isa 
neck. 
for a 
281). 
B 
ints. 
nuch 
own 
991) 
r the 
and 
d of 
hing 
tical 
ents 
tion 
tive 
sing 
ired 
tion 
ined 
the 
ved 
ated 
cial 
rent 
ally 
y an 
ion. 
heir 
1 in 
be 
hen 
Ork 
'ate 
ator 
the 
"th 
der 
; in 
ing 
Vu, 
for 
gle 
ich 
Its, 
ect 
tly 
  
represents buildings. Using approximate 
values for the orientation parameters in order 
to increase the efficiency, the system matches 
the images of the object lines with straight line 
segments, as features extracted from the 
image. This is done in several steps for all 
available control points. The final spatial 
resection is performed with the straight line 
segments and evaluated wilh respect to 
precision and reliability. 
This method implies the use of features instead 
of distinct points for orientation. Orientation in 
photogrammetry in the traditional approach is 
more or less completely based on distinct 
points. However, geometric features in object 
space can also be straight or curved lines and 
planar or curved surfaces. These features and 
their projection into image space can be used 
for image orientation. Especially for digital 
techniques of image analysis this aspect is 
essential, because lines are easier to extract 
than point features by automatic procedures. In 
order to rigorously accommodate the 
extraction and correspondence of these 
features suitable photogrammetric formulation 
must be performed. Research and practical 
work based on the correspondence of general 
geometric features, beside points, is being 
performed in computer vision and 
photogrammetry. 
Haala et al (1992) discusses the use of 
relational matching to match relationa! 
descriptions of images and maps. For 
instance, the top of the seat and the front of the 
back may be at right angles to each other. The 
structural descriptions of the images are 
obtained by thresholding selected channels of 
colour images and subsequent thinning of the 
linear structures. The structural descriptions of 
the landmarks were obtained by digitising 
maps, but, in principle they could also have 
been derived from a Geographic Information 
System (GIS). 
33.3 Feature extraction 
Feature extraction is the subject of intense 
interest at the moment but except in very well 
defined areas, there is little prospect of a 
robust solution to the majority of problems. 
McKeown, (Dowman ct al 1993) notes the 
major problem to be the complexity and 
variability of the scene interpretation task. In 
other words, it is very difficult to design a 
system which will cope with a wide variety of 
common scene characteristics. Two major 
research directions are apparent. The first is 
the definition of basic concepts and 
relationships and the design of tools to fit in 
with these concepts. The second is the use of 
multiple integration techniques using several 
341 
different types of data and knowledge based 
algorithms with a GIS. 
The work of Forstner (1992) relies on 
complex object models, their inter- 
relationships with their parts and other objects 
and the variation over time. Molenaar and 
Fritsch (1991) works along similar lines but 
has a primary interest in the data within a GIS. 
The automated extraction of linear features 
have been attempted by hierarchical texture 
analysis (Moller-Jensen 1990), and by search 
techniques like dynamic programming (Nonin 
1992, Gunst and Lemmens 1992, Maitre and 
Wu 1989) requiring initial approximation of 
the location of some features (or connections); 
these initial approximations could come from a 
GIS or from edges extracted with kernel filters. 
Nonin (1992) has described a system based on 
work by Maitre and Wu (1989) and operated 
by ISTAR, designed for the extraction of 
linear features. The operator identifies the 
features approximately and a dynamic 
programming based algorithm determines their 
exact position. INRIA are working with IGN 
(France) and CNES on the problem of 
extraction of roads. Laser scan are also 
experimenting with an iterative system based 
on their VTRAK system. 
The classification and segmentation of land 
use features no longer encourages per-pixel 
techniques that do not utilise neighbouring 
information such as texture. There seems to be 
greater interest in systems utilising a priori 
data from digital maps, GIS, (Janssen and 
Amsterdam 1991, Bolstead and Lillesand 
1992), DEMSs and other knowledge sources 
like human experts (Middelkoop and Janssen 
1991). ICC have combined data from a 
number of sources which include previous 
classifications, satellite data and topographic 
maps, the classification is done using neural 
networks and a 5% improvement on previous 
results is obtained. 
The extraction of buildings require high 
resolution imagery (e.g.. aerial photographs or 
possibly the high resolution Russian imagery 
DD5) and is not achievable by conventional 
per-pixel classification techniques. Murukami 
and Welch (1992) have attempted an expert 
system approach, Gulch (1991, 1992) used a 
rule-base on initially segmented lines and 
regions, and supported by consistency checks, 
but Shufelt and McKeown (1990), for the most 
successful results so far reported, fused three 
shape-from-shading and one edge-corner 
techniques, taking advantage of redundant data 
and also giving consideration to conflicting 
information