The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
436 
2. RELATED WORK 
The topic of this paper comprises two important aspects: first, 
the automatic scale-dependent adaptation of object models and 
secondly, the modelling of local context objects and their 
incorporation into the extraction model for the primary object, 
which in our case complies with roads. 
2.1 Automatic adaptation of object models to a coarser 
image resolution 
The task of automatically adapting image analysis object 
models to a coarser image resolution was only rarely tackled 
previously. Heuwold et al. (2007) presented an algorithm for the 
adaptation of object models to a coarser target resolution. The 
adaptable models can consist of arbitrarily oriented line-type 
(ID) object parts and/or 2D object parts of area-type and are 
represented in form of semantic nets (Tonjes et al., 1997). At 
first, the high-resolution object model is decomposed into 
groups of object parts, which influence each other mutually 
during the reduction of image resolution (this influence is 
denoted as interaction). Based on linear scale-space theory 
(Witkin, 1983; Koenderink, 1984), the scale behaviour for each 
of these interaction groups is predicted using analysis-by 
synthesis. The scale change analysis employs the blob detection 
algorithm by Lindeberg (1994). The analysis takes into account 
possibly occurring events in scale-space and predicts the shape 
and radiometric attributes of the individual object parts in the 
target resolution. These predicted low-resolution attributes serve 
as parameters for the respective feature extraction operators. In 
the last stage (fusion), all object part groups with their predicted 
appearance and adapted operators are composed back to a 
complete object model for extraction of this object in target 
resolution images. 
While the adaptation of landscape objects with all types of 
object parts can be carried out, using the described approach the 
appearance and approximate relative location of the object parts 
must be specified in the object model to be adapted. Thus, an 
adaptation is up to now not possible with this algorithm for 
objects with unknown position. However, experiments with 
adapted object models for roads in image data of suburban areas 
(Heuwold, 2006) pointed out the need for incorporation of local 
context in order to achieve a high completeness for road 
extraction, as road extraction often suffers from the presence of 
nearby objects and their influence on the image data. 
As far as we know only a single study concerning the scale 
behaviour analysis of local context objects exits (Mayer & 
Steger, 1998). A road lane with a vehicle is analysed in linear 
scale-space in order to choose a suitable image resolution for 
which an occluding vehicle does not disturb the extraction of 
the road. The scale for which the vehicle vanishes could be 
predicted analytically. However, the method of this study is 
only applicable to simple scenes, because the scale-space 
behaviour of more complex scenes cannot be predicted 
analytically. 
2.2 Exploitation and modelling of local context 
Local context objects can hinder the extraction of a landscape 
object by occlusion of or casting their shadow on the landscape 
object one is interested in, as context objects normally are 
situated very close to or on top of it. On the other hand, context 
objects can also support object extraction, if their pure existence 
is inferred. For instance, if a car is found in the image, with high 
probability a road must exist at this place. For road extraction 
the incorporation of local context proved to substantially 
support extraction robustness in complex scenes (Baumgartner 
et al., 1997; Mayer, 2008). 
As local context can greatly support a successful object 
extraction from remote sensing images, much work has been 
carried out studying the influence and exploitation of local 
context for image analysis. In general, reasonable context 
relations depend on the type of the individual context object and 
can be very diverse. For our application (roads) some of the 
proposed context relations are: occludes, casts shadow on, is 
close to, is parallel to for the most important context objects 
vehicles, rows of trees and buildings (Butenuth et al., 2003; 
Hinz & Baumgartner, 2003; Hinz, 2004; Gerke, 2006). The 
context objects to be used in the road extraction process can be 
extracted from images using object models, such as in (Hinz & 
Baumgartner, 2003) or derived from GIS, e.g. in (Zhang, 2004). 
With growing demand for traffic monitoring in recent years 
many studies aim for vehicle extraction resulting in ample 
literature. Thanks to the availability of more visual remote 
sensors with high spatial resolution, explicit modelling with 2D 
or 3D models for such relatively small objects like vehicles 
became possible. Related work on vehicle extraction from 
visual image data can be divided into two groups based on the 
type of model used (Stilla, 2004): either appearance-based 
implicit models using training data or explicit models 
describing the vehicle in 2D or 3D. Most explicit approaches 
use edge extraction and subsequent grouping to a 2D rectangle 
applying methods such as the Generalized Hough Transform, or 
classification in a feature space using a Bayes net (Liu et al., 
1999), or a 3D wire frame model for the vehicle (Zhao & 
Nevatia, 2003). Explicit object models with high amount of 
vehicle details for high-resolution aerial images were proposed 
by Hinz (2004): A car is geometrically modelled by a 3D wire 
frame representation containing front, windshield, roof and 
hood as substructures. Colour constancy between hood and roof 
is included in the model as radiometric feature. Based on sun 
direction and image orientation parameters a shadow projection 
is computed from the wire-frame model in order to adapt the 
expected saliency of extracted edges. 
3. ADAPTATION OF OBJECT MODELS 
CONSIDERING LOCAL CONTEXT OBJECTS 
In this section a new concept for the automatic adaptation of 
image analysis object models considering local context to a 
lower image resolution is introduced and explained in detail. At 
first, several possibilities for the adaptation concept are 
discussed leading to the developed algorithm. Secondly, the 
adaptation method and its prerequisites regarding the modelling 
of local context and the incorporation of local context into the 
extraction strategy are described. 
3.1 Adaptation concept 
The exact position of local context objects in an individual 
scene is usually unknown beforehand. Roads, for instance, can 
be flanked by trees at their margin, but it is unknown where and 
if there are trees. Vehicles can even occur on all parts of a road, 
anywhere. While for the instance net of a particular scene the 
exact relative location of the objects is known, the concept net 
of the primary object needs to be very flexible with regard to the 
relative position of potential local context due to this location 
variability of local context. For our approach a scale-dependent 
adaptation of the concept net is desired. However, the behaviour 
of objects during scale reduction greatly differs for different 
relative object positions. For example: while for object A and B