inbul 2004
The white
Structures.
trong grey
e proposed
ections. To
d edges are
icription of
TURES
lges of the
means rec-
traction of
ar features
connection
ath for the
e between
final road
tions.
the Steger
es are split
Id. This is
»wing step
re fused to
ynstructed.
curvature.
if they sat-
| must be
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
Construction
Construction
Verification of
Line and road section
; ; of road CU
extraction evaluation of fi using line
; section
connection strength image Fusion of Global Road
- linear grouping of
Construction i i
Construction Verification of features road sections network
Edge and road road section
extraction evaluation of 2 using gradient
connection SCHON
image
Figure 3: Workflow of road extraction using collinear features
Figure 4: Extracted lines (left) and edges (right). The original
image is given in figure 2.
linear feature => line or edge
endpoint
connection (straight line)
snake optimized connection
road section \
endpoint
linear feature
Figure 5: Road section, consisting of linear features and a snake
optimized connection.
Gaussian smoothed image, with ¢ = 2.0, in x and y direction,
e the linear features must have a minimum length (117, ,)
e there is a minimum and maximum length for the connection
(453)
The evaluation parameters 4; are not simply thresholded, but they
are interpreted as fuzzy values. The fuzzy variables are combined
into one evaluation value (jjcoN) with the Fuzzy AND operator
(1) (Zadeh, 1989).
HCON Hc, ^uo, ^ue ^in AHBLS ^ BL,
= MIN(HC, ; HC2 ; HC; BEA Eas La) (1)
Connections with a combined evaluation value above a given
threshold serve as basis for the construction of road sections.
3.3 Construction of Road Sections
To determine the actual path of the connection, the two adjacent
endpoints of the linear features are used as start points for a
ziplock snake (cf. fig. 5).
Snakes, also called Active Contour Models, were introduced by
(Kass et al., 1988). A snake is described by geometric (Fy)
and photometric (Fexr:) energies, with Esnake = Fint + Pext-
The goal is to minimize the energy by varying the path of the
snake. Due to the photometric energy the snake is pulled to
image features, whereas the geometric energy usually controls
the tension and rigidity of the snake.
(Neuenschwander et al, 1995) crafted the term ”ziplock
snake”, for which the optimization is performed from both sides
inwards. The advantages of this approach when using it for
bridging gaps in roads are that the given information about the
endpoints 1s exploited well, while local minima, which arise
especially in the middle of the gap due to a bad prediction of
the road path, are avoided (Laptev et al., 2000). During the
optimization process the active parts of the snake, where the
image information is exploited, move step by step from both
sides towards the center.
As roads can appear as bright lines and image edges, a line
strength and a gradient image, respectively, are used as photo-
metric energy. The line strength image is calculated in form of
the maximum negative eigenvalue of the Hessian Matrix (2) for
each pixel.
Jax Jaq
H(z,y) = y y y
duy JYY
(2)
Here Guz, gyy and gay, represent the second derivatives of the
