International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
Figure 5: Line sweep for surface outline determination in a per-
spective view: Points on the surface (red), the center (black
dot), the four sweep lines (black lines) and their sweep directions
(black arrows).
all isolated points on the plane. Second, we perform a cluster
analysis to obtain different surfaces on the same plane. Then for
each cluster, we construct the convex hull, as we found that it
gives a rough estimation of the surface outline. Finally, to deter-
mine a more accurate estimate of the surface outline in the form
of a rectangle, we perform four line sweeps in every cluster start-
ing at its center, two along v and the other two orthogonal to v
and the surface normal (Fig. 5). Candidates for the borders of
the rectangle are taken from the convex hull and are scored by
counting the number of inliers supporting these lines. The best
line model per sweeping direction is chosen and a least-squares
fit is performed for its inliers. In our experiments, we were more
successful with this procedure than by analyzing the point den-
sities of inliers at each sweeping step. This is due to the very
high point density variations on doors and windows. Finally, we
check, if there are close parallel and overlapping surfaces. In this
case, we merge the points from those surfaces and determine the
outline of the combined surface.
Construction of Surface Adjacency Graph: In this step, we
derive topological information for the surfaces and adjust the cor-
responding parameters. After checking all pairs of distinct sur-
faces, we recursively connect adjacent surfaces considering only
pairs of non-parallel surfaces. If a vertical surface border ends
inside another surface, both are connected to each other and the
smaller surface is interpreted as building part, such as a balcony
or an oriel. If two surfaces have very close borders, we connect
the surfaces, interpreting them as walls. This is repeated until no
further surfaces can be connected. After these steps gaps may
still remain in the boundary of a building. They must be closed,
if a closed polyhedral model for the building should be achieved
(Fig. 6).
5 EXPERIMENTS
We implemented our algorithms using the point cloud library by
Rusu and Cousins (2011) and we tested our approach on several
point clouds from image matching. For image orientation and
derivation of a sparse point clouds, we used software similar to
Snavely et al. (2006); Bartelsen and Mayer (2010). The sparse
point clouds often consist of approximately 50 to 100 thousand
points. These can be highly unequally distributed and often have
huge gaps in the data, especially in the shadow areas. Addition-
ally, we used the semi-global matching by Hirschmitller (2008) to
derive significantly denser point clouds containing points in the
range of a million and above.
148
Figure 6: Rectangles derived from the segmented planes to serve
as the basis for constructing an adjacency graph.
We have tested our examples on a dozen point clouds derived
from images at a small village in Southem Germany, and one ad-
ditional data set showing the castle of Ettlingen, taken from the
benchmark data set of (Strecha et al, 2008). On the one side,
we tested our approach on data derived frm simply shaped build-
ing, and on the other side, we used more difficult objects as L-
shaped buildings, or buildings with highly decorated facades or
with many other parts, such as balconies, oriels, and stairs. We
present some examples on the performance of our algorithm in
figs. 7 to H2.
During our experiments, we restricted ourselves during all
RANSAC searches by using a unique number of 1000 iterations.
However, due to variation in scale from one dataset to another,
we adapted the RANSAC threshold according to a known metric
scale of each input point cloud. Thus, the absolute value of this
threshold is approximately 20 cm.
When inspecting our results, we always find the major four verti-
cal walls af each building. Attached building parts, such as bal-
conies and oriels, are not included in our building models, so far.
We are able to detect smaller surfaces corresponding to minor
building parts, but we did not consider them when constructing
the surface adjacency graph. We will carry on with this refine-
ment step, when modeling building on the next level-of-detail.
Then we also consider to close the building models by additional
roof and ground surfaces to produce consistent LOD 2 models.
Figure 7: Left Image of building consisting of four major walls
standing on a slope. Right: The derived point cloud consists of
76000 points, the four major walls were successfully detected,
segmented and connected.
6 CONCLUSION
We presented an automatic approach for deriving cuboid-based
building models from point clouds reconstructed from multiple
Intern
Figure &:
standing
273-000 ;
segmente
Figure 9:
having a
derived |
walls we:
images.
inance a
with poit
ing scale
robustne:
In future.
nizing th
are able
ing, viel
LOD 1 1
with mot
future wi
such as s
Agarwal
and Sz
43(6),
Almansa
ing Fx
Trans:
25(4),
Figure t
showing
sists of €
tected, s
