anticipated parallax range is, in the terms of the 
scale of the object, 10mm. For an object dimension 
of 250mm, this suggests that 25 target points across 
the object in each direction (or about 600 targets 
across the surface of an object which fills the 
entire view of the cameras), would ensure that all 
conjugate targets and only conjugate targets fell 
within the expected parallax ranges. This figure 
corresponds to a target every 20 pixels in each 
direction on an image of 500 pixel dimension. 
Projected grid lines are usually at about this 
spacing and are about 1.5 mm wide on the object. 
In order to meet the cost and simplicity criteria, 
all image processing is carried out on an ordinary 
IBM-compatible PC with maths co-processor and a VCA 
monitor. 
3.3. Computational Stages 
A suite of programs, written in FORTRAN, has been 
prepared to carry out the various stages of the 
computations: 
i) to display the camera moving images before 
adopting them; 
ii) to grab and store the images at the chosen 
moment within 0.1 seconds of each other; 
iii) to detect targets on each image, in order to 
produce a finite list of points at which precise 
matching should be successful; 
iv) to classify the targets, thereby providing 
information to facilitate subsequent matching; 
v) to pair possible conjugate target points; 
vi) for stereoscopic image correlation to provide 
image co-ordinates; and 
vii) for surface reconstruction and display. 
3.4 Target Detection Classification. 
Various procedures which search for features suitable 
for stereo matching within the pattern have been 
explored, including the use of template matching and 
the interest operator described by Fórstner & Gülch 
(1987). 
A regular pattern is commonly supposed to cause 
confusion and erroneous matching, although it has 
been used in other medical and close-range measuring 
devices, e.g. Rüther (1989), Rüther & Wildschek 
(1989), Trinder et al. (1990), Deacon et al. 
(1991). The grid has been used successfully here to 
provide distinct target points, with known 
characteristics, at a pre-selected spacing. Areas 
within the grid pattern possesses a symmetry, which 
is obvious at the grid intersection points, but which 
is also apparent at the centres of the areas between 
intersections and at the centres of the lines between 
intersections. This is so even when the pattern is 
distorted by the perspective, and this property can 
therefore be utilised for target point selection. 
Indeed, tests showed that around 2000 target points 
could typically be detected in the image by finding 
symmetry, but the image processing is laborious. 
For a coarser resolution, the peaks in intensity 
across the pattern in the filtered image provide a 
very simple means of detecting the grid 
intersections. Such a procedure permits very fast 
target detection but usually gathers only around 400 
targets. 
    
3.5 Target Classification 
The interest operator referred to in Section 3.4 is 
subsequently used for classification of features 
detected by the symmetry of the pattern. The target 
types are distinguishable by the characteristics of 
the "error ellipses" derived from the image intensity 
gradients in the vicinity of the targets. Grid 
intersections can also be differentiated from the 
points at the centres of the grid squares on the 
basis of their reflectance intensity. 
3.6 Pairing of Points 
Pairing of prospective match points on both images of 
the stereo-pair is based on parallax limits defined 
by epipolar geometry, within some bounds governed by 
the expected variations in depth across the object 
surface and with some small tolerance (typically a 
couple of pixels) to allow for error in locating the 
same target in each image and any imprecision in the 
relative orientation. If a large number of targets 
have been detected, a number of targets in the right 
hand image may be paired with any one target in the 
left hand image, this number depending on the pre-set 
parallax limits. These limits can be varied 
according to the degree of convolution of the object 
surface. The multiple target matches than have to be 
contended with in the precise matching - see Section 
3.7 Image Matching 
To correlate conjugate targets, least squares signal- 
based matching was chosen in preference to feature- 
based matching because of the former's superior 
precision. Tests of least squares image matching 
with the regularly patterned surfaces showed that it 
works reliably; sub-pixel precision matching can be 
undertaken not only at the grid intersections but 
also at other types of target points. The incidence 
of poor matches, blunders and mismatches of like 
features can be limited if run-time parameters are 
carefully selected. Moreover, the statistics of the 
least squares match, particularly the variance 
factor, can be used to distinguish mis-matches from 
correct matches, in which case the grid spacing can 
be reduced to improve resolution. The pull-in range 
of the least squares matching was found by tests to 
not be an obstacle to the least squares matching in 
this case. 
The established 8-parameter least squares matching as 
given by, for example, Albertz and Kreiling (1989, 
p260), is used but with modifications to accelerate 
processing. Geometric constraints are not 
incorporated on the assumption that this requires 
that the target point co-ordinates detected in an 
earlier stage already satisfy the geometric 
constraint by lying exactly on the epipolar lines. 
Internal match precision on the images of the grid on 
skin varies, but rarely exceeds 0.3 to 0.5 pixels, or 
about one part per thousand of the image dimension on 
skin surfaces. For an object at one metre, this is 
about 0.2 mm. In fact, if the precision exceeds a 
certain pre-selected tolerance level, the matched 
point is rejected. 
A number of strategies which have improved the 
reliability and/or precision of the least squares 
matching in recent years, have been published since 
the early exposition of the least squares matching 
theory (e.g. Ackermann, 1984), but they have not been 
incorporated into this work because of their 
perceived detrimental effect on the speed of 
computation. These include combined object surface 
and radiometric modelling (see Weisensee & Wrobel, 
1991, Heipke (1990,1992), Wrobel (1991), among