In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C., Tournaire O. (Eds). IAPRS. Vol. XXXVIII. Part 3A - Saint-Mandé, France. September 1-3. 2010
ROBUST MATCHING OF AERIAL IMAGES WITH LOW OVERLAP
M. Mizotin a , G. Krivovyaz a , A. Velizhev a , A. Chemyavskiy b , A. Sechin c
a Moscow State University, Dept, of Computational Mathematics and Cybernetics, Russia - {mmizotin, gkrivovyaz,
avelizhev} @graphics.cs.msu.ru
b State Research Institute of Aviation Systems (FGUP GosNIIAS), Moscow, Russia - achem@gosniias.ru
c Racurs Co., Moscow, Russia - sechin@racurs.ru
Commission III, Working Group III/l
KEY WORDS: Image Matching, Image Registration, Aerial Triangulation, Low Overlap
ABSTRACT:
This paper addresses the problem of aerial image matching. We analyze existing approaches to this problem and show that, though
the modem algorithms cope with the task quite well, their results are deteriorated in case of low overlap and significant rotation
angle between the images. A two-stage feature-based image matching scheme is presented. It is shown that preliminary stage of
simplified (shift-rotation) model estimation is crucial in case of low overlap and influences significantly the whole matching scheme.
We introduce a novel method of shift-rotation model estimation, based on the voting procedure in parameter space, which allows
finding the correct model in case of extremely difficult input data (overlap area is less than 10%) without using any additional
information. Finally, the experimental results of our method on both synthetic and real data are presented. We compare our results
with one of the state of the art SAC-based model estimators and show that our algorithm outperforms existing methods in case of
very small overlap.
1. INTRODUCTION
Automatic aerial image matching, or aerial triangulation, is the
field of active research at the moment. Matching is necessary
for further creation of orthophotomaps and 3D modelling, both
finding an application in such areas as geodesy, cartography,
Earth monitoring and others. With the rapid development of
geoinformation systems and the growth of input data amount, a
fully automated technology of aerial image processing is
becoming the main goal in this area.
A lot of different techniques have been introduced recently and
commercial software is already available on the market. But the
existing state of the art algorithms still are not able to ensure the
stable work of automated systems. Because of the complexity of
aerial images obtaining process, some difficulties inevitably
appear in the input data. Among them are the cases of very
small overlap area and significant rotation angle that usually
occur when matching a whole block of images divided into
sequences. The aim of our work was to develop a technique that
outperforms existing matching methods in these cases. Note that
image scale is assumed to be constant as it is a typical situation
for aerial images. See Figure 1 for an example of input data.
The structure of the paper is the following. In Section 2 we
discuss existing approaches to aerial images matching. The two-
stage feature-based matching scheme is given in Section 3 and
the novel voting-based method of shift-rotation model
estimation is introduced in Section 4. We present our
experimental results in Section 5. Section 6 concludes the
paper.
2. RELATED WORK
Aerial triangulation operates with hundreds or even thousands
of images but actually this complex procedure is based on
image pair matching. Aerial images are usually very large
(hundreds of mega pixels) making it necessary to use
hierarchical scale pyramid.
Figure 1. An example of input aerial images. The overlap area
(5.7%) detected by the proposed method is outlined