International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
4.6 Fusion of SAR data from different views
Starting from the already found best SAR illumination the
remaining set of simulations is scanned for the best
complementing SAR measurement. Again, the portion of useful
data from the object class of interest is maximized. This means
that the objects have to be visible from at least one viewpoint.
The procedure is repeated twice based on the result of the
previous step. The analysis is carried out independently for the
object classes buildings and roads. The benefit for a fusion of 2,
3, and 4 SAR acquisitions can be seen in Table 2.
SAR measurements 1 2 3 4
Roof area (undistorted) 52 73 82 87
Road area (undistorted) 39 47 56 62
27 10 4 2
Road area (shadow)
Table 2. Portion (in 96) of visible roof and road area for the
optimal single and set of several SAR acquisitions
5. DISCUSSION
5.1 Optical data
The detection algorithm was tested on a series of high
resolution aerial images (ca. 15cm ground resolution) of
complex downtown areas. No preclassification of regions of
interest had been carried out. It shows that the vehicle detection
achieves a high correctness of 87% but only a moderate
completeness of 60%. This is a clear consequence of the
stringent parametric model. In contrast, the (independently run)
queue detection reaches a higher completeness of 83%,
however, as can be anticipated from the weaker generic queue
model, the correctness is below 50%. The effect of fusion and
recovering missed extractions increases the completeness up to
79% while the correctness is almost the same as for the vehicle
detection.
Failures occur in regions where the complete road is darkened
by building shadows. This could be overcome by pre-
classifying shadow regions, so that the vehicle model can be
adapted accordingly. Further improvements, mainly regarding
the vehicle detection scheme, include the optional incorporation
of true-colour features and the use of a model hierarchy and/or
geometrically flexible models similar to Olson et al. [1996] and
Dubuisson-Jolly et al. [1996]. The use of multi-view imagery to
separate moving from parking vehicles and to estimate vehicle
velocity would be another avenue of research.
5.2 Infrared data
A comparison of data sets recorded at different time of day,
weather condition, and temperature has shown that the
appearance of roads and vehicles in images differ significantly.
For detection of stationary cars the vehicle model has to
consider the different illumination and temperature situations.
Subject of future work will be the modelling of vehicles under
consideration of the different data statistics.
5.3 InSAR data
The analysis of buildings and roads in urban areas is limited
due to the SAR sensor principle. With a single SAR
measurement useful data can be acquired usually for a minor
part of the object areas only. It was shown that this limitation
can be overcome by taking additional SAR data from different
aspect and viewing angles into account. In case of the roads the
best result is achievable for an illumination from exactly north
and west with a viewing angle of 45°. These directions coincide
with the main road orientations. The third illumination direction
is along the tilted long road from 210° anti-clockwise towards
north (off nadir angle 6 = 50°) and the fourth from east to west
(0 = 60°). However, even in the case of the fusion of four SAR
images still more than a third of the road area can not be sensed
without distortions. Assuming the layover problem might be
resolved using techniques of SAR Polarimetry and/or
Interferometry, it is worthwhile to focus on the shadow areas
alone. In the last row of Table 2 it is shown that only 10% of
the road areca would be occluded combining two carefully
chosen SAR acquisitions.
6. CONCLUSION
Due to the high resolution of the used optical data a detailed
three-dimensional modelling of vehicles was suitable and
powerful for discrimination of these objects from other objects
of the scene. The recognition is supported by utilizing the
illumination situation and analysis of the shadow which is
important in case of low contrast between vehicle and road
surface. Exploiting context information by analysis of vehicle
queues increases the reliability of recognition. In the presented
approach map information was not utilized, but can easily be
incorporated. Analysis of monocular data allows an assessment
regarding movement or activity of vehicle only indirectly by
assignment of vehicles to traffic lanes or parking space.
Particularly at night thermal infrared sensors yields suitable
data for traffic monitoring. The used images of the video had a
lower resolution than the aerial imagery. Vehicles have to be
modelled as elongated blobs. Because of the simple model,
context knowledge from maps and about parking structure was
exploited to achieve a reasonable discrimination of vehicles
from other objects of the urban scene. Thermal images allow
the analysis of recently moved vehicles.
SAR sensors systems are able to monitor urban scenes day and
night and almost independent from weather. In contrast to the
visible and infrared imagery which was taken in nadir view
SAR imagery has to be taken in oblique view which is an
inherent part of the sensor principle. Due to this fact in urban
areas streets close to buildings or between buildings may be
partially invisible or distorted by layover. However, the side-
looking illumination by SAR causes inherent artefacts
particularly in dense urban areas. An analysis of multi-aspect
SAR data offers an improvement of the results. The SAR
acquisition directions can be locally optimized based on DEM
and map data of a GIS.
Intel
——
Brei!
mon
Dres
mov
scen
Dub
segn
PAN
Eck:
Traf
syst
End
mul
End
aper
Gui
Ana
fero
and
Giei
groi
103
Gor
data
Haa
flov
HC
Hin
gen
Patt
Hin
resc
mat
Kir
Mo
mo:
813
Liu
and
Lat
Me
Ge
Sys
Mi
Fal
252
Mi
resi
In:
