790 
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
2. METHODOLOGY 
Photogrammetric techniques define the shape, size and position 
of objects using images taken from different viewpoints. 
Photogrammetric reconstruction is based on the collinearity 
equation, which states that the image point, the perspective 
centre and the corresponding object space point are collinear. 
The internal orientation parameters (IOPs) of the implemented 
camera, which include the principal distance of the camera (c), 
the coordinates of the principal point (x p , y p ) and distortion 
parameters, are accurately recovered through a camera 
calibration procedure. The exterior orientation parameters 
(EOPs) define the position (X 0 , Y 0 , Z 0 ) and the orientation (to, 
cp, k) of the reconstructed bundle relative to the object space 
coordinate system. The EOPs simulate the actual position and 
orientation of the camera at the moment of exposure. A 
photogrammetric system first identifies a pair of conjugate 
points in overlapping areas between two 2D images acquired 
by calibrated cameras. Reducing the search area for conjugate 
features achieves better results and reliability. Epipolar 
geometry is commonly used to constrain the search in matching. 
Conjugate light rays can be reconstructed after identifying 
conjugate points. The intersection of two conjugate light rays 
defines an object point in 3D space. According to the above 
concepts, the proposed procedures for 3D reconstruction 
modeling require the following fundamental procedures: Image 
acuqisition, epipolar transformation, matching and intersection, 
(see Figure 1). 
2.1 Image Acquisition 
First, the utilized cameras undergo a calibration and stability 
analysis procedures. The objective of the calibration process 
(Habib and Morgan, 2003) is to derive the cameras’ internal 
characteristics including principal point coordinates, principal 
distance, lens distortions, etc. The stability analysis, on the 
other hand, aims at verifying that the estimated internal 
characteristics do not significantly change over time (Habib et 
al., 2006). 
Provided with initial estimates of EOPs and a test field with 3D 
points which are measured in advance, bundle adjustment can 
perform an estimation which minimizes the re-projection error 
by adjusting the bundle of rays between each camera centre and 
the set of 3D points. The EOPs of the camera at the moment of 
exposure can be obtained through a bundle adjustment 
procedure. 
Since the human face is a relatively homogeneous surface, few 
conjugate features can be identified. To overcome such a 
limitation, structured patterns can be projected onto the face 
during image acquisition to increase the density of identifiable 
points on the facial surface. The pattern projection technique 
was selected for several reasons. First, the pattern projection is 
especially useful for providing artificial landmarks in 
homogeneous areas by projecting a light pattern on the face. 
Second, this setup is relatively fast, inexpensive and enables 
acquisition of 3D information using easily available and low 
cost digital cameras. The additional cost is limited to a 
projector. Eleven 3 by 3 sub-blocks were used for the encoding 
pattern. The sub-blocks were randomly selected and arranged 
for this pattern (Figure 2). To minimize ambiguity, the 
sub-block should not be repeated within a certain radius. 
After an imaging environment with pattern projection is setup, 
a subject with projected pattern can be imaged by using 
calibrated cameras with known IOPs and EOPs at different 
locations. 
Figure 1. The flowchart for 3D reconstruction system design 
using pattern projection 
Figure 2. The designed pattern for pattern projection 
2.2 Epipolar Transformation, Feature Extraction and 
Matching 
The acquired images have to be pre-processed in order to 
perform better matching. The captured images are first 
resampled through Epipolar transformation. The main objective 
of epipolar transformation is to generate normalized images 
with corresponding points on the same rows. Epipolar 
trasnformation of frame images is a straightforward process. 
The resampling process involves projecting the original images 
onto a common plane in an orientation determined by the 
parameters of the original images. Original and normalized 
scenes share the same perspective centre. During normalization, 
the two optical axes should be parallel to each other and 
perpendicular to the baseline. For a normalized pair, we can 
search the conjugate points along the corresponding row. 
During matching process, conjugate points can be more reliably 
detected by identifying features on the surface. In the proposed 
algorithm, feature extraction can be achieved by using the 
Harris comer detector (Harris, 1988), which is a popular 
interest point detector due to its reliable invariance to: rotation, 
scale, illumination variation and image noise (Schmid, 2000). 
Searching conjugate points in stereo digital images can be 
automated using image matching procedures based on well