|. Istanbul 2004 
  
coded DSM). 
DISCUSSION 
xtraction in two 
t al., 2001). As 
> been extracted 
system is able to 
ith rather dense 
been evaluated 
lotted reference 
ccording to the 
en, we achieve 
ctness of about 
be linked into a 
cteristics yields 
detour/shortcut 
'ectness) values 
| Data set II: 
81.6 
95.0 
2.5 
1.05 
0.95 
84.0 
100.0 
  
  
  
  
  
  
  
  
  
pad axes. 
segments have 
= 7 by. This is 
nes in both im- 
construction of 
> can be seen at 
| of Data Set II 
road have been 
able to extract 
cular road axis 
1ere the imple- 
nsequence, the 
1 Data Set I to 
y be referred to 
n (Hinz, 2003). 
system extracts 
deficiency ex- 
r vehicle types 
>ck of our sys- 
ns. Hence, be- 
ing connection 
:d towards the 
. As a final re- 
| completeness 
initely encour- 
fact that these 
tise of the sys- 
(as it is surely 
International Archives of the Photogrammetry, Remote Sensing 
  
uem 
8 pean 
eru S 
Leng oy ic v 
(a) External evaluation of road axes. 
^m. u 
"E 
  
        
m ua 
(b) External evaluation of lanes. 
Figure 6: External Evaluation of Data Set I: Reference matching 
extraction (bold); missed reference (thin). 
true for every experimental fully-automatic system at present). In 
this field, we are still at the stage of fundamental research and 
there are still many questions left open and still many steps to go 
so that a state of maturity is reached to envisage a transition to 
operational use. 
6 OUTLOOK — BEYOND ROAD EXTRACTION 
In the last section of this paper, we will show that results like 
those obtained above can give valuable support for other applica- 
tions. We exemplify this by two complementary approaches for 
monitoring traffic in urban areas. The first approach uses optical 
data similar to that used for road extraction, while the second one 
is designed to extract vehicles from thermal infrared data. In con- 
trast to most related work on car detection, both approaches rely 
upon local as well as global features of vehicles. 
6.1 Car Detection in Optical Imagery 
To model a vehicle for high resolution optical data, a 3D- 
wireframe representation is used that describes the prominent ge- 
ometric and radiometric features of cars including their shadow 
region. The radiometric part of the model is adaptive because, 
during extraction, the expected saliencies of various edge features 
    
   
    
  
  
  
  
   
  
    
  
  
  
  
  
  
  
   
    
   
   
   
   
   
   
   
  
   
    
   
   
   
   
   
    
    
   
    
   
   
   
    
    
   
  
  
   
  
   
     
   
   
   
   
   
  
and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
à XE Ed "- *. E bo 
" ades 
ed lanes. 
   
Figure 7: Extraction and evaluation of Data Set II. 
are automatically adjusted depending on viewing angle, vehicle 
color, and current illumination direction. The extraction is carried 
out by matching this model "top-down" to the image and evaluat- 
ing the support found in the image. On global level, the detailed 
local description is extended by more generic knowledge about 
vehicles as they are often part of vehicle queues. Such groupings 
of vehicles are modelled by ribbons that exhibit the typical sym- 
metries and spacings of vehicles over a larger distance. To make 
use of the supplementary properties of local as well as global 
features, the algorithms for vehicle detection and vehicle queue 
detection are run independently first. Then, the results of both are 
fused and queues with enough support from the detailed vehicle 
detection are selected and analyzed for rectangular blobs to re- 
cover vehicles missed during the previous steps (see Fig. 8 a). De- 
tails regarding the implementation of this approach can be found 
in (Hinz, 2004b). 
Typical problems are posed by cars that are not part of a queue 
and whose sub-structures (hood, windshield, etc.) give not 
enough evidence for a successful detection. However, the inte- 
gration of intermediate or final results of road extraction helps es- 
pecially to find such cars, since the road information around a car 
now supplements the (missing) evidence of a car's sub-structures. 
Figures 8 b) and c) show an example of extracting a car between 
the ends of two lane segments. 
6.2 Car Detection in Thermal Imagery 
Compared to optical data, thermal imagery has generally a lower 
resolution and usually a worse noise level because of the higher 
sensitivity of the scanner. However, thermal sensors show also a 
number of advantages—most notably their night imaging capa- 
bility and their potential to derive temperature and temperature 
differences of objects, thus allowing for inferences about the cur- 
rent activity of objects even if they are not moving. For these 
reasons, thermal imagery has become a very attractive alternative 
for monitoring vehicle activity.