789 
RECONSTRUCTION, REGISTRATION, AND MATCHING OF 3D FACIAL MODELS 
FROM STEREO-IMAGES 
Yu-Chuan Chang and Ayman F. Habib 
Department of Geomatics Engineering, University of Calgary, 2500 University Drive NW, Calgary, Alberta, Canada 
T2N 1N4 - (ycchang, habib)@geomatics.ucalgary.ca 
WgS: WGV/6 
KEY WORDS: Reconstruction, Registration, Photogrammetry, Matching, Biometrics, Computer Vision 
ABSTRACT: 
Interest in biometric systems has dramatically increased worldwide due to the rising public demand for more reliable and effective 
criminal identification and effective surveillance systems. Biometric facial recognition has been studied intensely over the past 
decade due to its potential applications in government, law enforcement and business. Previous studies in facial recognition have 
tended to focus on using two-dimensional (2D) images. The performance of current 2D image-based facial recognition systems, 
however, is clearly unsatisfactory due to their sensitivity to changes in facial expression, body position, lighting, sensor conditions 
and other disturbance factors. Further, some 2D systems are easily fooled by simply presenting a photo in front of the sensor. The 
move from 2D to three-dimensional (3D) recognition technology is expected to improve performance in real-life environments. This 
study developed a low-cost imaging system to capture overlapping imagery which was then used to construct a 3D facial model. The 
generated 3D model was then registered and matched with available 3D models in a central database for personal verification or 
identification purposes. The experimental results showed that low-cost digital cameras, after proper calibration, can construct 
accurate 3D facial models when combined with an active pattern projection system. Preliminary experiments also demonstrate the 
feasibility and robustness of the proposed automatic surface registration and matching procedure. This study discusses the 
performance, advantages and limitations of the proposed method. 
1. INTRODUCTION 
Biometric measurements are being studied for security 
applications as an alternative to Personal identification number 
(PIN) codes and cards. In cooperative environments, speech 
and face modalities are well accepted by individuals but have 
yet to demonstrate acceptable performance and reliability. A 
previous profile analysis (Beumier, 1995) demonstrated the 
adequacy of geometrical information for automated personal 
authentication. Such information is derived by analyzing rigid 
areas of the face, such as the forehead, nose and chin and are 
generally unaffected by makeup or lighting conditions. This 
explains the success of many profile works (Chellappa, 1995). 
A facial 3d description requires more geometrical information, 
especially where grey level features are lacking as in the chin, 
forehead and cheek regions. Such an analysis would benefit 
from actual 3D measures without scale or rotation influence. 
Depth information also helps to distinguish the face from 
background elements. These advantages make 3D geometrical 
approaches an important complement to grey level analysis. 
Previous research in 3D facial recognition has yielded two 
main techniques: the stereo acquisition technique and 
laser-based acquisition technique. Most recent research have 
utilized laser scanning systems, which are typically more 
accurate but also more expensive and time consuming 
(Achermann, 1997). The acquisition of a single 3D head scan 
can take more than 30 seconds, which is a major limitation of 
laser-based systems. This relatively long scan time introduces 
errors due to the object’s breathing and movement resulting in 
errors in 3D reconstruction of a face. 
The stereo acquisition technique employs two or more cameras 
positioned and calibrated to simultaneously acquire images of 
the subject (Lao, 2000). The location for each point in 3D 
object space can be computed by using a photogrammetric 
procedure. This method has the lowest cost and highest ease of 
use. A simultaneous acquisition system using 
stereo-photogrammetric technologies does not have serious 
motion problems that a laser-based acquisition system has. 
However, photogrammetric systems are not widely used in 
commercial applications for facial model recognition due to 
their unsatisfactory automation of matching, which is a process 
for identifying conjugate landmarks of an object’s surface in 
stereo images. 
Landmark locations used for matching in previous work can be 
found either manually (Lao, 2000) or automatically (Yacoob, 
1994). The correct localization of the landmarks is crucial to 
many algorithms, and judging the sensitivity of an algorithm to 
localization errors by its description alone is usually not 
possible. Nevertheless, automatic landmark localization 
remains an unsolved problem. Among the possible optical 
acquisition systems (Jarvis, 1993), structured light has emerged 
as the optimal solution for identifying landmarks in 
homogeneous areas. 
This work presents an efficient and automatic algorithm for 3D 
model reconstruction from facial images. A photogrammetric 
model is proposed for surface reconstruction from stereo pairs 
using pattern projection. Its resolution, speed and adequate 
facial coverage make it attractive for numerous practical 
implementations. The next section briefly describes the 
photogrammetric principles for 3D facial measurements. The 
proposed system design for homogeneous surface model 
generation is then described and experimental results are 
presented. Finally, conclusions regarding the effectiveness and 
possible uses of the proposed system are summarized.