ALGORITHM DEVELOPMENTS FOR AUTOMATED OFF-LINE VISION METROLOGY 
J. O. Otepka ^, H.B. Hanley ? C.S.Fraser ^" 
“ Institute of Photogrammetry & Remote Sensing, Vienna University of Technology, 1040 Vienna, Austria - johannes@avt.at 
? Department of Geomatics, University of Melbourne, Victoria 3010, Australia - c.fraser(gunimelb.edu.au, 
hanley@sunrise.sli.unimelb.edu.au 
Commission V, WG V/1 
KEY WORDS: Vision metrology, automation, sensor orientation, close-range photogrammetry, algorithms 
ABSTRACT: 
The goal of developing automated off-line vision metrology systems has largely been accomplished over recent years. The emphasis 
of research into measurement automation has thus moved on to implementation of more flexible and robust algorithms and processes 
in order to both improve performance and better accommodate data errors. This paper describes key features of the automatic 
measurement operation adopted for the Australis off-line digital close-range photogrammetric system. Particular aspects covered 
include target detection and validation within image scanning; the design of an exterior orientation device to support initial image 
orientation; image point correspondence determination based on the geometry of epipolar planes; and the integration of these 
processing phases with both a preliminary and final bundle adjustment. As well as discussing the fundamental models and 
computational schemes adopted, the paper will address error handling strategies. It will also consider practical aspects of these 
developments and briefly report on results of experimental performance tests. 
1. INTRODUCTION 
In general terms, the process of image measurement and 
photogrammetric triangulation in automated off-line vision 
metrology (VM) systems follows that of traditional multi- 
station close-range photogrammetry: image coordinates are 
measured and labelled, preliminary exterior orientation (EO) 
and object point triangulation are performed, and bundle 
adjustment with sensor self-calibration is employed for the final 
object point determination. With the automated approach, 
however, the traditional sequential point-wise process for image 
mensuration has given way to image scanning, where in a single 
operation all potential targets are detected and validated. 
Moreover, through the incorporation of specific target patterns, 
the concepts of automatic resection via an EO device and 
network building through coded targets have been realised, as 
has robust feature point correspondence determination. 
The data processing scheme for off-line VM systems now 
generally follows a sequence of image scanning with target 
detection and validation; identification of target groupings 
constituting EO devices and coded targets; initial EO; 
correspondence determination, object point labelling and initial 
triangulation; and finally bundle adjustment. With the 
implementation of such an automated computational sequence, 
there is a need to take account of the impact of data errors. 
When compared to photogrammetric surveys with manual 
image measurement, two features of the automated process 
stand out as far as data errors are concerned: there are likely to 
be more observational blunders, and it is probable that these 
will arise through the addition of invalid targets rather than the 
omission of valid observations. As a consequence, the automatic 
measurement process must prove itself to be robust and reliable 
in the presence of both different image qualities, and many 
more data errors than might otherwise be expected. 
In this paper the integrated image measurement and 
photogrammetric triangulation process of the Australis software 
system for off-line VM (Fraser & Edmundson, 2000) is 
described, with particular reference to the goal of achieving 
robust performance in the treatment of inherent data errors. The 
coverage includes the operations of image scanning and initial 
EO determination, image point correspondence determination, a 
3-stage approach to photogrammetric triangulation, as well as a 
brief account of the results of performance tests. A more 
comprehensive account is provided in Otepka (2001). 
2. IMAGE SCANNING 
2.1 Target Quality 
The first data processing step within an automated VM process 
is the detection and measurement of artificial targets in each 
image. Whereas image recognition remains a significant 
challenge in image processing and artificial intelligence, the use 
of high-contrast targets has considerably simplified the 
development of straightforward target detection algorithms for 
VM, though there is a necessity for the image scanning 
algorithms to take account of differing image quality. Under 
ideal conditions, the use of retro-reflective targets will yield 
near binary images and ‘high-quality’ targets, as indicated in 
Figure la. 
In many practical instances, however, the contrast difference 
between the targets and background may not be as distinct, for 
example when utilising non retro-reflective targets or when 
operating in daylight conditions. It is thus necessary to account 
for such ‘low-quality’ images, exemplified by Figure 1b, as far 
as is practicable within the automatic image mensuration 
process. A more complex detection and validation process is 
warranted for low-quality images, which by their nature can 
lead in many instances to lower photogrammetric triangulation 
accuracy. 
In this section the procedure developed for image recognition 
and detection, target validation and centroid measurement 
within the Australis system is described, with the discussion 
being confined to the use of greyscale images only. The strategy 
adopted takes into account the anticipated target quality 
variations, as indicated in Figure lc, by including distinct 
processes for high- and low-quality images. 
An analysis of the imaged targets in Figure lc shows that 
irrespective of the contrast difference between the targets and 
—60- 
the 
ed 
the 
ed 
ea 
to 
be 
ad 
Lu 
Fig 
lov 
The 
the 
din 
anc 
in, 
Go 
nel; 
pos 
vali 
adv