Automatic DEM generation by 
multi-image feature based matching 
Hans-Gerd Maas 
Institute of Geodesy and Photogrammetry 
ETH - Hoenggerberg, CH - 8093 Zurich, Switzerland 
Phone +41 - 1 - 633 3058, Fax +41 - 1 - 633 1101, e-mail gerd@geod.ethz.ch 
Commission III 
Keywords: digital terrain model, feature based matching, epipolar lines, multi-image 
ABSTRACT: 
One of the major difficulties in the automatic generation of digital elevation models (DEMs) is the procurement of 
approximate values. This presentation shows a new development based on the extraction of discrete points by an interest- 
operator and epipolar line intersection techniques in multiple overlapping images. The method can be considered an 
extension of well-known automatic DEM generation approaches from a stereo technique to a multi-image technique, and 
it solves the problem of provision of approximate values inherently. In contrast to most stereo-based techniques, the 
approach is not based on image pyramids, but on the consequent exploitation of the geometric strength of multiple 
images, implemented via the intersection of epipolar lines. Although not outlined for a complete DEM generation yet, the 
method may be very valuable for a hypothesis-free generation of good approximate values for other (e.g. area-based) 
DEM generation techniques or for the refinement of automatically generated DEMs degraded by smoothing effects. 
Applications of the technique to DEM generation in mountain regions based on scanned overlapping imagery of a regular 
block with 80%/60% overlap and on stillvideo imagery taken from a helicopter showed good results and blunder rates of 
0 - 0.5% without introducing assumptions on the maximum steepness of the terrain and without any post processing. 
INTRODUCTION 
The automatic generation of digital elevation models 
(DEM) has gained large attention among photogramme- 
trists in the past 10-15 years. A wide variety of 
approaches has been developed and presented in the liter- 
ature, and automatic DEM generation packages are mean- 
while commercially available in several digital 
photogrammetric workstations. A very rough classifica- 
tion of methods distinguishes between area based tech- 
niques computing correlations between greyvalue patches 
of images and feature based techniques establishing corre- 
spondences between extracted features of images. Tech- 
niques based on least-squares-matching (Gruen, 1985) are 
sometimes considered a combination of both. 
A critical point in all automatic DEM generation methods 
is - besides the recognition (and avoidance) of objects 
above the terrain surface like houses or trees - the procure- 
ment of approximate values in regions where no pre- 
knowledge on the terrain surface is given. Most methods 
employ image pyramid techniques or interactively 
selected starting points to solve this problem: 
* Image pyramid: in a coarse-to-fine approach DEMs 
are generated at successively finer resolution levels of 
the images, starting with a horizontal plane as approxi- 
484 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
mation for the highest level of the pyramid; this 
approach is e.g. implemented in the commercial 
DEM generation package MATCH-T (Krzystek, 1991) 
and was also used by (Baltsavias/Stallmann, 1992). 
* Interactive starting points: an operator sets one (or a 
few) starting points interactively, and the matching 
procedure moves into all directions from the starting 
point(s), assuming a certain maximum terrain slope. 
This approach is e.g. being used in the industrial 
surface measurement package INDUSURF (Schewe, 
1987). 
Both techniques work well in most cases, but may show 
problems in regions with steep surface gradients, occlu- 
sions or certain frequency patterns in the image texture. 
Especially in area based methods the steps in image 
pyramids are sometimes too big, so that approximate 
values cannot be considered sufficient for the consecu- 
tive pyramid level and points get lost. 
The method presented in this paper does not require 
either one of these techniques for generating approxi- 
mate values, but provides a solution which is indepen- 
dent on the availability of approximate values or pre- 
know 
rough 
maxi 
is re 
follox 
extra 
corre 
epipc 
long 
leadii 
ambi. 
whicl 
uses 
limiti 
pyrar 
of M 
prob] 
micly 
outli 
the t 
robus 
break 
dens 
point 
In th 
epipc 
due t 
by « 
imag 
by s 
base: 
The 
imag 
metr 
ters | 
an ej 
the c 
this | 
the « 
certa 
to d. 
for t 
beco 
on tl 
men 
the : 
ofte1 
area 
betw 
epip 
requ 
and 
ities 
able 
not |