from the image, including the 1-D case of scan lines,
(segmented) edges, or specific geometric objects.
(2) The control strategy that specifies how to find a po-
tential match.
(3) The criteria for determining (selecting) the best match
from several candidates. The matching criteria are
measures of similarity between different features.
2.1 Area-Based Matching
In area-based matching (ABM) a rectangular area (templet)
of one image is compared with an area of the same size in
the other image. Fischler [13] shows that if images differ on-
ly due to horizontal and vertical displacement then *unnor-
malized’ cross-correlation is the optimal matching method.
Since ABM techniques use image patches they are sensitive
to perspective distortion (relief distortion), to changes in il-
lumination and contrast, and to occlusions and shadows. Of
all the positions compared, the one that renders the best
similarity measure between the templet and the search win-
dow is chosen as the match position. The similarity crite-
rion can be checked by either searching for the maximum
cross-correlation coefficient, or by minimizing the gray level
differences using least-squares adjustment (LSM).
Area-based matching (ABM) schemes offer these advan-
tages:
e Flexible mathematical model: LSM is the method of
choice in photogrammetry, because it provides a gener-
al approach to area correlation by offering a tractable
mathematical model (least square adjustments). It is
easy to use multiple images whereby all image patch-
es are matched simultaneously. It enables photogram-
metrists to apply familiar mathematical and statistical
principles [19]
Simple matching algorithm: Both, cross-correlation a-
nd LSM are considered simple algorithms with well-
known procedures for fast implementations.
e Small storage resources: Only the templet and the sear-
ch window need be kept in memory resulting in very
small memory requirements.
High accuracy: The accuracy of matched points is high.
Ackermann reports in [1] accuracies of points (geomet-
ric targets, fiducial marks) with a standard deviation
of 3.7 um.
Area-based matching methods suffer from the following
problems:
e Break lines: It is assumed that the template and the
search window cover a smooth surface area. If this
assumption does not hold, for example when a break-
line crosses the surface patch, then the matching re-
sults may be wrong. Breaklines possess rich informa-
tion about the surface. Unfortunately, ABM performs
poorly on these interesting areas.
712
« Matching *meaningless' points: ABM methods match
pixels on the basis of gray levels differences. Pixels
have no explicit information about interesting areas of
the object space. Therefore, the matching results (3-D
position of points in object space) are on the same low
level of abstraction as the original image and have no
meaning associated. It may be that totally uninterest-
ing areas are matched with a very high accuracy.
e Photometric differences: ABM methods have difficul-
ties with images of different radiometric properties.
The radiometric differences between images may result
from using different cameras, images from different e-
pochs, or from different reflections of bright objects
such as water bodies, etc.. Rosenholm concludes in
[32] that the radiometric quality of the images is crit-
ical for gray level matching with the LSM method.
Geometric differences: One of the basic assumptions
of image matching techniques is that the two windows
(templet and reference window) cover the same area
in the object space. This is only the case if the surface
is parallel to the camera base. In real situations the
two windows cover different areas, hence different gray
levels, which affects the matching results (see Horn [22]
for more details).
* Problematic texture: In areas such as grass, or in areas
with repetitive patterns, there is a problem to deter-
mine the position of the best match (e.g flat correlation
surface).
2.2 Feature-Based Matching
In feature-based matching (FBM) selected features of each
image are first determined on the basis of distinctive im-
age values. The features so determined may include points
(feature points), corners (intersection of feature lines), and
edges. After the location of features is determined a relation-
ship between conjugate features is established (matching).
This process is usually performed on the basis of similari-
ty of the feature attributes, for example shape, orientation,
gradient, etc..
Some of the advantages of feature-based matching (FB-
M) are summarized below:
« High reliability: Generally, FBM produces more reli-
able results than ABM because of the distinctive prop-
erties of features. Also, features (particularly edges)
are derived over a large spatial extent and thus add to
the robustness.
* Captures important information: Feature possess mo-
re explicit information about the object space than
the raw gray levels. Matching zero-crossings, for ex-
ample, renders the 3-D location of potential object
boundaries. This stems from the relationship between
discontinuities in the surface and gray level disconti-
nuities (edges). Matched edges are an essential step
toward image understanding and object recognition.
