model 
Therefore, now fixed 3-D wireframe models are used for the 
model-based control point location. The control point models 
are not any more treated as being stochastical. An approach 
for the automatic determination of the 3-D control point mo- 
dels is described in [BRAUN C. 92]. Available approximate 
orientation values enables us to reduces the searchspace of 
the matching process. The approximate geometry of the ob- 
ject up to a few degrees in the image, and its approximate 
location (^ 50 pixel ) is known. 
The whole automatic orientation procedure consists of follo- 
wing 6 steps (cf. Fig. 1). Step 1 to 3 are done seperately for 
each controlpoint leading to a good approximate position of 
the controlpoints in the image. Step 4 to 6, the final evalua- 
tion of the orientation parameters, and the selfdiagnosis is 
done commonly for all controlpoints. 
1. Projection of the 3-D control point model into the 
aerial image using the approximate orientation values. 
This projection leads to a 2-D wireframe control point 
model in the sensor system (cf. Fig. 1 b). 
2. Extraction of straight line segments [SCHICKLER 
W.92] in a subsection of the digitized aerial image (cf. 
Fig. 1 d). The position of these subsections whith re- 
spect to the whole image is derived from step 1. Its 
size depends on the precision of the approximate ori- 
entation values and the size of the control point buil- 
ding in the image. The inner orientation of the subsec- 
tions is known, therefore the extracted line segments 
are determined in the image koordinate system. In 
case of color images the available color edge extraction 
[BRUGELMANN R./FORSTNER W. 92] is applyed. 
3. Probabilistic clustering technique to determine ap- 
proximate position of the controlpoint model in the 
subsection of the aerial image. This 2-D matching pro- 
cedure leads to a preliminary set of matching candida- 
tes between image and model edges (cf. Fig. 1 e). 
4. Ransac technique [CF. BOLLES R. C. / FISHLER M. 
A. 81] to find incorrectly located control point models 
and to predict a more likely set of matching candidates 
(not yet implemented). 
9. Robust estimation to clean the whole set of prelimi- 
nary correspondencies (cf. Fig. 1 f). This 3-D matching 
procedure is a final common fit of the 3-D controlpoint 
models to the image (cf. Fig. 1 g). It is a robust spacial 
resection using homologous straight line segments in- 
stead of homologous points leading to the orientation 
parameters of the aerial image. 
6. Selfdiagnosis by analysing the final result whith re- 
spect to precision and sensitivity considering the geo- 
metric configuration of the control points. This enables 
AMOR to decide whether the automatical determined 
orientation parameters are acceplable or have to be re- 
jected. 
3 The matching procedures 
In many model-based object loaction concepts the matching 
problem is solved by a heuristic tree search procedure [c.f. 
GRIMSON W. / LOZANO-PÉREZ T. 87] The restrictions 
(cf. Section 2) allowes to apply a special 2 step matching 
procedure which is robust with respect to outliers and even 
in cases of weak information in the image able to locate the 
control point models correctly. 
3.1 Finding Candidates for Correspon- 
dencies via Clustering 
The task of this procedure is to determine the approximate 
position of the projected model in the subsection of the 
image, to determine a set of matching candidates between 
image and model edges. 
For each match of an image edge with a model edge the po- 
sition of a reference point is calculated. This point can be 
chosen arbitary but has to be fixed. As the length of the 
matched edges in general are not equal, a range of expected 
positions is added to the accumulator, taking the uncertain- 
ties of the extracted image edges into account. The final ac- 
cumulator is a discretized version of the probability density 
function of the reference point, with a maximum for its most 
likely position. With this position one can derive a set of 
matching candidates between image and model edge. 
The propabilistic clustering has been descriped in [SESTER 
M. / FÓRSTNER W. 89]. It has be modified to treat all 
kinds of matches shown in Fig. 2. 
If the given set of model features is unique, this 2-D matching 
technique is able to locate the model correctly even in cases 
where only a few (3 to 4 edges) corresponding image featu- 
res are present. Unfortunately the uniqueness of the control 
point models is not always given due to weak model infor- 
mation. This may lead to partially wrong corespondencies or 
totaly incorrect locations. 
3.2 Representation of the Straight Line 
Correspondencies 
The clustering procedure results in a set of preliminary mat- 
ches. Unfortunately the length of an image edge usualy is 
not the same as the length of the corresponding model edge, 
due to occlusions, weak image information, shadows and so 
on. The final fitting of the models to the image is done by a 
robust spacial resection using homologous straight line seg- 
ments, which has to cope with homologous lines also in case 
the length of model and image edge is not the same. 
There are 7 possible cases of matching an image edge to a 
model edge: 
  
  
  
a b c d 
e f g 
Fig. 2 Cases of matching an image edge (thin lines) to a 
model edge (thick lines) 
The representation of these partial matches can not be done 
by individual point correspondencies, the image edges have 
to be treated as units holding the line information with pos- 
sibly undefined start and (or) end point position. 
This seperation of the line and point information can be done 
representing the image edge segments in the following man- 
ner: 
593