IMPROVEMENT OF MAPPING ACCURACY BY APPLYING TRIPLET MATCHING
TO SPOT IMAGERY
Ryosuke Shibasaki
Public Works Research Institute
Ministry of Construction
1, Asahi, Tsukuba, Ibaraki, Japa.n
Shunji Murai
Institute of Industrial Science
University of Tokyo
7-22,Roppongi,Minatoku, Tokyo, Japan
Commission IV/3
ABSTRACT
One of serious problems of stereo matching based on image correlation is
that it is impossible to verify whether a recognized conjugate point is correct
or not.
The past experiences showed that a point with ma.ximum correlation is
not always a correct conjugate point. In order to improve this weakness,
authors proposed two kinds of triplet matching methods, in which the coordinate
of ground point is given as that of a intersection of three stereo rays.
These
methods were applied to SPOT images with three sights, that is. right, center
and left. The results of test cases using SPOT images showed matching accuracy
and reliablity were improved by applying the triplet matching methods in
comparison with a conventional stereo matching method based on image correlation
though computing time increased by 52'"'--92 percent.
INTRODUCTION
Though many stereo matching methods have been proposed, improvement of
accuracy and reliability still remains one of serious problems. With
conventional stereo matching method based on image correlation, it is impossible
to verify how likely recognized conjugate points are correct.
One of the solutions is multi-point matching method [IJ, where conjugate
points are determined to let weighted sum of image correlation and smoothness of
terrain shape determined from neighboring conjugate points be optimal. In multipoint matching, the apriori information on smoothness of terrain shape is
lected on the weight value and used to verify the correctness of conjugate
points.
approach is triplet matching. Position of ground point is
ned from three stereo rays. In case of SPOT image with three sight, that
is, right, center and left images, a conjugate point is likely correct, only if
the point
ch is determi
from a stereo
r of left and center images
coinci
th
point
another stereo
ch is
r of r
and center images.
Authors al
proposed triplet matching methods
with simul
linear array sensor data [2J.
applying triplet matching to SPOT images are
In
showed it's validity
is study, the result of
TRIPLET MATCHING ALGORITHMS
The following two tr
were proposed
let matching algorithms
on image correlation
compared with a conventional stereo matching method.
Independent Matching Method (Fig. 1)
Two pairs of stereo rays, that is, ray No.1 and No.2, ray No.1 and No,3 are
independentlY matched.
Discrepancy between two computed heights, Z12 and Z23
is checked as shown in Figure 1. If height discrepancy 1::,.Z is less than a
threshold value.
result of the triplet matching can be considered
reli
Ie. If the discrepancy exce
the threshold, the matching is repeated
until either the criteria is satisfied or iterration count exceeds a limit.
(2) Simultaneous Matching Method (Fig. 2)
Three rays are simultaneously matched under the condition that three rays
should be intersected at a point, as shown in Figure 2.
In both methods, shape of correlation windows is changed so as to cancell
difference of parallax within each window using neighboring conjugate points as
the matchi ng procedure proceeds from 'coarse' to 'f i ne ' . The appropr i ate size
of windows
threshold value for height discrepancy were determined from
several trials carried out in the test cases.
Ray NO.1
Ray No.2
Ray No.3
Ray NO. 1
Ray No.2
Ray No.3
\
\
\
Z23]
Independent Matching
for a Conjugate Point
~Z
Z12
Search Range of Simultaneous
Matching for a Conjugate Point
Figure 1 Concept of Triplet Matching
( Independent Matching)
Figure 2 Concept of Triplet Matching
( Simultaneous Matching)
TEST AREA
Test cases for triplet matching were carried out using three SPOT images of
a mountainous area around Mt.Fuji.
Figure 3 shows location of test area.
Figure 4 shows date and sensor angles of three SPOT images ( left, center and
right images).
Photo 1,2 and 3 are SPOT images of the test area.
TEST CASES
Height accuracy of digital terrain models obtained through stereo or
triplet matching was used for evaluating matching accuracy.
