MULTIPLE IMAGE MATCHING
Peggy Agouris
Toni Schenk
Department of Geodetic Science and Surveying
The Ohio State University, Columbus, Ohio 43210-1247
USA
Commission III
ABSTRACT
Digital photogrammetry is concerned with the development of algorithms to automate photogrammetric tasks.
The majority of efforts though are focused on single stereopairs. This paper addresses the task of simultaneously
matching conjugate windows from multiple overlapping images. After establishing a theoretical understanding
of the problem, we introduce several approaches and present the associated mathematical principles. We report
on the advantages and disadvantages of each one, discuss various implementation issues and in conclusion, we
examine potential applications in photogrammetric procedure.
1. INTRODUCTION
Digital photogrammetry has recently emerged as one of the
most promising and multi-faceted photogrammetric subfields.
A solid body of research work and a wide array of topics have
laid the foundation for the evolution of the photogrammetric
procedure. Among the research topics, automatic matching
is one of the most challenging.
Digital image matching attempts to identify sets of conju-
gate entities from two or more overlapping images. From
the diverse set of matching techniques [Lemmens, 1988],
least squares matching is a popular choice [Ackerman, 1984].
Even though there already exists substantial work on this
subject, most efforts have been focused on the stereomatch-
ing case, which involves a single pair of images. This paper
deals with simultaneously matching windows from multiple
overlapping images using least squares techniques. The sig-
nificance of this issue lies in the impracticality of handling
single models at the time when processing large blocks is
common practice in the photogrammetric industry. Success-
ful and efficient completion of multiple image matching is
expected to contribute significantly in the transition of digi-
tal photogrammetry from an experimental to a production-
oriented status.
Significant research in the area of multiple image matching
can be found in [Grün & Baltsavias, 1988],[Heipke, 1992]
and [Helava, 1988]. In this paper, we present alternative ap-
proaches to the subject by introducing geometric constraints
and performing matching in the object space. The general
least squares matching procedure is discussed in detail and
is subsequently expanded to accommodate multiple image
windows. We explore the theoretical issues of the proposed
approaches and establish the corresponding mathematical
principles. Then, we report on their advantages and disad-
802
vantages from a photogrammetric point of view and address
several implementation issues.
2. LEAST SQUARES MATCHING
Least squares matching techniques attempt to match win-
dows of pixels by establishing a correspondence between
them which minimizes the differences of their gray values.
Assuming gr(zr, yr) to be a window of n, x n; pixels in the
left image, and g%(x%,Y%) an equal size approximation to its
conjugate position in the right image, the objective is to es-
timate a new location of the right image window gn(*n, yn)
such that the gray value differences
gi (2r, VL) = gn(2n, yn) — e(z, y) (1)
are minimized. The estimation is performed by the trans-
formation of the coordinates (2%,y%) and resampling of the
corresponding gray values. The coordinates of the two win-
dows are related through a perspective transformation to
a common surface patch in the object space. Taking into
account the very small size of the windows to be matched,
their coordinates are assumed to be related to each other by
a 6-parameter affine transformation
TR = a1 + az + azyg, (2)
and
Yr = by + baxz + bayz (3)
With linearization, the equations
gi(21, yr) — e(z,y) = gR(=R, YR) (4)
become
gi(n, yr) — e(z,y) — gn(z5, yn) - gn.dvn -- gn,dyn (5)
f