In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds), IAPRS. Vol. XXXVIII. Part ЗА - Saint-Mandé, France. September 1-3, 2010 
Snakes or parametric active contours are a well-known concept 
for combining feature extraction and geometric object 
representation (Kass et al., 1988; Blake & Isard, 1998). They 
explicitly represent a curve with respect to its arc length. In the 
standard formulation they cannot handle changes in the 
topology such as splitting and merging of entities (Mclnemy & 
Terzopoulos, 1995). This is not a problem for the adaptation of 
the 2D vector data to ALS features, because the initial topology 
is taken from the GIS data base and should be held fixed during 
the process. For that reason the network-snake algorithm of 
Butenuth (2008) is used with new definitions of the image 
energy functions. The basic concept of snakes is widely used in 
image and point cloud analysis as well as GIS applications. For 
example, Burghardt and Meier (1997) suggest an active contour 
algorithm for feature displacement in automated map 
generalisation, and Cohen & Cohen (1993) introduce a finite 
elements method for 3D deformable surface models. Borkowski 
(2004) shows the capabilities of snakes for break line detection 
in the context of surface modelling. Laptev et al. (2001) extract 
roads using a combined scale space and snake strategy. 
In order to extract roads from ALS data, Rieger et al. (1999) 
propose twin snakes to model roads as parallel edges. This 
integration of model based knowledge stabilises the extraction 
and is able to bridge gaps in the structure lines in the vicinity of 
roads, which are often not continuous in nature. Road extraction 
can also be improved by fusing ALS and image data, e.g. (Zhu 
et al., 2004), as well as GIS data (Oude Elberink & Vosselman, 
2006). The ALS intensity values, assumed to be a by-product a 
few years ago, can also be exploited in the extraction process. 
Roads have usually small intensity values and can be 
distinguished well from other objects by this feature along with 
the fact that they are situated on the DTM (Alharthy & Bethel, 
2003; Clode et al., 2007). ALS data have also been used to 
detect bridges (Clode et al., 2005; Sithole & Vosselman, 2006). 
In our previous work we used network snakes (Butenuth, 2008) 
for adapting 2D road vector data to ALS intensity and height 
data (Goepfert & Rottensteiner, 2009). The image energies 
consisted of a combination of the ALS intensity, the DSM 
heights, and a smoothness term derived from the DSM. As 
roads are situated on the terrain, smoothness should be derived 
from a DTM. Furthermore, using the raw DSM heights for the 
image energy, the method cannot be applied to areas with 
undulating terrain. Another problem of the existing method is 
that it might be negatively affected by buildings and bridges. 
Buildings sometimes have similar ALS intensities as roads, 
which in densely built-up areas may cause the snake to be 
caught in a local minimum. Considering bridges is essential 
because they have a disturbing effect on the road that passes 
underneath the other one. By the new definition of the image 
energy our method should become more generally applicable. 
2. METHOD 
2.1 General Work Flow 
In this paper a top-down method using the concept of network 
snakes for adapting road networks from ATKIS data base to 
ALS data is proposed. The initialization of the snake and 
therefore the internal energy are obtained from the vector data, 
whereas the ALS information defines the new image energy 
forcing the snake to salient features (cf. section 2.3). Compared 
to our previous work (Goepfert & Rottensteiner, 2009), we 
improve the image energy by terms related to the smoothness of 
the DTM and by terms derived from building outlines and 
bridge positions. Extracted buildings are used to act as 
repulsion forces in the image energy, whereas bridge detection 
is performed in order to determine confident areas for the 
correct road position. After defining and weighting the different 
terms of internal and image energies the iterative optimisation 
process is started modifying the position of the network snake. 
The change of the position of the contour in the current iteration 
is used to determine the convergence of the algorithm. 
Afterwards, the new position of the contour should match the 
corresponding features for the road network in the ALS data. 
2.2 Snakes and network snakes 
It is the general idea of snakes that the position of the contour in 
an image is determined in an iterative energy optimisation 
process. An initialisation of the contour is required. Three 
energy terms are introduced by Kass et al. (1988). The internal 
energy E in , defines the elasticity and rigidity of the curve. The 
image energy E imagc should represent the features of the object 
of interest in an optimal manner in order to attract the contour 
step by step to the desired position. Additional terms (constraint 
energy E co „) can be integrated in the energy functional forcing 
the contour to fulfil predefined external constraints: 
E'snau, = J(£ inl (v(s)) +E imege {v(s)) + E con {v(s)))ds (1) 
0 
where v(s) = (x(j), v(i)) is the parametric curve with arc 
length s. In order to obtain the optimal position of the snake in 
the image, the energy functional in Equ. 1 has to be minimised, 
e.g. by variational calculus. The internal energy can be written 
as (Kass et al., 1988): 
£iB|(r( - )) _ «(M-lv,(^)| : +/?fG-|v, t G)| 2 (2) 
where v s and v„ are the derivatives of v with respect to s, and a 
and /? are weights. The first order term, weighted by a, is 
responsible for the elasticity of the curve. Due to the arc length 
minimizing effect, high values of a result in very straight 
curves. The second order term, weighted by /?, forces the snake 
to act like a thin plate and determines the rigidity of the curve. 
High values of [i cause a smooth curve while contour parts with 
a small /7 are able to model the behaviour of corners. Using the 
idea of network snakes (Butenuth, 2008) allows exploiting the 
initial topology of a network of lines during the energy 
minimisation process. The individual lines of the network are 
connected via nodes of an order higher than two at the junctions 
of these lines. The internal energy has to be modified so that 
this initial topology is preserved in the resulting line network. 
This means that at junctions, the elasticity term in Equ. 2 is 
disregarded (a = 0), whereas there is one smoothness term 
(weighted by /7) per line intersecting at the junction node 
(Butenuth, 2008). The image energy has to be defined in a way 
to ensure that the snake is attracted to image features that are 
characteristic for the object to be extracted. Thus, model 
knowledge is to a large degree incorporated into the image 
energy. Our new definition for the image energy, including the 
integration of a building mask, the extraction of bridges, and 
planar features in the DTM is explained in Section 2.3. We do 
not use any constraint energy terms in our method. 
2.3 Image energy 
The image energy E image consists of three different components, 
namely a general ALS energy E ALS , a building term E buM that 
repulses the snakes from buildings, and a bridge term E Mdge 
attracting the snake to bridges detected in the ALS data: 
164