Petra Zimmermann 
  
2 DESIGN OF A NEW SYSTEM 
We decided to "redesign" the AMOBE I system to make it extensible and went from the former procedural language to 
an object-oriented software design. Also the processing is no longer like in a flow diagram fixed, but adaptive through 
the information derived from multiple cues and the class structure. Our system provides a set of classes for feature 
extraction, data storage, reasoning and visualisation, including user interaction. The core is an easy-to-expand 
repository, "toolbox", of basic features and algorithms. Each step enables an exchange between data and processing in 
2D and 3D to improve results on these levels. 
Common systems are often based on only few cues. The requirements of our system are that it should be able to- 
integrate as many cues as necessary and possible, further it should be able to 
= Adapt algorithms and parameters depending on scale, density of buildings, degree of details etc. 
= integrate models and knowledge 
* integrate user interaction, visualise results 
= process top-down and bottom-up 
" integrate control mechanisms 
= and be easy-to-expand through additional algorithms 
2.1 Motivation 
The idea behind this new system is to develop a framework for extracting as much as possible basic robust features 
from different type of data or different cues. These basic features should be extracted and stored for all further 
processes, that means for building detection, recognition and reconstruction processes in 2D and 3D (Table 1). As basic 
cues we assume colour, texture, colour edges, grey-level edges, shadow and elevation information from either DEM or 
DSM. Generally we consider buildings as "blobby" regions with long straight edges and high homogeneity in colour 
and texture domain. 
  
  
  
  
  
  
  
  
  
  
  
  
  
Cue Region- or | Derived Derivable features for | Objects that could be recognised | Information for 
edge-based | information for | recognition reconstruction 
detection 
Colour Region- Colour Roofs, trees, roads as | Homogeneous colour: regions for region-based 
(RGB, based classification, regions and also the | Roofplanes, roads, vegetation (Hue | matching 
HSV, Boundary | homogeneous boundaries green) boundary edges with 
L*a*b*) information | areas, similar areas information about the flanking 
edge-based Inhomogeneous colour: region colour 
small objects, different materials 
Grey-level | Edge-based | Grey-level gradient | Long straight edges, Long straight edges: edges for edge-based matching 
Edges useful for reconstruction | Large (man-made) objects, ridges, 
Gradient in the boundary of roofs, roads, 
Colour colour channels, | Small edges 
Edges gives additional Small edges: 
edges Trees, cars, small objects 
Texture Region- Homogeneous May differentiate | Homogeneous texture: additional information for 
based regions: entropy, | between inhomogeneous | Roofplanes, roads, water, bridges, | region- or edge-based matching 
homogeneity, vegetation areas and | shadow from buildings on e.g. 
correlation, homogeneous man-made | roads 
uniformity objects, additional 
direction, contrast, | attributes for segmented | Inhomogeneous texture: 
regions and edges Vegetation, small objects, shadow 
from trees 
DTM Region- Slope, aspect Gives underlying terrain, | Gives underlying terrain, useful for | basic coarse terrain 
based useful for blob extraction | blob extraction characteristics 
DSM Region- Blobs = regions | Coarse location and | High elevation: gives coarse model of 
based with high elevation | shape information of | Trees, bridges, buildings, noise buildings, ridges, main axes, 
compared to their | roofs, e.g. ridges, size of shape, extend, useful as 
neighbourhood these objects Low elevation: constraint in matching 
Roads, cars, places, lawn, shadow, 
small buildings 
Shadow Region- Colour Occlusions useful for | Occlusions useful for further | indicates "dangerous" region 
based classification in | further reconstruction to | reconstruction to avoid errors where edges and regions may 
Edge-based | HSV space and | avoid errors be distorted through shadows 
elevation in the 
neighbourhood 
  
  
Table 1: Applied cues 
reconstruction 
and data and the derived information 
for building detection object 
recognition and building 
  
1064 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 
buil 
veg 
Lim 
Fig 
ima 
The 
fror 
Scal 
ove 
mot 
  
Fi