AUTOMATIC GENERATION OF FACIAL DEM s. 
F.A.S. Banda,* J-P. Muller,* S.N. Bhatia, M. Bukhary. 
* Department of Photogrammetry and Surveying, University College London, 
Gower Street, London WCIE 6BT, United Kingdom. 
INTERNET: banda\muller@ps.ucl.ac.uk 
Department of Orthodontics, Kings College School of Medicine and Dentistry, 
Kings College Hospital, Caldecot Road, London SES 8RX, U.K. 
ABSTRACT. 
There are a handful of methods currently employed to produce three dimensional computer models of the human face. Among these, 
stereophotogrammetry and laser profiling have been used extensively. Unlike other methods in use the method presented utilizes 
multiple stereo models of four charge-coupled device (CCD) captured images and for comparative purposes four Rollei 6006 images 
to cover the whole face. Using a least squares grey level stereomatcher and seedpoints automatically generated, a dense disparity map 
is produced by a matching process which grows out from the seed points. From camera parameters which are determined by the 
bundle adjustment method the disparity map is transformed to an absolute co-ordinate system of the control points within the object 
space. The resulting data set is used to provide surgeons with pre and post surgical management information including profiles of the 
face, angles and distances between strategic features. 
KEYWORDS: Digital Elevation Models, automated seedpoints, area correlation matching, facial surface models. 
1.0 INTRODUCTION. 
Orthodontists and maxillo-facial and plastic surgeons require 
knowledge of the shape and size of the human face to estimate 
population norms and to evaluate changes with growth and 
cosmetic facial and jaw surgery. À careful metrication of the 
facial surface is needed to meet the above requirements 
(Balagh et al, 1990). 
A number of methods involving a matrix of mechanical 
probes, laser holography, Moire fringe patterns and 
stereophotogrammetry have been investigated as possible 
ways in which three dimensional records could be made of 
human heads. Each method has its own merits and demerits 
ranging from accuracy requirements, safety factor to the 
subject, complexity and cost of analysis. The method 
employed in this paper is stereo multi-camera 
photogrammetry. However, for medical tasks photogrammetry 
is confronted with problems where the differences to be 
measured between the original and changed face have to be 
done in the absence of identical points or areas of the face. The 
special difficulty lies in the exact definition of a reference co- 
ordinate system as everything on the human face is changing 
or imprecisely defined. Since the photogrammetrist and 
surgeon have different aims, the photogrammetrist has to find 
what information (lengths, angles, areas, volumes, shifts, 
rotations, inclinations, scale changes, asymmetries etc) is 
needed and how accurately it has to be determined and 
measured. Consequently, a suitable method to visualize the 
results of the measurements or computations has to be 
developed so that the surgeon immediately sees what he needs 
to see as well as provide simple tools which require minimal 
training to enable the surgeon to make photogrammetric 
measurements. 
2 EQUIPMENT SETUP, IMAGE ACQUISITION AND 
PROCESSING. 
The system described has been installed in the Orthodontics 
department of Kings College Dental School and consists of 
four Pulnix CCD cameras and four Rollei 6006 cameras 
mounted on a rigid semi-circular bracket as shown in figure 1. 
These cameras are used to acquire images of a subject ın a 
convergent manner at the same time in order to get complete 
coverage. 
Projector 
  
Cameras : Cameras 
1 Subject 4 
t1 
Calibration 
target 
Fig. 1: Setup of cameras and projector on the rig. 
To capture images, the subject is made to position his/her head 
within a three dimensional control target consisting of twenty- 
five control points distributed around its volume. Some fine 
scale texture in the form of either grid lines or random texture 
is projected onto the subjects face. Once the operator is 
satisfied with the position, the cameras are triggered 
simultaneously with the help of a switch. The four CCD 
cameras are connected to a Matrox IP-8 acquisition board with 
2Mbytes of on board memory installed on an 80386 IBM 
compatible PC running at 25MHz. The board comes with 
software primitives for performing multi-frame acquisition. 
The size of each image is 512x480 pixels and a pixel 
quantisation of eight bits. The distribution of control points 
over the target is such that at least eight control points are 
imaged on each frame, hence providing an over determined 
problem when solving for the orientation parameters of the 
cameras in the bundle adjustment. Once the stereo imagery has 
been obtained, they are processed to reduce noise and 
correlated using a coarse-to-fine area based stereo-matcher and 
in the conventional manner using stereoplotting. 
The production of surface models from pairs of stereo images 
may be subdivided into three independent stages. Firstly, a 
stereo-matching procedure (Otto & Chau, 1989; Muller, 1989) 
is used in order to identify a dense array of conjugate points. 
The output from the stereo-matching stage is a dense Digital 
Disparity Model (DDM) or a set of 2-D [x,y] correspondences