2. COMPLEXITY OF INTERPRETATION OF 
AERIAL IMAGES 
Image understanding, easily performed by human 
operators, is the  bottle-neck in  automatizing 
photogrammetry. The traditional approach for 
automated interpretation is to describe each object 
with a number of descriptive attributes. Their values 
are compared with values of a prototype, stored in a 
model base. This approach is realistic when only a 
small number of parameters is necessary to describe 
each object in an unique way. It is always possible to 
develop a procedure that nearly perfectly interprets a 
test-image with respect to a certain task-domain. How- 
ever such a method is often developed by an exhaus- 
tive trial-and-error process and therefore it relies 
heavily on characteristics of the test-image, like types 
and manifestation of objects. If these characteristics 
are stable, the procedure may operate successfully on 
future similar images. Interpretation of images that 
even just slightly deviate from the prototype tends to 
fail. The assumed stability of conditions is realistic in 
industrial environments. Typical in an industrial 
environment is the limited number of objects in object 
space and their isolated location. 
Interpretation of aerial images is a very hard task, due 
to the following characteristics of the image: 
* Objects in images of natural scenes need to be em- 
bedded in more extensive contexts, rather than 
treating them as independent objects. Many objects, 
like bridges and fly-overs, receive their meaning in 
the context in which they appear. So we need 
recognized objects to create a context for the inter- 
pretation of other objects. 
* The same type of objects can appear in a wide 
range of representations in an aerial photograph. So 
it is very difficult to build a prototype that des- 
cribes each type of objects unambiguous. 
Moreover the definition of an object is task-domain 
dependent. Additionally the appearance of one and 
the same object is not stable, due to changes in 
recording circumstances, like sun angle and season. 
So, in images of natural scenes the assumption about 
stability is violated, resulting in poor results for 
images that even just slightly differ from the test 
image. Consequently the traditional trial-and-error 
procedures are unsuited for solving the interpretation 
problem of aerial images completely. 
3. CONCEPTS FOR INTERPRETATION OF 
AERIAL IMAGES 
The need of context to interpret images of complex 
object spaces, brings us to the conclusion that a multi- 
stage recognition procedure is required. That means, in 
simple terms, that a small part of the image may 
already have been recognized, while other parts of the 
image are still in the raw image data stage. The recog- 
nized parts influence or may even guide the further 
recognition process. 
478 
Because of the variability in appearance of objects, we 
need an integration of different segmentation tech- 
niques, guided by a control level. Hence we adopt the 
usual two level approach (Nazif and Levine, 1984; 
Tenenbaum, 1973). The task of the low level is 
segmentation and the high level's task is interpretation 
of these results. 
Interpretation at the high level may be looked at as a 
process of hypotheses generation, testing and updating. 
For all possible changes that a road may undergo, 
hypotheses should be generated. 
Testing these hypotheses means following a path in a 
tree with many branches, where at each branchpoint 
one has to take a decision which branch to follow. It 
is our believe that the path should depend on the 
characteristics of the data, the type of object and the 
task domain. In this way knowledge from different 
sources is integrated. 
During the interpretation process new knowledge 
becomes available, derived from the image data. 
Consequently hypotheses have to be updated during 
the interpretation process. Therefore the condition part 
of the decision rule, examined at each branchpoint, 
should not be fixed, so it can be adapted. Parameters, 
like the values of thresholds, need to be determined 
during the interpretation process. Our idea is to use the 
a priori information present in the database, to deter- 
mine among others initial values of these parameters. 
The order of hypotheses testing is also of importance, 
introducing the problem of the creation of a 
hypotheses hierarchy. 
The next sections will describe how the multi stage 
and two level approach can be used for automatized 
road updating. 
3.1 Multi-stage approach 
The concept that different parts of the image are in 
different stages of processing, requires a strategy to 
guide the major steps of the hypotheses hierarchy. It is 
task-specific and depends on input and required out- 
put. Our input is a road database of the old situation 
and a scanned aerial photograph of the present 
situation. 
The basic real-world knowledge of our strategy is that 
roads form a connected network. A network can be 
considered as a graph, consisting of nodes and arcs 
between nodes. Nodes are for example crossings and 
fly-overs. Arcs are road segments that start and end in 
a different node or at the side of the image. 
Using the multi-stage approach, arcs and nodes will be 
extracted in different stages of the process. Arcs have 
features that are easier to extract than features of 
nodes. Features of nodes are also more dependent of 
the type of node. Hence arcs are examined first in 
each step of the strategy, which in fact is an interpre- 
tation of road segments. They create a context for 
interpretation of nodes, so a hypothesis for the location 
and type of the node can then be made. This simplifies 
segmentation and interpretation of the nodes.