al
in
Hf
3. RESULTS.
A pair of stereo images were acquired using the close range
vision cell at the Orthodontic Department, Kings College
Dental Hospital London. These images (only with a grid
projection) are shown in figure 4 and figure 5. The stereo-
matcher was applied to the images in a conventional manner
and in a pyramidal manner using both manually and
automatically generated seedpoints. Fig 4 shows the left and
right images with the seedpoints while figure 5 shows the
matched areas superimposed on the pair of images. Table 1
shows the number of matched points and the time taken to
stereo-match the images on an unix workstation. Figure 6 is a
plot of the processing time against matched points
Pattern Number of | Matches CPU time
Projected seedpoints
Regular Grid 12 3671 410.6 s
Random 34 4611 512.6s
Table 1: Stereomatcher results using automatically
generated seedpoints on the 512x480 images.
Image Size Res ition Matches ng
32x30 16 273 1235
64 x 60 8 1 845 109.0 s
128 x 120 4 8 783 761.5 s
256 x 240 2 9 676 1 393.3 s
512 x 480 10 099 1 664.1 s
Table 2: Results of the coarse-to-fine matching.
2000
1000 4
Exponential function
fitted over the points.
CPU time(sec)
" T- —_ —_ T—T —_ — ey
0 2000 4000 6000 8000 10000 12000
Matches
Fig 6. A plot of CPU time against matches of the
coarse-to-fine matching.
A triangulated surface consisting of Delaunay triangles through
the set of 3-D points is then generated. The
surface which is generated consists of triangles using a subset
of the stereomatched input points as vertices. Not all of the
input points are used. If the difference in height between a
point at position x,y and the value of z at x,y on the surface of
the triangle encompassing that point is less than a tolerance
value given as an option on running the triangulation program,
then the point is considered unnecessary and is omitted
(DeFloriani, 1989). Hence a surface is generated using the
Fig 8. Example of a facial DEM generated from
output of the stereo-matcher.
minimum amount of data needed to represent the surface to
within the specified tolerance figure 8.
4. CONCLUSION.
A system has been described which can be constracted to
generate automatically a large set of stereo correspondences so
that a facial surface model may be produced. The measurement
accuracy of this system meets the requirements for surgical
planning and treatment monitoring. In facial applications, the
widespread acceptance of 3-D non-invasive biostereometric
systems depends on the ease with which these systems can be
used. The biostereometric system advocated here uses
autoseeding area correlation matching algorithm in conjunction
with a pre-calibrated cell. The system is therefore well suited
for use by medical personnel who are not skilled in either
Photogrammetry or computing. The speed of the system could
be improved in the future by using transputer elements based
on T9000 ( see Zemerly et al., 1992)
5. ACKNOWLEDGEMENTS,
This project is being funded mainly by the Swedish
International Development Authority (SIDA) and partly by
research funds of the Orthodontic Department, Kings College
School of Medicine and Dentistry, London. The authors
would like to thank Andrew Deacon.
6. REFERENCES.
1. Allison, D., Zemerly, M.J.A., Muller, J-P. 1991.
" Automated seedpoint generation for stereomatching
and multi image registration." Proc. IGARRS,
Helsinki, Finland.
2. Balagh, B., Rasse, M., Waldhausl, P. and Forkert, G.,
1990. " Photogrammetric surveys of human faces for
medical purposes." | Close-Range Fhotogrammetry
meets Machine Vision.Zurich, Switzerland. p 704.
3. Deacon, A.T., Anthony, A.G., Bhatia, S.N., Muller,JP.,
1991. " Evaluation of a CCD-based facial measurement
system." Medical Informatics. 16(2) pp 213 - 228.
4. DeFloriani, L. 1989. " Surface description: A pyramidal
data structure for triangle-based surface description."
IEEE Computer Graphics and Applications 3;
pp 67-78.
