Part B3. Istanbul 2004
Fthe vertical baselines.
nt in the gradient direc-
reduces in a significant
are much smoothly de-
ing "intelligent" thresh-
h easier. The estimation
n the maximum residual
lier and Deriche, 2002).
1al regression is that the
; parameters can be de-
erging gives a minimal
erance on the polygonal
;s when the merging has
given by the user. Once
ie parameters 0 and p of
ll as the variance covari-
hated by using the results
inder the assumption that
: detector have a variance
)
> ratio signal/noise in the
nts for two images.
ted line segments.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
3.3 Vanishing Points Detection
Due to the effects of perspective projection, the line segments
parallel in the 3D world intersect themselves in image space. De-
pending on the orientation of the lines the intersection point can
be finite or infinite and is referred to as vanishing point. There
are many methods for vanishing points detection in the images.
A good review is given by V. Heuvel, (Van den Heuvel, 1998).
Here the vanishing points detection is based on an extension of
the method proposed by Shufelt (Shufelt, 1999) generalised to
multi-head viewing (Figure 7). As seen in (Figure 7 - a) the image
segments of each relatively calibrated view (to an image segment
corresponds a 3D plane intersecting the sphere) are accumulated
on the same plane tangent to the north pole (considering that the
horizontally oriented cameras are on the equator). The most dom-
inant groups of converging lines segments in the image will pro-
duce maxima on the accumulator. To avoid aliasing problems,
each segment (A and B on Figure 7 - b) accumulates between its
uncertainty bounding segments with an ad-hoc weighting func-
tion to take into account the uncertainty on the contour and the
segment extractions, the approch is based on edge error modeling
to guide the search for vanishing points. This fuzzy accumula-
tion thus indirectly takes into account that long segments should
weigh more than small ones. The width of the the segments band
indicates the degree of uncertainty in the orientation of an edge.
(a) (b) (c)
Figure 7: (a) Planar method of vanishing points detection us-
ing the Gaussian sphere. (b) Bounds for line segments in image
space. (c) Bounds for line segments for a fuzzy accumulation on
the tangent plane at the pole of the Gaussian sphere.
Figure .8 shows the accumulator where only two images of one
vertical baseline have accumulated. Of course, for obvious ge-
ometric reasons, the highest gain in precision will occurs when
mixing images from the vertical and the horizontal baselines
when the full system will be operationel.
(a) (b) (c)
Figure 8: (a) & (b) The detected vertical vanishing points for
cach MMS image of the vertical baseline independently. (c)
Multi-view vertical vanishing point accumulation.
Once the vertical vanishing point has been found we accumulate
the non vertical segments on a cylinder tangent at the correspond-
ing equator (considering the oriented horizontal cameras at the
equator) to find all horizontal vanishing points (Figure 9). The
accumulator cells do not need to have a high angular resolution.
Each vanishing point corresponds to a different 3D plane orien-
tation relatively to the image planes. Thus horizontal segments
can be classified and associated to a plane direction. These plane
directions will be used, as shown further, to infirm the 3D planes
extracted from the DSM generated from the vertical baseline
Figure 9: The detected horizontal vanishing points of MMS im-
age pair corresponding to the local maxima of an accumulation
on a cylinder tangent at the equator.
Figure 10: The line segments associated to princicipal orienta-
tions in the scene.
3.4 Digital Facade Models
We have now estimated for each rigid capture the pitch and the
roll of the moving platform. Let us now estimate the relative yaw
and pose of captures at time (t) and (¢ + dt) with the help of the
short vertical base line.
Our short stereo vertical baselines acts as a very precise range
measurement unit. [n our case even with a short baseline favour-
ing image matching, one meter baseline provides a relative depth
accuracy of 5 millimetres on a facade at a distance of 10 meters
(with a disparity estimation accuracy of 0.25 pixels). A dense
raster-based Digital Facade Surface Model (DFM) is processed
by a dynamic programming optimisation method matching glob-
ally conjugate epipolar lines (Baillard, 1997) integrating edges
with subpixel accuracy and adapted to landscapes with disconti-
nuities.
Figure 11: Dense Digital Facade Model computed from the short
vertical stereo baseline at (t) and (tdt).
3.5 Extracting 3D Planes
3D planes and the set of 3D points belonging to these planes are
extracted in the 3D dense DFM with a robust region growing al-
gorithm mixed with a robust estimator RANSAC of Fischler and
Bolles (Fischler and Bolles, 1981). The aim is a robust detection
of the dominant facade planes. We randomly select a triple of
points and evaluate the resulting hypothesis of planes, we perform
a RANSAC based plane detection algorithm in a local neighbor-
hood.
This means that, assuming that we have a sufficient overlap be-
tween two acquisition of the rig, the rotation between two poses
can be estimated by finding the matching planes subsets.