function approach and
to Cho[1995].
a tree, two well known
and forward checking.
problem is rather large.
> unit primitives. Tree
5 and explores fruitless
els of a tree have many
terested in ordering the
tives with fewer label
the tree.
stency of current unit-
low the current level in
ire is modified to be
m in this study. While
el pairs, the modified
number of valid future
s stored in a 2-D table:
label primitives. At the
ward checking sums up
t-label pairs. The total
ird checking is utilized
on.
>s of a stereopair differ
, difficulties in feature
:lational matching must
ng must be controlled,
is a cost function. If the
to nils, the total cost
ides a trivial solution.
does not contribute any
«tent of its usage is not
el pairs that benefit the
pping.
cheme utilizes the A*
ering and the modified
tching problem reaches
\* search is selected in
ching scheme employs
d forward checking and
ge to reach the solution
] up the convergence of
is are locally matched
tion is achieved using a
illustrates the entire
ching scheme.
s the line primitives in
oh some matched pairs
r matched pairs, there
arge search window,
onding features. These
s: a geometric approach
e affine transformation
airs. It is assumed that
criptions is small or
imitives. The detailed
ha 1996
1. Compute the affine transformation between two sets of
matched end points.
2. Estimate the residuals in y coordinates between the
original points and the transformed points and compute
the standard deviation of residuals in y coordinates.
3. Eliminate the points whose residuals in y coordinates
exceed three times the standard deviation.
In this approach, the y coordinates of the matched points are
only considered because the x coordinates correspond to the
object height of the points.
While the geometric approach is designed to detect the
blunder-like mismatches, the radiometric approach is to detect
mismatches more rigorously. The radiometric approach
employs the correlation technique using the normalized
correlation coefficients for the matched end point pairs. Any
matched point pairs satisfying the condition p < po,
are eliminated from the set of conjugate point pairs. The
correlation coefficient threshold p is set to 0.6 in this
study..
threshold
4. EXPERIMENTS AND RESULTS
To assess the feasibility of the proposed relational matching
scheme, a software prototype was developed and experiments
input
stereopair
mies il
Feature Extraction
&
Postprocessing
| 2-D Binary Relation Construction
from 2-D Tree Technique
Extracting
Node Relations
Match Node Relations
by A* Search with
Modified Forward
D Checking
Yes
Mismatch
Detection
Matched
No >. Node Relations
| User Defined |
Base Vector
Compute
| Base Vector
En
e
Compute the Variation of
Geometric Attribute Values by |:
Analytical Function Approach
Creating ;
Initial Matched Pool |
Unit Ordering |:
Local Relational
Matching with
Global Consistency
Modified Forward Checking:
Checking
Globally Matched
Solution
Mismatch
Detection
Compute the Relative |
Orientation with |
Matched Corner Points,
Figure 5: Flowchart of the proposed matching scheme.
with two stereomodels have been performed. Because of a
limited space, the author describes one of data sets used, and
reports and analyzes the major results. All computations, such
as generating image pyramids, image subsampling, feature
extraction, and relational matching were performed on an
Intergraph workstation InterPro 6000.
The stereopair consists of two digitized images depicting the
campus of The Ohio State University at a scale 1: 4000. The
diapositives were scanned by the Intergraph Photoscan with a
resolution of 30um. An image pyramid was generated using a
Gaussian kernel. Images with a resolution of 512 x 512 pixels
were used to extract corner points and straight lines. Figure 6
shows the stereopair superimposed with corner points.
4.1 Feature Extraction
Corner points and straight lines were extracted from the images.
For the Forstner interest operator, f = 1.5, q = 0.75, 97%
confidence level and 7x7 window size, also for nonmaxima
suppression. The linear straight lines were obtained by the
straight line extraction algorithm developed. There are four free
parameters required in this algorithm: minimum line distance,
norm distance, and gradient orientation and magnitude
115
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
