ss of building
e appropriate
base in Fig. 3
itives is con-
cation of the
below). The
ion, possibly
surfaces have
process con-
relations and
ilding model.
e a complete
r, e.g. walls
netric model.
iction step is
select
surface
select
surface
model
object
traction
rification
not ok
teractive
rrection/
xtraction
'S modeling.
ractions: (1)
ace; and (2)
ions for task
2D or 3D, or
mage. These
fine-pointing.
surface prim-
reflect what
lagery. Diffi-
!posite mod-
base is small
ns for intuit-
could also be
4.4). We are
iat additional
pe (e.g., wall,
ctory, school)
irban, CBD),
iction.
mi-automatic
le extraction
urface model
trategy illus-
tlined below.
puts to each
search 2D
[re DSM — window —= feature
ri generation definition extraction
images !
ZI —» model-based 44 mali)
object reconstruction 3D eate
model feature matching
"cloud"
gi oh reconstructed
object
Figure 5: Process flow in surface extraction.
1. Acquisition of a digital surface model (DSM) of the
scene. DSMs automatically generated from stereo im-
agery provide reasonable results if the sampling density
of the DSM in object space is high (Baltsavias et al,
1995). Airborne laser scanning technology promises to
be a viable alternative in the near future (see Sec. 2.2).
2. The 2D (or 3D, in the case of stereoviewing) point-
ing to the surface of interest in an image is used as
the centre of a search window. This window is pro-
jected into all overlapping images using the DSM (see
Fig. 6). The dimensions of the original window are
critical insofar as the entire surface to be extracted
should be contained within it and its associated pro-
jected windows. Minimizing the area encompassed has
operational consequences: image segmentation will be
faster (if performed on-line) and there will be fewer
image features to consider during the feature matching
and surface-forming steps.
te A
\
q
/
+ 7
(c)
Figure 6: Exploiting DSMs in feature matching: (a) principle;
(b) example DSM; (c) projection of manually measured lines
in the upper left image into three overlapping images using
the DSM.
3. Polymorphic feature extraction is conducted in a seg-
mentation step to extract points (corners, junctions),
straight line segments (edges) and homogeneous re-
gions (texture, colour, intensity) in the search windows
defined as above in the images. Image feature attrib-
utes and relations may also be extracted, e.g. in the
form of an attributed graph (Henricsson, 1995) to as-
sist in latter steps, although at the cost of significant
demands on memory and processing time.
4. 3D object reconstruction can best proceed using 3D
features, thus the next step is to derive 3D information,
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
in particular 3D lines, by multi-image matching of the
extracted 2D image features. The DSM is employed
along with the epipolar constraint for the determination
of search spaces.
5. The matching step produces a "cloud" of 3D linear
and point features. Search is conducted for structures
of 3D lines which satisfy the constraints of the selected
surface primitive (e.g., parallelity, orthogonality).
4.4 Implementation Issues
A number of issues with respect to operational implementa-
tion motivate the proposed reconstruction strategy above.
e Where interaction is required, real-time system re-
sponse is needed. Delays, even if short but frequent,
will inevitably lead to its poor acceptance. This implies
the development of computationally efficient tools and
minimisation of search area.
e Demanding operations, such as DSM generation, im-
age segmentation and attributed graph computation,
should be carried out off-line. In addition, systems
should support the batch processing of repeated se-
quences of processing steps, e.g., when a row of same-
shaped buildings or surfaces is to be extracted.
e The system should support customization, e.g. in form
of macros for sequences of processing steps that may
have general application. This extends to the user be-
ing able to contribute composite surfaces to the sur-
face primitive database specific to the user's domain
(see also Sec. 5).
e |t is important that tools for automated reconstruction
include se/f-diagnosis to provide information useful to
user in the verification of reconstruction. This inform-
ation should include quantitative accuracy measures to
relieve the user of laborious checking.
e Capabilities for storing and accessing large image data-
bases, moving fluently between mono and stereo view-
ing, colour and intensity image display, and switching
between multiple overlapping images are necessary.
5 SUMMARY AND OUTLOOK
The conceptual framework for a semi-automatic building
reconstruction methodology from photogrammetric imagery
was presented. This methodology is novel in its suggestion
of an interaction model implementable in an operational con-
text. The key component is a generic object modeling schema
based on composites of primitive surfaces, i.e. CPS modeling.
It was shown that CPS modeling is well suited to both: (a)
interaction, users can convey a decomposition of the building
structure in terms of the visible surfaces in the imagery in a
natural way; and (b) for the automatic extraction of building
shape.
Successful developments of automated procedures for the fol-
lowing tasks can be accommodated in this methodology and
will further reduce interaction requirements:
e DSMs can be used in some circumstances for automat-
ically detecting buildings, i.e. as blobs on the terrain
(Baltsavias et al., 1995; Haala and Hahn, 1995).
e The analysis of the DSM blob detected for each build-
ing may be exploited to automatically select appropriate