sed 
her 
are 
ind 
g is 
ess 
sed 
Or 
his 
ith 
ler 
the 
jen 
ro- 
of 
5 a 
he 
ist 
zes 
ie 
re- 
Ins 
he 
re 
0- 
ch 
cal 
on 
ol- 
)r- 
at 
le. 
ld 
A 
NO 
nd 
right photo planes by collinearity equations and then trans- 
formed into image coordinates, yielding the two search lines. 
As a next step, the correlation windows are moved along the 
search lines and the coefficients are computed. Under ideal 
conditions there is only one pair of conjugate points point- 
ing to the desired surface location, and for this point the 
correlation coefficient is likely to have a maximum value. 
The analysis of the correlation curve around the maximum 
gives some indication about the reliability of the maximum. 
Using a band of parallel search lines can further improve the 
reliability of the results. In this case a correlation ridge is 
obtained, and its analysis can lead to more reliable results. 
In general, independent global matching should be applied 
to increase the confidence level of the matching results. 
Global Search 
In this case the correlation window is moved within the en- 
tire patch area, and a 2-D correlation function is computed. 
The shape of the correlation function may help to confirm 
or drop our hypothesis about a location, although it is not 
necessarily feasible to determine directly the desired loca- 
tion. 
Surface Warping 
If true surface data are available, then the distortion of 
the image patches caused by terrain relief can be totally 
removed. In this case, cross correlation is concerned only 
with texture information, and reliable results are obtained. 
Surface data may be known for the previous patches, but in 
general the true surface is never available, and therefore it 
must be approximated. Based on our experience (Schenk et 
al., 1990), with ¥ — S feature based matching, enough reli- 
able surface points are obtained. Nevertheless, the surface 
data are still quite sparse and a surface interpolation al- 
gorithm must be used. Through hierarchical iterations the 
surface approximation usually improves. It is appreciated 
that in this process the interpolation algorithm itself plays 
an important role (Al-Tahir, 1992). A bad approximation 
strategy can slow down the convergence or even make the 
surface diverge. 
2.4 Least Squares Matching 
In general, least squares matching can be similarly con- 
strained as cross correlation. The general approach is: 
g2(z,y) = ho + hy - g1(ao + a1 + a27, bo + bay + box) (1) 
Eq. 1. can be simplified if epipolar geometry exits: 
92(z) = ho + hy - g1(ao + a,2) (2) 
The surface data can be used to set better initial values for 
the adjustment procedure. 
3. EXPERIMENTS, RESULTS 
A prototype version of the proposed solution in the DOG 
project has been implemented on Intergraph workstations. 
Most of the modules are operational. As of writing this 
paper extensive tests have been performed, and the system 
controller tables have been built. 
The first observations are: 
e the major modules perform as expected 
403 
e the basic cross correlation matching works well for rea- 
sonable test data (for example, with fixed X and Y 
increments it automatically collects DEM grid points) 
e the automatic parameter tuning is difficult and needs 
good initial values 
the deterministic approach of the system controller is 
not optimal 
e it is quite complex to define the rules 
In summary, the preliminary test results are encouraging. 
Theoretical investigations are needed to analyze the results 
and to parametrization the confidence level. On the imple- 
mentation side, the growing number of rules and module 
sequences justifies the use of an off-the-shelf expert system 
(Schenk and Toth, 1991). The accuracy tests will include 
large numbers of varying image data, and more indepen- 
dent operators are needed to provide the ground truth for 
performance evaluation. 
4. REFERENCES 
Ackermann, F., 1984. “Digital Image Correlation: Perfor- 
mance and Potential Application in Photogrammetry,” Pho- 
togrammetric Record, Vol. 11, No. 64, pp. 429-439. 
Al-Tahir, R., 1992. “On the Interpolation Problem of Au- 
tomated Surface Reconstruction,” International Archives of 
Photogrammetry and Remote Sensing, Vol. XXIX. 
Helava, U.V., 1988. “Object-Space Least-Squares Correla- 
tion,” Photogrammetric Engineering and Remote Sensing, 
Vol. 54, No. 6, pp. 711-714. 
Kaiser, R., 1991. “ImageStation: Intergraph’s Digital Pho- 
togrammetric Workstation,” Digital Photogrammetric Sys- 
tems, Wichmann, pp. 188-197. 
Schenk, T., Li, J-C., and Toth, Ch., 1990. “Hierarchical 
Approach to Reconstruct Surfaces by Using Iteratively Rec- 
tified Images,” Proc. Int. Soc. of Photogr. and Remote 
Sensing ISPRS, Symp. Comm. V, vol. 28, part 5/1, pp. 
464—470. 
Schenk, T., Li, J.C., and Toth, Ch., 1991. “Towards an Au- 
tonomous System for Orienting Digital Stereopairs," Pho- 
togrammetric Engineering and Remote Sensing, vol. 57, no. 
8, pp.1057-1064. 
Schenk, T., and Toth, Ch., 1991. "Knowledge-Based Sys- 
tems for Digital Photogrammetric Workstations," Digital 
Photogrammetric Systems, Wichmann, pp. 123-134. 
Schenk, T., and Toth, Ch., 1992. "Conceptual Issues of 
Softcopy Photogrammetric Workstations," Photogrammet- 
ric Engineering and Remote Sensing, Vol. 58, No. 1, Dp. 
101-110. 
Zong, J., Schenk, T., and Li, J-C., 1991. "Application of 
Forstner Interest Operator in Automatic Orientation Sys- 
tem," Proc. ASPRS-ACSM Annual Convention, Vol.5, pp. 
440-448. 
Zilberstein, O., 1991. "Solving the Correspondence Problem 
in Aerial Imagery Using Relational Matching," Phd disser- 
tation, Dept. of Geodetic Science and Surveying, The Ohio 
State University, Columbus, OH.