Fig. 5: Aerial image with overlaid segmentation results for a) roads and b) borders between forests and grassland. c) Height map
with elevated forests, showing the 3D —surface mesh of selected objects from a) and b) as overlay.
Segmentation is solved iteratively using maximum —a—post-
eriori estimation (Gimmelfarb, 1991; Tónjes, 1994).
Finally the contours are smoothed using a contour model
based on Gibbs random fields which favours smooth contours
(Mester, 1988). Figure 5b shows the segmentation results for
two classes of textures.
5. RECONSTRUCTION
5.1 Data driven Reconstruction
The initial reconstruction is data driven and employs
photogrammetric stereo vision. The correspondence analysis
uses normalized cross correlation as cost function for match-
ing of homologous points to determine the height dependent
parallax. A Smoothness constraint is exploited by subse-
quently interpolating continuous regions. Finally the parallax
map is transformed to a height map using binocular camera
geometry (Koch, 1995).
5.2 Model driven Reconstruction
The model driven reconstruction exploits prior knowledge
about object geometry to restrict the parameters for recon-
struction. While data driven reconstruction uses uses only a
few and general geometric constraints, interpretation offers
the facility to exploit object specific geometric properties. In-
terpretations yields the segmentation of aerial images in vari-
ous regions, as forests, grassland, and roads. The location of
these regions is stored in image masks. Scene reconstruction
uses these image masks to apply object specific constraints to
the height map obtained by stereoscopic correspondence
analysis. The prior knowledge forces a height step between
forests and grassland or roads. Further the object semantic
controls mesh generation. Roads are approximated by a sepa-
rate mesh to ensure a continuous course. At the edges of fo-
rests a vertical mesh for the height step is inserted (fig. 5c).
The semantics attached to the model parts allow an object
specific post processing. This offers the facility to refine the
objects artificially by adding details which are invisible to the
sensor from computer graphic libraries. E.g. for close views
872
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
synthetic trees with fine transparent leaf structure are placed
in front of forest edges (fig. 6).
Fig. 6: Close up view of forest edge
6. RESULTS
Figure 7 shows the synthesized view of a landscape model re-
constructed from a pair of overlapping aerial images. The
model generation considered object semantics: roads are
represented by a separate surface mesh and exhibit a continu-
ous course. At the edges of forests a height step was inserted.
The model of the Sieber Valley in the Harz (2km x 2km) con-
sists of approximately 13.000 Polygons and a texture map of
2048x2048 pixel. For interactive exploration of the scene in
real time the model can be visualized on a graphic computer.
Fig
The
Sce:
knc
the
van
mo
tive
olu
tak
hur
edg
Fur
inte
pri
top
Acl
Cle
For
Gir