IMPLEMENTATION OF SEQUENTIAL ESTIMATION FOR SINGLE-SENSOR VISION METROLOGY
Kenneth Edmundson and Clive S. Fraser
Department of Geomatics
University of Melbourne
Parkville, Victoria 3052, Australia
email: kle 9 sunburn.sli.unimelb.edu.au
Clive Fraser 8 mac.unimelb.edu.au
Commission V, Working Group 1
KEYWORDS: Algorithms , On-line Triangulation, Precision, Sequential Estimation, Vision Metrology
ABSTRACT
Investigations into on-line triangulation, robot vision, image sequence analysis, and autonomous vehicle navigation
have established the merits of sequential estimation methods in aerial and non-topographic photogrammetry utilising
both film-based cameras and digital sensors. These merits generally focus on enhancement of the speed and
efficiency of the triangulation procedure through the incorporation of quality control and observational error detection
into the measurement procedure. The on-line quality control of industrial measurement with vision systems utilising a
single sensor such as a CCD camera is a natural extension for sequential techniques. This paper examines how the
sequential estimation process may be incorporated into single-sensor vision metrology for typical industrial
photogrammetric inspection. Issues investigated in the context of the industrial application include the sequential
nature of data collection and adjustment, the influence of normal equation structure on system response, generation of
approximate values, additional parameters for systematic error compensation, blunder detection procedures, and
datum establishment.
With regard to datum establishment, a factorisation method for recursively updating the
equation system obtained in a free-net adjustment by inner constraints is suggested.
1.0 INTRODUCTION
Multi-camera, stereo configurations have, up to now,
been the focus of the bulk of the research effort in close-
range vision metrology (VM). The recent commercial
availability of large-area, high-resolution CCD cameras,
coupled with the proven advantages of a single metric
camera for high accuracy measurement, has heightened
the potential for the single-sensor VM system in industrial
inspection. The performance of industrial measurement
tasks such as localised inspection, re-work, and fit
checking is additionally enhanced through the use of VM
systems containing an on-line link between the camera
and an external computer. While real-time three
dimensional measurements are not achievable with the
single camera system, the near real-time image
measurement capabilities associated with digital
imagery, in combination with sequential estimation
techniques such as on-line triangulation (OLT) can
provide rapid data turnaround.
The acceptance of CCD cameras for industrial
photogrammetry continues at a pace which is
constrained primarily by questions of accuracy related to
the typically reduced format and resolution of CCD
sensors as compared to medium and large format film-
based metric cameras. Studies presented in Fraser &
Shortis (1995) and Maas & Kersten (1994) have
indicated that CCD and still video cameras such as the
Kodak DCS200 (and DCS420) can yield acceptable
accuracies for many industrial measurement tasks. One
consequence of the lower resolution afforded by CCD
sensors is that significantly more images may be
necessary to achieve a precision comparable to a
network obtained with a metric film camera. The
potential of the VM system eases previous limitations in
the number of images that may be readily processed and
allows their incorporation into the network with minimal
time expenditure. After an optimal convergent camera
station network is in place, the use of multiple exposures
is the principal means of improving object space
precision. Here, OLT can serve as a mechanism for
recursively monitoring object point accuracy. The
photogrammetrist, while still on site, can interactively
strengthen the network geometry until the desired level of
accuracy is obtained.
OLT methods have typically focused on the detection and
removal of gross errors. By incorporating quality control
and observational error detection into the measurement
process, the speed and efficiency of the overall
triangulation is enhanced. A comprehensive historical
background of sequential estimation as applied to OLT
can be found in Gruen (1985). Several recent studies
apply these methods in non-topographic applications.
These include robot vision (Gruen & Kersten, 1992),
image sequence analysis (Kersten & Balfsavias, 1994),
and autonomous vehicle navigation (Edmundson &
Novak, 1992). The suggested application of sequential
estimation in OLT to industrial quality control (Kersten et
al, 1992) has for the most part remained unexamined.
This paper, which builds upon work reported in
Edmundson & Fraser (1995), explores the utilisation of
sequential estimation for OLT in single-sensor vision
metrology. We begin with a re-examination of the
mathematics behind the general sequential estimation
problem focusing on the computational algorithm, an
orthogonal transformation technique known as Givens
133
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part BS. Vienna 1996