Local features are grouped into three classes: 
e Interest points can produce target matching locations 
(Zong et al., 1991). 
e Edges can be matched (Schenk et al., 1991). Although 
their localization is weak, the matched edges are rel- 
atively reliable. Therefore they may render good ap- 
proximations for other methods with precise localiza- 
tion properties. 
e Texture values and texture based segmentation are 
useful for detecting shadows, water bodies, and the 
like. Matched segments provide good global matching 
constraints. 
In the context of our DOG project, matching procedure 
refers to the matching method in general. We use four meth- 
ods: 
€ Cross correlation 
e Least squares matching 
e V — 5 feature based matching 
e Symbolic matching 
Cross correlation and least squares matching methods are 
discussed in the next sections. The V — S feature-based 
maíching is a very reliable technique to obtain numerous 
matching points with fair localization accuracy. In our im- 
plementation a scalable LoG operator generates the edges. 
The edges are sorted and transformed into the V — S domain 
where a global matching takes place. Symbolic matching is 
very useful for global matching (Zilberstein, 1991). Due to 
implementation issues its use in our project is limited to the 
1-D case. We presently use it in profile matching. 
Matching strategy refers to constraints how the extremes are 
sought. Based on the original DOG objective — automati- 
cally adjust Z (find conjugate points, compute Z coordinate 
and drive the floating mark) — the conjugate location is con- 
fined to the vertical line which translates into two lines in 
the image planes. Since occlusions may block out certain 
segments of the constraint line, global methods are supe- 
rior to cope with this case. Although the terrain is mod- 
elled in least squares matching by the shape parameters, a 
significantly better approach is to use a priori surface ap- 
proximations (Schenk et al, 1990). Based on the surface 
data the patches are warped and in this format, basically 
free from terrain relief distortion, normal cross correlation 
is used. Since the surface data are quite sparse it is very 
critical how the surface interpolation algorithm performs. 
The most crucial component of our DOG project is the eval- 
uation and performance analysis of the results. The evalu- 
ation module serves as a system controller to implement a 
data driven algorithm. Typical operation tasks are: 
1. The two patches directly go to the statistical module 
(the preprocessing is skipped unless specified by the 
user). 
2. The results of the image patch analysis are compared 
to data obtained from the neighboring patches; if sig- 
nificant differences are found then the other three fea- 
iure extraction modules are activated, otherwise the 
402 
same sequence of module processing is executed used 
for the neighboring patches (it may also include other 
feature extractions) . 
3. After executing a matching function (a defined se- 
quence of module operations) the matching results are 
compared to the results of the neighboring patches and 
to predefined global parameter values. If everything is 
all right the new Z value is computed and the process 
terminates. 
4. If in (3) the results are not satisfactory then based 
on built in rules the matching sequence - in whole or 
part — is modified and a new computation starts; this 
is repeated until a satisfactory solution is found; if all 
strategies are exhausted, the process terminates with 
a message that matching is impossible. 
5. If in (2) the comparison fails, the system controller 
assigns the initial matching strategy according to the 
results of the three feature extraction modules, then 
point (4) is executed. 
6. Upon termination, the parameter set and the last pro- 
cessing sequence are saved for the next application of 
the DOG function. 
In summary, the system controller can be considered as a 
data driven, self-organizing, adaptive system which finds the 
optimal matching strategy to any data input from a list 
of prestored computation sequences. The performance of 
module computations is measured in the usual parameters, 
like the absolute and relative value of the correlation coeffi- 
cient in cross correlation, or residuals and variances in least 
squares matching, or the relative number of matched edges 
in ¥ — S feature based matching, etc. The system controller 
can be viewed as two tables, one containing matching se- 
quences and the other consisting of rules on how to evaluate 
the results. Tables are extended to include experimental re- 
sults, thus the system can learn. Under normal conditions 
the neighboring image patches are similar enough and the 
search process for optimal strategy is called only where there 
are significant scene changes in the images. 
2.3 Cross Correlation 
Cross correlation matching can be used in all three geo- 
metrical constraint strategies. Less important details, such 
as window size, which may be determined from statistical 
properties, are omitted, and the typical epipolar condition 
(correlation window becomes a line) is assumed in the fol- 
lowing description. 
Line Constrained Search 
The search line can be easily determined from the XY coor- 
dinates of the floating mark, and from the assumption that 
the current Z value is quite close to the real surface value. 
Thus the Z coordinate of the possible P surface point should 
be in the range: 
Zp 
d 
Zp 
Ze t dZ 
Z. — dZ 
where Z, is the current elevation of the floating mark and dZ 
is the search range defined by global constraints. The two 
extreme points, Z and zs are projected to both left and 
right 
form 
Asa 
searc 
cond 
ing t 
COITe 
The 
gives 
Usin, 
relial 
obtai 
In ge 
to in 
G 
In th 
tire p 
The : 
or dr 
neces 
tion. 
Su 
If tri 
the i 
remo 
with 
Surfa 
gener 
must 
al, | 
able : 
data 
goritl 
surfa« 
that i 
an im 
strate 
surfa« 
2.41 
In ge 
strain 
g»(a 
Eq. 1. 
The s 
the ac 
A pro 
projec 
Most 
paper 
contre 
The fi