above, 
crude 
project 
the im: 
autom: 
This is 
pixel re 
1mm 
adjust 
applie« 
conver 
metho: 
observ 
small 
the prc 
200 m 
large i 
and m 
also a 
the py 
metho 
all ime 
It aim: 
adjust 
remait 
genere 
known 
recalle 
alread 
a min 
obsen 
  
Figure 
The mr 
with 
Furthe 
are c: 
can © 
Howe 
and 
unacc 
2.4K 
Once 
is sL 
Additi 
cente