OPTIMISATION OF BUNDLE ADJUSTMENTS FOR STEREO PHOTOGRAPHY
Bruce King
Department of Civil Engineering and Surveying
University of Newcastle, Australia
Commission 5
ABSTRACT
Conventional bundle adjustments ignore the invariant geometric relationships that exist between camera pairs in a bundle of
stereo photography. Two models for optimising conventional bundle adjustments to take advantage of these relationships
are developed. These models are compared with a conventional bundle adjustment. Initial results indicate that both
presented models yield improved accuracies when compared to a conventional bundle adjustment.
KEY WORDS: Optimisation, Bundle adjustment, Stereo photography.
INTRODUCTION
In the field of non-topographic photogrammetry bundle
adjustments are used in a wide range of applications.
Whether implemented as a DLT or self-calibrating model
the purpose of the bundle adjustment is to minimise the
residuals of all the observations. Each camera thus finds a
position and orientation that reflects this minimisation. This
approach is well established for single camera imaging
geometry as typified by Fraser (1991).
When the imaging is performed by a stereo camera
system, inherent in each pair of photographs is the
invariant geometrical relationship of the two cameras. f
such a bundle of stereopairs were reduced by a
conventional bundle adjustment the invariant camera
relationships between the two images of each stereopair
would be ignored in favour of optimising each camera's
position and orientation based on the observations and
their random errors.
This paper reports on two methods that have been
developed to optimise a conventional bundle adjustment
for use with stereo photography so that the invariant
relationships are retained for all stereopairs. The
mathematical models are developed and the results of a
trial comparing the two models with a conventional bundle
adjustment are presented.
THE BUNDLE ADJUSTMENT
Slama (1980, Ch2) gives the generally accepted model of
the observation equations for a conventional bundle
adjustment. These are of the following form:
V.BA «C (1)
= 458 Al - e
V=iV B=|-1 0 A=} 1 C- Cl.
0 1 A G
V = vector of plate observation residuals,
V = vector of exterior orientation parameter
observation residuals;
V - vector of object coordinate observation
residuals,
B = matrix of partial derivatives wrt exterior
orientation parameters;
B = matrix of partial derivatives wrt object
coordinates;
A = vector of exterior orientation parameter
corrections;
A = vector of object coordinate corrections;
e = vector of plate observation descrepancies,
C = vector of exterior orientation parameter
descrepancies,
C - vector of object coordinate discrepancies;
The structure of the camera parameter portion of the
normal equation matrix produced by the least squares
solution to this model is shown in figure 1.
Figure 1 Normal equation structure of the camera
parameter portion of a conventional bundle adjustment of
stereopairs.
Each camera is represented by a 6x6 symmetric sub-
matrix. The total number of camera parameters to be
solved for a bundle of s stereopairs is 12s. An overview of
photogrammetric bundle adjustment programs can be
found in Karara (1989, Ch6). Bundle adjustments
developed in this study were based upon this model.
MODELLING OF CAMERA INVARIANCE
The invariance that exists between the cameras of a
stereopair may be divided into two relationships: