Herbert Jahn
STEREO MATCHING FOR PUSHBROOM STEREO CAMERAS
Herbert JAHN
DLR
Institute of Space Sensor Technology and Planetary Exploration
Herbert.Jahn G dlr.de
Working Group III/2
KEY WORDS: Photogrammetry, Stereopsis, CCD Cameras, Image Matching, Parallel Processing.
ABSTRACT
A parallel stereo matching algorithm is presented which is mainly thought for the processing of images from pushbroom
stereo cameras. The algorithm is designed for non-epipolar geometry, because of disturbances of flight attitude and
velocity. Existing epipolar algorithms can give a first estimation of disparities in epipolar (x-) direction, but the
recursive algorithm can also start with zero-disparity initial condition if the disparities are not too big. The algorithm
minimizes locally a certain least squares distance of a stereo image pair using the method of steepest descent leading to
a recursive disparity updating. To diminish ambiguities a pyramid with Gaussian image smoothing together with other
measures (e.g. exploiting the ordering constraint and applying edge preserving disparity smoothing) is used. The
presented matching algorithm is parallel in space and sequential in time. Therefore, when suitable parallel processing
hardware (with one processing element assigned to each pixel) will be available then real-time stereo processing
becomes possible. Some examples demonstrate the capabilities of the algorithm but also the remaining difficulties.
KURZFASSUNG
Ein paralleler Algorithmus zur Stereo-Bildzuordnung, der hauptsüchlich für die Verarbeitung von Bilddaten von
Pushbroom-Zeilenkameras gedacht ist, wird prásentiert. Der Algorithmus wurde wegen vorkommender Stórungen der
Fluglage und —-geschwindigkeit für nicht-epipolare Geometrie konzipiert. Vorhandene epipolare Verfahren kónnen für
eine erste Schátzung der Parallaxen in epipolarer (x- ) Richtung verwendet werden, aber der rekursive Algorithmus
erbeitet auch ohne derartige Schátzwerte, wenn die Parallaxen nicht zu grof sind. Der Algorithmus minimiert lokal
einen gewissen Abstand eines Stereo-Bildpaares durch Anwendung der Methode des steilsten Abstiegs. Dies führt zu
einem rekursiven updating der Parallaxen. Zur Verminderung von Mehrdeutigkeiten wird eine GauBsche Pyramide
zusammen mit anderen MaBnahmen (Reihenfolgebeschrinkung — ordering constraint, kantenerhaltende Glittung von
Parallaxen) verwendet. Der Algorithmus ist räumlich parallel und zeitlich sequentiell und kann daher, wenn geeignete
Parallelverarbeitungs-Hardware (mit einem Prozessorelement pro Pixel) verfügbar ist, Echtzeit-Stereoverarbeitung
gewährleisten. Einige Beispiele zeigen die Fähigkeiten des Verfahrens und seine Mängel.
1 INTRODUCTION
The successful experiments with the digital stereo cameras HRSC (High Resolution Stereo Camera), WAOSS (Wide
Angle Optoelectronic Stereo Scanner), WAAC (Wide Angle Airborne Camera) and with a prototype of the first
commercial digital aerial camera ADC (Airborne Digital Camera) performed at the Institute of Space Sensor Systems
and Planetary Exploration of the German Aerospace Center (DLR) have shown that high quality stereo reconstruction
with pushbroom cameras is possible.
To generate cost efficient and high quality 3D data products the key problem continues to be the matching of two or
more image stripes. One needs a very fast matching algorithm in order to process efficiently the huge data amounts
generated.
Because of aircraft attitude and velocity variations the image geometry is not strictly epipolar. Therefore, available
efficient epipolar algorithms, e. g. the algorithm of Gimel’farb (1999) using dynamic programming techniques, can be
used only as a first approximation. For refinement a very fast and precise non-epipolar algorithm is needed.
436 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.