ne vision 
| medical 
a risk of 
1 reliable 
d on the 
Or serves 
ng points 
of point 
s well as 
rator and 
uities in 
ind a cost 
Range 
Ss kind of 
igate the 
nodelling 
method 
'spective, 
nd hence, 
bth maps 
| textured 
lel which 
l., 2002). 
ilti-Photo 
ecause it 
ver, the 
ind hence 
atching is 
tions are 
| the light 
e same in 
esult, the 
ut, as an 
siderable 
eading to 
; images. 
vhere the 
flexes for 
00). The 
specially 
ffer from 
d be kept 
tment by 
ct of the 
le extend 
intensity- 
based techniques, however, have been successfully applied to 
aerial imagery where the assumptions of an area-based 
matching are approximately satisfied. 
In this research, we introduce MEDPHOS (MEdical Digital 
PHOtogrammetric System) and propose a strategy for a robust 
trinocular vision system fulfilling the demands of the medical 
society particularly for wound analysis. The main goal in this 
research has been improving the robustness of the multi-camera 
system by means of geometric constraints. 
2. SYSTEM DESCRIPTION 
MEDPHOS consists of three synchronised digital cameras and 
a projector (Fig. 1). A more detailed technical description of the 
system can be found in (Boersma, et al., 2000). 
  
  
Figure 1. The Trinocular Vision System 
2.1 Configuration 
Originally, the goal of introducing the third camera was to 
reduce the uncertainty regarding the goodness of local matching 
and having more redundant data to improve reliability (Maas, 
1999). In our method the main reason is making the 
correspondence problem entirely constrained by T7 rifocal 
Epipolar Geometry concepts and improving the robustness by 
rejecting non-homologous triplets. Our method requires at least 
three cameras and cannot be applied to the stereo case. The 
ambiguities in matching cannot be solved solely by geometric 
constraints in a system based on two cameras. The trinocular 
vision system has rich geometrical constraints that can be used 
to reconstruct the 3D wound faithfully. The expected number of 
ambiguities in a three-image arrangement is (Luhmann, 2000): 
(a? - p 
N VOU n (emm (1) 
a Fsina bys bi 3 
Where 7 is the number of points, F is the image size, € is the 
tolerance or width of the epipolar band, a is the intersection 
angle of epipolar lines in the third image, and b;s are the 
camera base lines. Minimisation of this equation leads to the 
optimal configuration b,2=b;3=b23 and a=60 degrees of the 
cameras arrangement with respect to minimum number of 
ambiguities which means a configuration of the three projective 
centres in a equilateral triangle. 
  
  
  
  
  
  
Figure 2. The epipolar band, after (Luhmann, 2000) 
Furthermore, There is a general trade-off in image matching 
between the requirements of precision (long base lines) and the 
requirements of reliability (short base lines). Based on the 
requirements of the project, a base-to-depth ratio of 1:2 was 
adopted. 
2.2 System Calibration 
In order to efficiently exploit the epipolar constraint as well as 
to compute the depth values at the final stage, the uncertainties 
of the image measurements and the quality of calibration 
process should be taken into consideration. Moreover, the 
epipolar line constraint holds as long as the lens distortions are 
not too large to be negligible or, more generally, as long as the 
perspective projection can be modelled by a projective 
transformation. If one does not perform a reduction of the 
observed pixel coordinates for lens distortion, the 
corresponding points still have to sit on lines that, however, are 
not straight anymore (Faugeras, 2001). For this reasons, a 
complete calibration procedure based on self-calibration bundle 
adjustment is conducted in MEDPHOS. 
2.3 Pattern Projection 
Image matching requires finding points of interest in the image. 
Interest here has several meanings (Schmid, et al., 2000): 
Distinctness. Points should be distinguishable from immediate 
neighbours. 
Invariance. The position and selection of the points should 
remain stable with respect to the perspective transformation. 
Uniqueness. The points should be globally separable to avoid 
confusion in the matching process. 
Image features or elements used in vision problems must be 
reliably extractable from images. Moreover, if these elements 
match, then their attributes should be (approximately) the same. 
As the wound surface does not show sufficient texture 
(features), we use artificial features by projecting structured 
light in the form of a dot pattern onto the wound surface. 
2.4 Image Preprocessing 
In close range photogrammetry, proper illumination is an 
essential criterion for the correct recognition and extraction of 
features of interest. In poor conditions, the feature extraction 
procedure delivers a lot of elements that can be neither 
construction details nor homologous elements. In order to 
improve the feature extraction, image filtering is applied as 
described in the following sections. 
—265— 
  
  
RTS" CBE ei aa