International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 Interna 
  
  
  
   
  
  
  
    
  
   
  
  
  
   
  
   
     
    
    
  
     
  
  
       
    
  
  
  
  
  
  
  
  
  
  
  
  
   
  
  
  
  
  
    
i (Road network] 4 I 
il nm mer à Wise Part-of relation 
1 P — connec ME 
| (Road inkk--------7---- #4 Junction J ——-  Specialization 
j ; AE. ren — Concrete relation 4.1 E 
| | Road segment J [c omplex junction | [Simple junction ---*# General relation between objects 
— poa 
CUT a The ex 
at craie ane Ÿ 3 Coarse scale i 
Fine scale | Lan y porallet oarse scale certain 
is Foie are Lane segment ) ploited 
orthogona L————À 1 
EN tractioi 
ORI Ye border On 
| Wake, | Ran [Pavement ; Context Contex 
Line-st E d) (Symbols) | Object se scene i 
Line-shaped Symbols ) ; T€ ~ el s S 
d | marking Lan : Ps oe t di Ere] text rel 
j Me a is aligned N Real possibl 
fil {Long marking (Short marking" 3 Work DAR s 
|] o0 00000 0cM—-—-----------—---------- T-mMe---- E-------------l---- the sys 
1 EVA a A = | : : 
i Long | Short ; SE E Cam Geometry | Road Object | Junction tion of 
i T E E CENTER concrete or concrete or |' in 1 
| coleredline] |colored line | symbols ) l'asptait région | lasphalt région Material x bons ir 
Te n: yat markin 
a or we we 2 Ent re meg en emp a ; divides : 
B : ; à tie : ie road m 
i Long Short Bright Elongated Compact |i| Bright homo- Bright mage = 
iaht li — i ; ; | Substructure ments 
bright line bright line symbols ) |. bright region bright region J {geneous ribbon] | blob QUESUCENTE : 
During 
(a) Road Model (standard case) are iter 
; ; twee 
Figure 1: Model for roads and their context. b: 
Ebner, 
material describing objects independently of sensor characteris- work C 
tics and viewpoint (Clément et al., 1993). In contrast, the image diate s 
level, which is subdivided into coarse and fine scale, comprises For de 
the features to detect in the image: Lines, edges, homogeneous etal., 2 
regions, etc. Whereas the fine scale gives detailed information, vn Ta there e 
the coarse scale adds global information. Because of the abstrac- x A MONS 
tion in coarse scale, additional correct hypotheses for roads can the Bes 
be found and sometimes also false ones can be eliminated based : them ii 
  
dian ^ 2 s 
on topological criteria, while details, like exact width and posi- (a) .engilis of mais 1i and saps issue. 
tion of the lanes and markings, are integrated from fine scale. In pt) 
this way the extraction benefits from both scales. 
     
i 
gap Unark 
   
3.2 Context Model 
Contex 
based i 
analvsi 
  
The road model is extended by knowledge about context: So- 
called context objects, i.e., background objects like buildings or 
vehicles, may hinder road extraction if they are not modelled ap- 
1 propriately but they can substantially support the extraction if 
  
  
(¢) Histogram of lengths (d) Histogram of curvatures 
  
Figure 2: Model for grouped markings: 
  
  
  
  
  
  
E Components used for extraction | Components used for evaluation 
they are part of the road model. We define global and local con- D Orientation difference of pairs of D Lengths of markings and gaps: con- > 
text: markings and gaps: limited stant within group 
> Gap length: limited > Overall curvature 8: low Extraci 
j Global context: The motivation for employing global context D Length of group: lower bound D Height variation of s: low of salic 
. las 4 V roads 
1 stems from the observation that it is possible to classify seman- 
4 tically meaningful image regions—so-called context regions— : a : 2 
where roads show typical prominent features and where certain — 0f a lane. In a very similar way, relations to sub-structures and 
relations between roads and background objects have a similar the integration of GIS-axes—though not used here—can be mod- Road 
importance and characteristics. Consequently, the relevance of elled. Figure 1 b) summarizes the relations between road objects, networ 
i , s 3 ; : » comple 
i different components of the road model and the importance of ~~ context objects, and sub-structures by using the concepts "Lane 
i N : 3 : 23 » : » : CE = 
i different context relations (described below) must be adapted to ~~ Segment” and "Junction" as the basic entities of a road network. 
16 : : : 
If the respective context region. In urban areas, for instance, rela- 
p ; ; . A > . 
i tions between vehicles and roads are more important since traffic 3.3 Model for Self-Diagnosis 
i is usually much denser inside of settlements than in rural areas. J A. ; 
| OL ; ; In order to enable the computation of internal quality measures 
We distinguish urban, forest, and rural context regions, which are j j rum oy «vi 
N den during extraction, the criteria (i.e. model components) defining i 
extracted by a texture-based segmentation (see (Baumgartner et Es ; - à; 
ped an object are divided into two different types. Model compo- t 
al., 1999, Hinz et al., 2001)). : d Y d 
nents of the first type are used to extract an instance of an object 
Local context: We model the local context with so-called con- and the components of the second type serve for evaluating its 
text relations, i.e., certain relations between a small number of ~~ quality. For guaranteeing an unbiased evaluation, model com- 
road and context objects. In dense settlements, for instance, the ponents belonging to different types should be independent from 
footprints of buildings are almost parallel to roads and they give each other. Figure 2 illustrates this for the case of grouped mark- 
therefore strong hints for road sides. Vice-versa, buildings or ings: Components listed in the left column are used to create a 
other high objects potentially occlude larger parts of a road or group of markings. The parameters listed in the right column are 
cast shadows on it. A context relation "shadow", for instance, computed from each marking group and matched to predefined 4 
can tell an extraction algorithm to choose modified parameter evaluation functions according to fuzzy-set theory. The fuzzy ag- à 
settings. Also vehicles occlude the pavement of a lane segment. gregation of all matches yields a confidence value indicating the uai 
Hence, vehicle outlines as, e.g., detected by the algorithm pre- reliability of the extracted object. A description of all involved 
sented in (Schlosser et al., 2003) can be directly treated as parts evaluation models can be found in (Hinz and Baumgartner, 2002). Figure