A COARSE-TO-FINE CORRELATION ALGORITHM CONSIDERING OCCLUSIONS
Symposium of Commission III
Rovaniemi, Finland
August 19-22, 1986
Susumu Hattori
Institute of Industrial Science, Univ. of Tokyo, Japan
Chuji Mori
Dep. of Civil Engineering, School of Engineering,
Okayama University, Japan
Osamu Uchida
Asia Air Survey Co. Ltd., Japan
Abstract: This paper covers a stereo matching algorithm suitable to
small or middle scale aerial images. Conventional matching algorithm
based on simple area correlation are often confronted with the
difficulties: 1) correlation deterioration due to terrain relief 2)
matching wanderings due to poor textures and 3) matching wanderings in
occluding areas. The authors propose that for resolving these
difficulties the so-called coarse-to-fine correlation method should be
employed. A sophisticated algorithm for stable and precise matching is
proposed based on coarse-to-fine correlation in the paper. It is featured
by 1) median-filtering of x-parallaxes to find occlusions and to take off
wanderings, and 2) narrow-band-pass-filtering of images to get high
matching precision through coarse-to-fine correlation.
1. Introduction
Area correlation has been widely investigated as a practical tool for
stereo matching of aerial images. But conventional techniques based on
simple correlation often fail to produce sufficient matching precision
and stability. The major reasons for it are as follows:
1) If correlation windows are large compared with spatial frequencies of
the terrain relief, correlation is suffered from perspective distortions
and true peaks are often hidden.
2) If correlation windows are so small as not to contain sufficient
textures, matchings tend to wander.
3) If occlusions are contained in images, matchings tend to wander.
Occlusions means areas that are revealed in one image but not revealed in
the other image because views are obstructed by some objects.
As to the difficulties 1) and 2), some authors have already tried to
solve. Panton et al./6/ suggested a method of adaptive windows: i.e.
search windows are shaped to be adaptive to corresponding correlation
windows according to terrain slopes that are predicted from previously
matched points. Furthermore window sizes are varied according to the
amounts of textures around the points to be matched. Pertle/8/
exploited a matching method of expressing the perspective distortions
and densitometric distortions in the windows in a linear form,
coefficients of which are estimated using the iterative least squares
method to decide conjugate points. Their methods, though very
attractive, seem to be still insufficient in matching stability.
On the other hand the difficulty 3) is inherent in area correlation.
Since area correlation is a method of comparing likelihood between two
finite areas, it is substantially hard to find occluding areas
(especially their boundaries) with high precision. Probably on this
account few investigations have been tried up to now to solve the
occlusion problem in the photogrammetric field. The authors/1,2,5/
asserts that for solving these difficulties so-called coarse-to-fine
correlation is very useful. Coarse-to-fine correlation is a multi-step
procedure from coarse matching in a low frequency domain to fine matching
in a high frequency domain. In this paper a sophisticated matching
algorithm for stereo aerial images is discussed with a particular stress
on how to find occlusions. Briefly to say, occluding areas and poor-
textured areas can be found step-by-step in the course of a multi-step
- 317 -