High Precision Image Matching 
or 
Digital Terrain Model Generation 
Armin W. Gruen, Emmanuel P. Baltsavias 
Institute of Geodesy and Photogrammetry 
Federal Institute of Technology 
ETH-Hoenggerberg 
CH - 8093 Zurich, Switzerland 
ABSTRACT 
The Adaptive Least Squares Correlation (ALSC), combining gray level correlation with geometric 
conditions, was applied to a grid sampling mode for generation of Digital Terrain Models (DTM) 
from aerial stereo models. 
The heights of targets of different geometric and contrast quality in large scale digitized images 
were determined using this technique and compared to results from manual measurements. 
Both solutions showed an average difference of 0.016-0.038% and a maximum difference of 
0.037-0.120%. flying height (hg), depending on the height approximation. Most cases 
investigated required an initial height approximation at the 1-275, hg level. An average height 
error of 0.2%. hg was cleared in each iteration step. The constraints imposed proved to define 
the movement of the gray level image patches, thus limiting the search area and probability of 
false correlation, and increasing the precision and reliability of height determination. 
1. INTRODUCTION 
One of the earliest attempts for automatic contour plotting is described in U.S. patent No. 
2,283,226 issued to H.B. Porter in 1942. Since then, numerous instruments and techniques 
have been developed for the automation of DTM generation. They can be divided into 
analogue, hybrid and digital techniques, with the recent tendency towards digital solutions, 
using specialized hardware in a parallel, distributed processing environment. Most of them made 
use of the epipolar plane condition and used crosscorrelation as matching criterion. 
Today the commercially available hardware and software is beyond the increasing needs for 
generation of DTMs and has not reached a stage of maturity yet. They can handle primarily 
medium and small scale images, lack flexibility and adaptability to the various terrain surfaces, 
make no or little use of additional information which could be favourably exploited, and have only 
little interactive operation capability. The failure rate is still high, control of the precision and 
reliability of the results is limited, discontinuities of the terrain and geometric and radiometric 
dissimilarities due to other factors remain a bottleneck. 
This paper addresses the precision matching aspect for automatic DTM generation. 
ALSC as a general and flexible image matching technique combines gray level correlation with 
object point intersection conditions and other sensor and object constraints. The system 
parameters allow for local geometric image shaping and radiometric corrections and are 
automatically corrected and optimized during the least squares iterations. The quality of the 
match and other system parameters can be assessed. The geometric conditions imposed lead 
not only to an improved matching procedure, but in addition provide for a simultaneous 
matching/object point positioning technique. Thus the conventional two Stage process is 
replaced by a one stage solution, which is capable of utilizing all available radiometric and 
geometric information at once. 
The general mathematical model used is described in Gruen1985, Gruen,Baltsavias1985. This 
algorithm is applied to a grid sampling mode for generation of DTMs from aerial stereo models, 
requiring determination of the heights at given X,Y ground positions.