Edward M. Mikhail 
work. These three points define an initial model containing four road segments and the intersection that they define. An (e.g 
example of this initial model is shown in Figure 26. Each intersection has four road segments and each road segment from 
connects two intersections. The verification problem is to determine whether each of the road segments exists. Figure 
27 shows a small portion of such a grid model which must be verified. Since actual road widths (in meters) vary from 
scene to scene, we allow the user to adjust the default width to fit the particular scene. A width refinement step, later in 
the extraction procedure, reduces the need for exact initial widths. 
  
  
  
  
Road Grid bas 
TT 
  
  
Figure 26. Initial Grid Model 
  
Figure 27. Small portion of road grid model 
  
  
  
  
The grid extraction procedure is composed of two phases, the first tests each intersection (using the four road segments 
for the model) to find which ones are supported by the image data. This hypothesize- and-verify phase propagates the 
grid across the entire scene and provides an initial geometric match for the scene. Figure 28 outlines this phase of the 
  
  
  
   
  
  
  
  
  
  
procedure. The second phase uses this initial match and tests triples of road segments (three consecutive road segments) imag 
to find the best location and width for each triple. Figure 29 illustrates this second phase. These results provide the input mult 
for further use of context and refinement using other data sources. 
A de 
(i.e. 
, : othei 
Grid Refinement featu 
the t 
Input Images 
Match Image and 4 V 
Model Features 
of Intersection Grid A of Road Segment 
Description Description Select Good Alit 
; Select Good Matches obj € 
umm Adjust Position ' bene 
Propagate Grid and Width corre 
asses 
Figure 28. Road Grid Initial Verification Figure 29. Road Grid Refinement throt 
of oc 
time. 
Use of digital elevation models, or DEM, helps in the refinement of the extracted road segments. While a DEM has 
many problems and may not be exact, it provides a good approximation of elevation to determine when a road segment The 
is higher or lower than the others in its extended street. Figure 30 shows a small portion of one image with the matched Rese 
road segments color-coded: consistent segments shown in grays, inconsistent segments in white. In this case, consistent inclu 
means that the average elevation of the road segment as given by the DEM is similar to the average for the extended Metl 
street, inconsistent means the elevation is much higher than the average for the extended street. gii 
allov 
The DEM is used in two ways, first the road segment is shifted (perpendicular to its primary direction) to a minimum over] 
elevation location. Rather than allowing arbitrary shifts, the distance is limited according to the quality of the geometric phot: 
match (with a perfect match the segment will not be shifted). The results of this refinement step are shown in Figure 3l 
where the consistency measure has been recomputed using the new locations of the road segments. Even with the shifts IM | 
er 
  
604 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.