Beijing 2008
The International Archives oj the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008
439
>cal context
for the initial
le road model
he relation to
m or optional
local context,
i vehicle.
i
i Mm
etocnlifti
5 for
in Figure 5, is
5 one of Hinz
1 suffice for
ncept. A large
bject parts as
ss are given in
’art-of
Spatial relation with
arallel [d]=] and
lerpendicular
listance [dx=]
3m/pixel
markings and
>ld (2006). The
growing for
scted by shape
id radiometric
ts of the road
ution yields an
:ing left at the
road extraction
Figure 6. Road extraction result (black/white) superimposed on
example image in 0.03m/pix and region of interest
(ROI) for local context extraction
4.2 Adaptation Process
Both object models for road and vehicle are adapted separately.
The adaptation of the road model was presented by Heuwold
(2006). The adaptation process for the vehicle object model is
explained in the following. The intermediate steps of the
analysis-by-synthesis for the scale behaviour prediction of the
vehicle are displayed in Figure 7. Note that the steps of the
analysis-by-synthesis are depicted here in a single synthetic
image for better illustration although there is no interaction
between the vehicle parts.
f.l
Figure 7. Intermediate steps of analysis-by-synthesis for
minimum (upper row) and maximum constellation
(lower row): synthetic image initial resolution L 0
(a.+e.), synthetic image target resolution L Rt (b.+f.),
blob extrema E 0 (red/white) and E, (green/black)
superimposed on L R , (c.+g.), blob support regions
Suppo (red/white) and Supp, (green) on L R , (d.+h.)
No interaction of the vehicle parts takes place for the example
target resolution /?,=0.2m, since they are situated too far apart
for this resolution change. In order to predict the scale
behaviour for the vehicles the minimum and maximum
configuration with regard to the defined attribute ranges from
the initial resolution is to be analysed. The scale behaviour
simulation for the vehicle is carried out with the maximum
contrast, since the initial object model does not contain any
restriction with regard to grey value, i.e. all values are possible.
Thereby, the most “optimistic” case is simulated, leading to the
maximum lifetime in scale-space. As explained in section 3.4
this procedure can be chosen, because the vehicle has an
optional context relation. Neither for the minimum nor the
maximum constellation a scale event takes place, but the
attributes of the vehicle parts have to be changed for the target
resolution. The attributes of the nodes and the spatial relation as
well as the shape parameters for the feature extraction operators
of the vehicle parts are automatically derived from the blob
support regions in target resolution Supp,.
4.3 Adaptation Result
The adapted road model is depicted in Figure 8. Details with
attribute values of the nodes and spatial relations can be found
in Heuwold (2006). The adapted vehicle object model is
displayed in Figure 9 showing all attributes of the individual
nodes and the spatial relations. Applied to the example image in
target resolution the adapted road object model of Figure 8 also
suffers an extraction failure in the area with the vehicle. This
gap can again be filled with the adapted vehicle object model
(Figure 9). Figure 10 shows both the road extraction gap and the
vehicle extraction.
Figure 8. Adapted road object model for target resolution
^?,=0.2m/pix
Vehicle
Front
[d. =2-4]
►
Trailer
Rectangle
Rectangle
wid=12-15
wid=12-15
len=10-12
len=18-87
gv= 1-255
gv=1-255
Figure 9. Adapted vehicle object model for target resolution
^,=0.2m/pix
Figure 10. Road extraction failure for example image in
i?,=0.2m/pix and region of interest (ROI) for local
context extraction
5. CONCLUSIONS
In this paper a method was presented for the automatic
adaptation of image analysis object models, which consider
local context objects, to a coarser image resolution. An example
for the adaptation of a road model with vehicle as local context
