describe some illustrative systems for some specific tasks and
discuss how these choice were made.
2. ISSUES IN MATCHING
Several important issues need to be considered in making the
choice of a matching strategy for a particular task. These
include the choice of representation level for matching,
whether the matching is done in 2-D or 3-D, whether it is
local or global and the method of matching itself. These are
discussed below.
2.1 Representation Level:
Perhaps the most important consideration in matching is to
determine the level (or levels) of representation at which
matching is to be performed. Some possible levels are:
i) Direct pixel intensity (or color) values
ii) Point features such as local variance or edges
iii) Grouped features such as curves, sets of curves or regions
iv) High level features such as surfaces and volumes
Clearly, the features to be matched must be computable from
the input data, must be invariant or quasi-invariant, and must
be in the same form as the model or forms that can be derived
from the model. The appropriateness of the level will depend
on factors outlined in the introduction earlier, however, we
can make some general observations about the choice:
i) Higher levels of representation require more complex
algorithms: e.g. for intensity matching, simple correlation
may be used but matching of surfaces may require graph
matching procedures. Computational requirements at
higher levels may be less due to the lower number of items
to be matched; however, this may be compensated by the
computational requirements of obtaining the higher level
representations.
ii) Higher levels of representation are more distinct and thus
the matching is likely to give less ambiguous results.
However, more ambiguities may be present in the process
of constructing the higher levels from the given image data
in the first place.
2.2 2-D vs. 3-D Matching:
In general, matching in 3-D is more constrained and the 3-D
features are more distinctive. Thus, if the input data is in a 3-
D form, there are obvious advantages to performing the
matching in 3-D. However, 3-D is not explicitly available in
intensity images, in fact, extraction of 3-D may be an explicit
goal of the matching process.
2.3 Local vs. Global Matching:
Another issue to consider is the extent over which matching
568
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
procedure should be applied at a time. Local matching can be
very precise, but ambiguous and the various local matches
may not be consistent with each other. Global matching is
more robust but not necessarily accurate in local areas unless
a single transformation relates the sources to be matched. In
general, we may need to make a compromise between the two
extremes or proceed from local to global matches (or vice-
versa) in stages.
2.4 Method of Matching:
Several techniques of matching are available, corresponding
to the kinds of representations used. Three kinds of methods
are listed below:
i) Area Correlation: here a single measure of match is
computed by applying a transformation and matching the
similarity between two sources.
ii) Feature Matching: This is a modification of the area
correlation method. À certain transformation is applied to one
source and the number (or amount) of matching features is
computed. Determination of what constitutes a match is now
more complex (i.e. when can two line segments be considered
to match and to what degree).
iii) Structure Matching: Here a group of features is
matched together, by considering not only individual feature
properties but also some explicit relations between them
(such as certain kinds of alignments, say parallelism). In
general, these methods employ graph matching techniques.
A good survey of these issues and approaches can be found
in ([6], [10]).
We now consider two specific kinds of tasks to illustrate the
specific issues in matching and some kinds of techniques that
have been used. First task is that of scene registration where
one (or more) of newly acquired images need to be matched
with a model (or map) of the site (constructed from earlier
images or other sources); this task is important for purposes
of change detection and model (map) updating. Second task
is that of matching two or more images for the purpose of
extracting 3-D models (maps). We will focus on cases where
the input data consists of panchromatic intensity images and
the scenes contain significant amount of man-made features
rather than just natural terrain.
3. SCENE REGISTRATION
In this task, a new image needs to be registered with maps or
models constructed from previous images. This operation is
needed for several tasks such as detecting changes in the
scene from the last time the models were constructed. Change
detection is important for many civilian tasks such as map
updating, urban monitoring and earth resource surveys and
also for military tasks of observing significant infra-structure
changes.
The complexity of this task varies greatly with the kind of
given ii
provide:
will con
terrain r
with go
matchin
characte
model |
transfori
distortio
that of «
general,
camera
paramet
range.
3.1 Ma
In this c
heights |
stereo m
Let us cc
image (i
We can
intensity
rather |
transforr
two-dimt
be condi
methods.
by some
intensity
squares |
One arra
metric cc
the best
efficient |
coarse re,
search at
the lower
achieved
approach
This tech
range im:
and inten
and the in
paramete:
aware of
that some
below wi
3.2 Mat
orientati
We now «
with a 3-]
Thesiten