The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
278 
much detail, it is impressive to see that their system achieves 
reasonable results for hundreds of buildings and runs on an 
operational basis. (Rottensteiner et al., 2005; Rottensteiner, 
2006) extract low-level features for building reconstruction 
from multispectral images and surface models. Building 
parameters are estimated in a consistent way by considering 
geometrical regularities employing soft constraints, while false 
and conflicting hypotheses are eliminated through a robust 
estimation. Recent approaches make also use of the distinct 3D 
characteristics of buildings and some of them focus also on 
facades, which can be included by acquiring terrestrial data. 
(Dorninger & Nothegger, 2007) present an approach for 3D 
segmentation, which can make full use of high resolution 3D 
points from laser scanning (up to 20 points per m 2 ) or image 
matching showing also at least partially the vertical facades. The 
approach relies on a full 3D representation of planes in 
parameter space and clustering the points in this space. Results 
for points from laser and image data prove that very high quality 
models can be generated. (Zebedin et al. 2006, 2007) show how 
facade planes derived from dense 3D points can be refined by 
sweeping the planes and projecting them into all views where 
they are visible. This does not only lead to very good estimates 
for the plane, but also for the texture as demonstrated by a 
number of convincing examples. An alternative to close the gap 
of missing structures of facades in classical nadir data and 
missing structures from roofs in terrestrial data is to capture the 
scene with an oblique looking sensor. (Hebei & Stilla 2007) 
show this for the case of an oblique-viewing laser scanner. Yet, 
also for this approach, a high precision multi-aspect registration 
of point clouds is necessary for further processing the data. 
A clear tendency can be seen that also recent approaches for 
road extraction integrate 3D information. (Clode et al. 2007), 
for instance, use high resolution LIDAR height and intensity 
data to delineate the road geometry in 3D. Primary road 
hypotheses are generated by classification employing colour 
intensities and height gradients. The result is then vectorized by 
convolving a complex-valued disk (so-called Phase Coded 
Disk) with the image. The Phase Coded Disk represents 
basically the local features of the road model. Center line and 
width of the road are obtained from the magnitude image while 
the direction is determined from the corresponding phase image. 
(Hinz 2004a, 2004b) employs a DSM and multiple-view 
imagery to extract urban road networks. The extraction is based 
on a detailed scale-dependent road and context model to deal 
with the high complexity of this type of scenes. The 
corresponding extraction strategy is subdivided into three 
levels: level 1 comprises the analysis of context, i.e., the 
segmentation of the scene into the urban, rural, and forest areas 
as well as the analysis of context relations, e.g., the 
determination of shadow areas and the detection of vehicles; 
level 2 includes the detection of homogeneous ribbons as 
preliminary road segments in coarse scale, collinear grouping 
thin bright road markings in fine scale, and the construction of 
lanes and carriageways from groups of road markings and road 
sides; level 3, finally, completes the road network by fusing 
road segments detected in overlapping images, iteratively 
closing gaps in the extraction, and exploiting the network 
characteristics to generate a topologically complete road 
network. A key feature of the approach is the incorporation of a 
scheme for internal evaluation. Hypotheses generated during 
extraction are internally evaluated so that their relevance for 
further processing can be assessed. 
Typically, also multispectral information can be exploited since, 
for both airborne sensors and spacebome sensors, the 
acquisition of multi-spectral data in the visible domain has 
reached a resolution regime, in which multi-spectral analysis 
can substantially support the extraction. For instance, (Mena & 
Malpica, 2005) and similarly (Zhang & Couloigner, 2006) use 
colour information to derive various statistical and textural 
parameters. Classifying an image based on these features yields 
potential road segments, which are cleaned and skeletonized 
into road center axes. Such approaches show limitations when 
applied to images of low resolution compared to the object size. 
Dealing with mixed pixels is thus an important issue if roads are 
to be mapped using satellite images. To cope with this, (Bacher 
& Mayer, 2004, 2005) developed a two-step strategy. First, 
training information for a supervised classification is obtained 
from an initial step of road extraction with very strict parameter 
settings. The results are fed into a multispectral classification to 
generate a so-called roadclass-image, which can be interpreted 
as an additional channel. These multi-channel data are 
processed simultaneously with the line- and network-based road 
extraction approach of (Wiedemann & Hinz, 1999). By means 
of this strategy, the linear properties of roads in each channel 
are exploited and - if available - supplemented with area-based 
colour information. An alternative is shown in (Ziems et al., 
2007), where colour information is employed to better identify 
false alarms when determining potential errors in existing road 
databases, e.g., if GIS road axes run through fields or bushes. 
To this end, a statistical analysis of the colour distribution 
derived from potential road areas in comparison with trained 
distributions is carried out. 
2.2 Integration of functional and temporal properties 
With the increasing availability of airborne videos and highly 
overlapping photogrammetric image sequences, also the 
integration of temporal features becomes feasible. These show 
great potential to add very valuable information additionally to 
the geometric and radiometric properties of objects. This trend 
can be seen in particular for road extraction where first 
approaches for road mapping by activity analysis and car 
tracking were recently developed. The work by (Pless 2006), for 
instance, aims at detecting temporal changes in stabilized 
airborne videos. It is based on a generic scheme to discern static 
background from active foreground on the basis of eigenvalue 
analysis. Especially in the case of dense traffic, active image 
regions correspond to the main roads. While this approach is 
mainly designed for the analysis of inner city areas, the 
Bayesian car tracking system by (Koch et al. 2006) is able to 
fuse multiple car tracks from different flight paths. By 
employing a comprehensive Bayesian model for the sensor 
characteristics as well as the detection and fusion scheme, the 
inherent uncertainties in the physical and mathematical sensor 
modelling and track hypothesis generation are handled in a very 
consistent way. As final step, the - potentially interrupted - car 
tracks are transformed into objects space and fused to 
eventually delineate a precise and topologically intact road 
network. 
2.3 Integration of scale-space characteristics 
The importance of incorporating the scale space behaviour of 
objects into automatic extraction approaches has been 
recognized already in the 1990s. It has been shown that 
different levels of image resolution can be linked to certain