Cases shown in
Table 1 were examined for comparison of height accuracy and computing time.
The
threshold value for checking height discrepancy in the independent matching
method was 40 meters. Window size used in both stereo and triplet matching was
9 x 9 or 7 x 7, which was determined from several trials.
Figure 3
1
test case is shown in Table 2
ing time
and 3.
F
those
terrain
actual digi ta.l
errors between
5
exist some biases
stereo or tr
hei
ause the bases were
three
im;:tr<es
be more appropriate
to be caused
in test cases,
icator for assessing
i ng.
There
are almost the same in all
error the sizes of whi
cases.
let
orientation errors of
iation (STD) of errors would
ing accuracy.
2
( 17.
1
=
3
Ri
ter
( 7,
o.
( 8,
1
= O. 52
nter Ima.ge ( 7.
The best result of stereo matching (
results
tr
let matching.
) come very close to
e
r, rest
the results
stereo matching
are much
lor to
of triplet
ng in terms of
STD error
maximum error ( error range). It is clearly seen in Figure 5 that the tr
let
matching methods could reduce the number of gross errors. It could be concluded
ng me
showed tter accuracy and reliablity
that the tr let
a
ing me
conventional stereo
ting time
on image carre ation.
triplet matching method increase
percent ( respectively
ng
mu taneous
ing
requires more time than
is
until
discrepancy is satisf ed or iterration count
either the
exc
ing me
ing
simultaneous
52
a lit.
ION
tr
let
ing
accuracy
reli iIi
on image corre ation,
d
a conventi
require more
tive especi
were e
ly in
ing gross errors.
ing
s
lm,D.,:
ly more accurate
ing time,
ng
than the
ndent
ing
Iti
Iuation of
nsing voL
i, R.
Com. I I I Sympos i um,
ing
triplet
ing t
reliable
requires more time.
ing
int
ing
Technique
imensional
tric
neering
6.
reo
iemL
ing
, 1986.
Stereo
CCD Linear Sensors,
Table 1
Case 1.
es
t
Conventional Stereo
ratio
B/H ratio
BIH ratio
Case 1.
O. 20
(
O.
O.
(
nter and Right
( Ri
and
ft )
ing
Simultaneous
Tabl
iug
2
curacy
ing
e 3
me
termining a
nt
ing
1
1
Bin
o.
}
O.
1
1
c
1
2
,)
;.,
2
"I
1
nt
Simultaneous
ng
telling
1.
ec
1
Error Size
1
-200------191m "'----_--.
-190------181m
-180------171m
-170------161m
-160------151m
Frequency
3
10
10
2
10
= O. 72
-150~-141m
-140------131m
-130------121m
-120------111m
10------10 1m
-100----- -91m
-90--- -81m
-80~ -71m
-70----- -61m
-60----- -51m
-50----- -41m
-40----- -31m
-30----- -21m
-20----- -11m
-10----- -1m
0----- 9m
10----- 19m
20----- 29m
30----- 39m
40----- 49m
50----- 59m
60----- 69m
70----- 79m
80----- 89m
90----- 99m
100----- 109m
110----- 119m
120----- 129m
130--- 139m
140----- 149m
150--- 159m
160----- 169m
170----- 179m
B/H = o. 20
B/H = 0.52
/
~~~-~-:~-
~~ Independent Matching
Simultaneous Matching
..I.-----f---'
Stereo Matching
Independent Matching
Simultaneous Matching
Figure 5
quency of
1
Errors
Contour Interval;
(1)
500
Ac
6
Fi
o
ng (
o.
1000
(2)
Fi
reo Matching (
Image
o. 52
)
Measured
Contour Interval;
(3)
Stereo
o.
i ng (
2500
(4) Tr
(5)
let
iplet
ing ( Independent Matching)
iog (
Simultaneou~
ing )
Figure 7 Contour Image of Measured DTMs
( Contour Interval; 40m )