cal Methods for 
w York. 
vation Models. 
n, February 6 - 
ace Restoration 
Models. Spatial 
Computer Vi- 
ich, September 
onserhaltenden 
rbeit am Insti- 
tät Stuttgart. 
FINDING 3D-STRUCTURES IN MULTIPLE AERIAL IMAGES 
USING LINES AND REGIONS 
Hakan Wiman, Peter Axelsson 
Department of Geodesy and Photogrammetry, Royal Institute of Technology 
S-100 44 Stockholm, Sweden 
e-mail: hakanw @ geomatics.kth.se, pax @ geomatics.kth.se 
ABSTRACT 
A framework for autonomous generation of 3D structures using multiple aerial images is presented. Objects of interest 
are man-made structures consisting of planar surfaces delineated by straight lines, for example buildings. The task is 
subdivided into several stages, including increasing object knowledge. By moving from the image domain to object 
space when searching for correspondences between image features, the advantages of using multiple images are made 
possible. One of these advantages is the added redundancy and reliability to the result. Also, by avoiding image-to- 
image processing, the search space is increased only linearly with the number of images. The Minimum Description 
Length principle is used both for feature extraction and for clustering of features in object space. Currently, only 
buildings with rectangular roof wings can be described. An example is presented for a buildings covered by six images. 
KEYWORDS 
1. INTRODUCTION 
Automated photogrammetric map compilation has 
become one of the largest research topics in the 
photogrammetric community. Most efforts have been 
made to automatically localise and describe man made 
object, s.a. buildings and roads. One of the reasons for the 
desire to automate map compilation is that it is labour, 
and thus cost, intensive. The fundamental change of 
medium for geographic information, from paper sheets to 
computer data bases, and the interactive way maps 
thereby can be created, also make new demands on 
geographic data capture. Such demands are made for 
three dimensional city models, which are asked for by 
many users for eg. city planning and tele 
communications. A 3D city model requires the entire 
building volumes to be mapped, which, without the aid of 
automatic mapping, is a most elaborate exercise. 
2. OBJECTIVES OF THE STRATEGY 
A strategy for autonomous generation of 3D structures 
was developed with the following basic objectives: 
- to use multiple images 
- to use parallel search strategies for evidence 
- to use object knowledge to constrain the search 
domain 
- to work in object space when possible 
When using multiple images, more information becomes 
available for the evaluation task. If evidence is weak in 
one image it may be found in another. Also, the problem 
of occlusion may be reduced significantly. A drawback 
with multiple images is however the complexity of the 
correspondence problem. Moving to object space when 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
image understanding, object reconstruction, object space modeling 
analysing more than two images simultaneously has 
several advantages, e.g., the strength of geometrical 
constraints, but the major one seems to be that the 
complexity of the correspondence search can be treated in 
an efficient and rational way. More images just adds more 
information while the search space remain the same. This 
is still only possible if the search procedures are designed 
as being independent of the order in which the evidence is 
collected, i.e., parallel procedures. The advantages of 
parallel procedures are thus twofold: (i) it is easier to 
avoid local minima since all evidence are treated 
simultaneously and (if) the search space is in principle 
independent of the number of evidence. For complex 
structures, s.a., buildings a number of generalisations and 
constraints must be imposed on data if such parallel 
search procedures are to be designed. These constraints 
come from the object knowledge and excludes objects 
outside a defined category. 
Following this discussion, we believe that moving 
to object space when solving the corresponding problem 
task is a necessity if the goals of such a system are to be 
met. 
3. FINDING 3D-STRUCTURES 
3.1 System Outline 
We will present a system for autonomous 3D description 
of buildings. There are no semantic rules, like e.g. 
windows are surrounded by walls, so in principle any 
imaged 3D structure fulfilling the postulated geometric 
and radiometric criteria could be described. We will 
however refer to buildings, since the main objective is to 
describe them in 3D. In order to reduce the complexity of 
the task, the following criteria have been formulated: 
- One building at a time is analysed. 
953