XIX-B3, 2012
terval of time from
rm analytic relative
ng the differential
has two significant
1jugate points in the
ise ground control
images can be used
ects.
ient the multiview
. It states that the
lie on the epipolar
ion is expressed as
GB)
and d4, d; are the
ective centre of the
presents the vector
on (3) is linearized
yaw values. Three
'ompute differential
ve centres and the
obtained from
supplied with the
Coefficients
ls are being used as
primarily due to the
ymplex; they need
good understanding
ther hand, rational
supported by major
Moreover, using
sensor model makes
er, it is important to
ensor model and the
puted using terrain
ysical sensor model
per the method
pace co-ordinate is
ordinate using the
| et al, 2004). The
mly spaced grid of
using the physical
orresponding image
ational polynomial
ynomial coefficients
>. Since the rational
er processing, it is
eo images obtained
v-1/2, and IKONOS
d for extraction of
ication of conjugate
d as mathematically
ransformed to well
raint. One possible
probable match by
ding reconstruction
racted and matched
s (Baillard C, Park,
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
S, Schmid,C.) Some researchers proposed use of colour
information for edge extraction and line segment stereo
matching[Ok, A.O, Scholze, S.], our approach does not attempt
to extract and match line segments for two reasons (1) only flat
rooftops are modelled which means that same height is assigned
to multiple points, thus, it is sufficient to match points on the
edge of the building to compute height of the building. (2)
Building boundaries are digitized, refined and connected on
nadir image. These boundaries along with the height
information analytically transferred to the ortho image as the
relation between geometrically uncorrected image, epi-polar
image and ortho image is established using the sensor model.
If two overlapping images are relatively oriented then the
disparity between two conjugate points is due to the topographic
relief. Based on the geometry either the set of points are
matched in original stereo pair or the other image is re-sampled
to form set of epi-polar images. In case of Cartosat-2, often the
angle about roll axis is significant; this means that the disparity
due to terrain variation will not be restricted to one dimension.
Image matching strategy has to take into account the image
acquisition process. The disparity map obtained in lower
resolution is utilized to guide the matching in next step; In both
cases geometric constraint is used to limit the search space. The
dense DSM is generated using area based matching techniques
performed on epi-polar images, while feature based matching
matches the edges of buildings in raw as well as in. epi-polar
images using the geometric constraints. For both the image
matching techniques, normalized cross correlation is used as a
similarity measure. In both the matching approachs, template
size is increased dynamically. The similarity measure of
normalized cross correlation initiates with a small template and
two thresholds, termed as noise and acceptance threshold. If the
normalized cross correlation coefficient for the template is more
than or equal to the acceptance correlation threshold, the point
is checked for forward and reverse matching. It is accepted as a
match point if the correlation coefficient is more than the
acceptance threshold for both the directions. If the correlation
coefficient is less than the acceptance threshold but higher than
the noise threshold, the template size for reference and search
space is increased and correlation coefficient is recomputed.
The noise threshold is selected as 0.4 and acceptance threshold
is selected as 0.9. Initial window size of the search image is 13
by 13 which is increased in steps of two pixels. The process is
repeated for at least three different sized windows. This process
is repeated for each level of image pyramid .The unmatched
points are transferred to the next level using interpolation. The
interpolation for unmatched point is not done for matching of
original resolution images.
2.6 Computation of Normalized DSM
A Digital Terrain Model (DTM) is the elevation model of the
landscape which does not include above ground objects. On the
other hand, a Digital Surface Model (DSM) includes the objects
with their heights above the ground as well as the topography.
The man-made objects with different heights over the terrain
can be detected by applying a threshold to Digital Surface
Model. The DTM is estimated using mathematical morphology.
The morphological operators help in bringing the background
terrain from the DSM. The above ground objects are detected
using the DSM and morph output. Two morphological
operators, namely “opening” and “closing” are used iteratively.
The size of the window depends upon the size of the building.
The normalized DSM is generated by subtracting the DTM
from the DSM. Segmentation and area filters detect the
buildings of desired size. Finally the building outline can be
constructed using the neighbouring gradients. For the better
definition of the building, the gradient information from the
gray image is also used. The DTM provides the ground height
which is used as an input for computing the building/object
height.
Fig. 6 Derived DSM
Fig. 7 Normalized DSM
2.7 Development of 2-D and 3-D Digitization Tool
Automatic detection of edges and grouping them to form a
meaningful entity has been an area of research for a long time.
Attempts to find a completely automatic and successful solution
for these problems is an active area of research. A 2-D
digitization option is used to manually digitize the building
outlines in one image. These manually digitized boundaries are
further refined using the Canny operator. In this case, the edges
are found only in the neighbourhood of the manually digitized
edge to ensure better localization. These edges are matched in
other images using the geometrically constrained image
matching procedure. In case the image matching procedure is
not able to find corresponding edges in another image,
digitization is done in 3-D viewing mode. User can move cursor
in Z direction and place each vertex of the feature at any
position in depth during digitization and fuse the cursors.
Positions of cursors in left and right images are recorded.
Options are available to draw points, lines and polygons. Fig. 2
shows digitized building boundaries for a portion of
Washington image.
Fig. (2): Digitized building boundaries
The process improves edge localization and minimizes the
effort of manual identification of edges precisely.
2.8 Refinement the Edges
Digitization of buildings is the major step in 3D site modelling,
fully automatic methods for building extraction are not enough
matured as a result we prefer semi automatic digitization
method. The precision of edges is attained by refining the
digitized edges to the nearest real edges which are lost in
manual digitization. This is performed in the following steps,
1. Digitize manually in the proximity of the intended building.
2. Refine the edges that are manually digitized in step-1 by
densifying and constructing a neighbourhood of each of
these points and searching for real edges in canny edge
image.
