Note that the
-coded by the
> analysis, we
butes of the
ilable to us,
used in the
elemental
able, with the
n.
Is on the rigor
atures yielded
mputer vision
ny sources of
odeled such as
ay be partially
inference from
Direct ranging
typically have
remains. This
iple, a building
n one structure
in 3-D, use of
of these tools.
tion conditions
n from sensors
ion with just à
1odalities. This
ypothesize and
| is hierarchical
ses are selected
re then verified
n examined for
inear buildings:
he images (the
e basic features
pod, Texas, site
s between lines
The set of lines
hypotheses art
ram). A pair of
used) and each
ble image lines
Edward M. Mikhail
Basic Description Method (MVS)
Figure 18. Near vertical (left) and oblique (right)
frame photographs of Fort Hood, Texas
The next step is to verify whether the selected
hypotheses have additional evidence in support of
being buildings. This evidence is collected from the
roof, the walls and the shadows that should be cast by
the building. Since the hypotheses are represented in 3-
D, deriving the projections of the walls and shadows
cast, and determining which of these elements are
visible from the particular view point is possible. These in turn guide the search procedures that look in the various
images for evidence of these elements among the features extracted from the image. A score is computed for each
evidence element. Each of the collected evidence parameters is composed of smaller pieces of evidence. A critical
question is how to combine these small pieces of evidence to decide whether a building is present or not and how much
confidence should be put in it. Results shown in this paper use a Bayesian reasoning approach.
Figure 17. MVS Extraction System
After verification, several overlapping verified hypotheses may remain. Only one of the significantly overlapping
hypotheses is selected. The overlap analysis procedure examines not only the evidence available for alternatives but
also separately the evidence for components that are not common. Figure 20 shows the wireframes of the detected
buildings from the pair of images. Note that while most of the buildings are detected correctly, some are missing.
The system presented above relies on image intensities from multiple overlapping images. The performance of the
building detection and description system can be greatly improved if information from other sensors become available.
As described above, our system can take advantage of multiple panchromatic (PAN) images even if they are not
acquired at the same time. We consider two other sources of a different modality.
The first source of additional information is digital elevation models (DEMs). DEMs may be derived from stereo PAN
images or acquired directly by active sensors such as LIDAR or IFSAR. The second source of information is from
multi- or hyper-spectral imagery, such as from the HYDICE or HyMap sensors, which is becoming increasingly more
available.
DEMs make the task of building detection much easier as the buildings are significantly higher than the surround and
accompanied by sharp depth discontinuities. However, DEM data may not be accurate near the building boundaries,
and the active sensors may contain significant artifacts. The spectral information makes it easier to decide if two pixels
Figure 19. Linear segments from image in Figure 18 Figure 20. 3D wireframe model of detected buildings
International Archives of Photogrammetry and Remote Sensing. Vol, XXXIII, Part B3. Amsterdam 2000. 601
