International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
/ Building \
N. pointer /
——{ Left image pre-processing )-4—
+ | ee PER OR
/
7
—-
D
sto: -
E Transfering to the building N
space -local coordinate +
System j
ge
Digital images
er oon
| Detecting the 2D buiding | — y3t le
| contour on the left image | 3 Se 2
AN mer.
( Height extraction )— ; | Ea
Merge and
> en oed 5
TU" generalize the MM
contour
SE re EIE TIT
TR stage
— Right image pre-processing er
horse — E = t
stage
rt Y
—r ms is gucci LEN
\
(
| . .. | Detecting the 2D building
i | contour on the right image
N
stage
| Se
|
|
( Transfering to the building '
—B space - local coordinate
system
stage
Figure 1: Flow chart of the semi-automatic algorithm.
The initial vote area is expanded to its neighbors (4 or 8)
provided that they fulfill the criterion that connects them to the
building roof. The building roof is extracted by an iterative
process. The values / and c are calculated at every iteration.
This process is carried simultaneously out in all directions. A
procedure is applied to the extracted building roof by filling the
holes and Sobel edge detector. The created roof edge is
converted from raster to vector.
2.4 Stage 3: Height Extraction
A procedure for finding the homolog point is performed for
every point on the left contour. The search for every point is
performed on a specific area limited by geometric conditions:
the epipolar-line equation and the possible maximum and
minimum building height.
Following is the calculation of the epipolar-line equation (Eq.2)
in the right image for each point (xij, yi; ) on the contour in the
right image (Thompson, 1968) where A is the correlation
matrix:
Xp
: ; (2)
[, yu FHA», |=0
All points on the left contour represent edge points and are
defined as “interest points”. It would therefore be efficient to
use the ABM (Area Based Matching) methods for finding their
homologue points. Cross-Correlation Matching (Eq.3)
calculates the matching at pixel level, where A is a template
from the left image, B is a template from the right image and
A,B are the averages of A and B. For greater accuracy, the
matching at the sub pixel level can be calculated according to
Eq.4 which uses Least Square Matching (Ackermann, 1984),
where Gr (X, Y; ) is the grayscale value in the target window,
G. (e) is the grayscale value in the search window, A is an
affine transformation model accommodating the geometric
716
difference between two images, and h,h, are used for the
radiometric difference modeling.
SEO a
xx(4-4) x xY[s-s5
G, (X,, Y)=h +h, G, (1.3, A) (4)
In calculating the building height one homologue point on the
contour is sufficient (assuming that the building roof is flat).
However, many observations will give a more accurate result.
Therefore, the height (Z) is calculated by using the MEDIAN of
the heights which has the highest value (over 0.994) in the
criterion for determining the optimal match.
2.5 Stage 4: Right Image Operations
After calculating building height (Z), it is possible to transfer
the first vote 7, — (x, , y,, ) to the right image in two steps:
transfering to a local coordinate by using Eq.5
(X4, X4,) * Z —— X,Y,Z and transfering to a right image
coordinate. X , Y, Z ——o9(xc,, yc, ) . It is performed by using
the co-linear equation (Eq. 6). From here on, Region Growing
can be applied in the same way as in the left image.
2.6 Stage 5: Mapping the Object Space
At this stage the building is mapped according to the
information extracted from the images in the previous stages.
The mapping is performed in three parts:
The first is transferring the extracted contours to the object
space using the extracted height. The second part consists of
merging the contours. The advantage is that the information is
derived from both images and many hidden details from one
image can be complemented by the second.
re
