ares correlation,"
pages 321-331,
ng with Simulta-
lines. PhD thesis,
Royal Institute of
ng," Photogram-
r, (8):1039-1048,
arity using regu-
, 1988.
rebra," Proceed-
Ill, International
eral base of fast
sing Symposium,
hotogrammetry.
| signal process-
1d Remote Sens-
ray algebra tech-
1.11] Symposium,
i, Finland, 1986.
system: An array
ammetry,” Pho-
te Sensing, (3),
cer, "Automated
ing algorithm of
rrammetric Engi-
J.
nent of digital
errain models,"
pages 573-587,
using the least
hree-dimensional
ng and Remote
g along vertical
urnal of Remote
erse visual prob-
AMI, 8:413-424,
thmen fur das
| Report 374,
inich, 1991.
vision) - a new
zital photogram-
n Conference of
'ammetric Data,
with Neural Net-
Jniversity of Sur-
1990. in chinese.
Automatic Point Transfer: a practical application
of optimal digital image matching based on
local invariant properties
X. Xu, ZG. Tan*
Graz City Council, Dept. 10/6
Commission III, WG III/2
Key Words: invariant properties, local descriptor, optimal matching, automatic point transfer
Abstract
One practical example in digital photogrammetric production is that if one point in the new image is chosen, the same position in old
digital image should be found and its geodesic coordinates be transferred to the new one automatically. Another well known
example is to automatic produce digital terrain model from stereo pair. One of the key steps of automatic point transfer or automatic
DTM production from aerial photos is introducing a robust matching algorithm. In our case, robust means even in tatteried area it
should give out satisfied results. In this paper a feature-based matching method which depends on locally invariant properties. is
presented. The characteristics of this algorithm are the first, the matching inputs are the radiometric and geometric noise invariant
properties of image patches; the second, the locally matching inputs are optimally used for extrinsic matching procedure. Some
practical results are presented.
1. INTRODUCTION
1.1 The goals
Generally after every aerial photography it is necessary to
follow the aerotriangulation procedure so that the new photos
can be used in the production. Our experience told us the
aerotriangulation was a time- energy consuming job, even with
optimal working procedures and excellent adjustment program
/Ganster1993/, /Ganster1994/. For a city like Graz, usually in
every 3 ~ 4 years one new aerial photographic flight is to be
carried out. So in the point of view of economic aspects, it will
be very interesting if the aerotriangulation can be removed from
the production chain. Now slowly digital image matching
techniques show their practical marvelous sides. If one position
in new image is defined by operator, the same position in old
image(s) which were taken several years ago and absolute
orientation parameters are known should be found and the
geodesic coordinates can be transferred to the new one. Let's
name this operation as automatic point transfer. On the base of
one high accurate aerotriangulation results of the old image sets
and available digital photos (the both sets of digital photos will
be used for Orthophoto production), automatic point transfer is
exact the right tool for this purpose.
It is true that in some areas under special conditions the DTM
can be driven from digital stereo pair by computation. But in
the typical european city Graz, it is almost unimaginable at now
to get DTM (urgent demand for orthophoto production) full
automatically. What we try now is that let feature-based
matching results (or combined with other matching methods)
make the measurement of DTM much more easier for human
operators, especially in the tattered areas like quarry field or
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
down town. The direct application is to use the matching results
as prepositions for DTM measurement.
1.2 Overview of different methods
Before this matching method was developed, lots of existed
methods were studied carefully /Xu1994/. Here let's have a
simple overview. Due to the complexity and severity of
different matching tasks, a great mounts of matching algorithm
have been developed during last several decades. Each kind of
them deals with special aspect of the problems. They can be
generally sorted into following four types:
1. signal based matching /Hannahl988/, /Dowman1977/,
/Ackermann1983/, /Benard1986/; 2.) low-level feature based
matching /Moravec1977/, /Foerstner1987/; 3.) high-level
feature based matching /Shapiro1980/, /Cheng1985/; 4.) hybrid
matching /Jordan1991/, /Hsieh1992/, /Xu1993/
Signal based matching methods (They are also named as area-
based matching) can produce results with very high accuracy on
the basis of rather precise initial values and on the cost of
computation time. Low-level feature based matching is
sometime used as first step of whole matching procedure to
supply the initial values for signal based matching. For this aim,’
the methods should be qualitative robust, i.e. matching results
with the same quality should be returned even with different
image contents. In application of aerial photos in european city
areas there are few methods which can produce the satisfied
initial values for signal based matching, e.g. least square
matching. High-level feature based matching has the difference
with low-level one in the way that it takes the relations between
ZG. Tan* : Institute for Applied Mathematics, Tsinghua
University, PR. China
983
