Face Biometrics and Tracking
The natural way of recognizing a person is by a look in her/his face.
Everybody is used to it, be it
the observer or the observed.
It is therefore obvious that biometrics need to include techniques to
process images or videos
containing faces since it is
one of the most common tasks that
we do with our vision
routinely, and with an amazing
accuracy.
Given images or live camera footage, the tasks of face biometrics are
to detect faces and to recognize the person's identity, or
at least to categorize
it with regard to
known/unknown,
age, gender - like a human would do. Then comes the
task of tracking a
person robustly in subsequent
images. In analogy with human
visual intelligence, the machines should
be able to track with or
without knowing the identity of the person. However, tracking
means
much more than this. It may
also mean to track body parts accurately,
e.g. eye,
nose, mouth, and limbs.
Accordingly, complex behaviors involving other objects can be recognized, for example, people carrying loads, danger of pedestrian-vehicle accidents.
The most significant
challenges in face biometrics
and tracking are probably robustness against
illumination changes and cluttered background. We apply different features,
including symmetry descriptors
and Gabor filters.
The latter demonstrably have
analogies in the
human visual cortex. We employ Support Vector Machines and Boosting techniques for pattern recognition.
Examples of successful face
localization in bad light
conditions are on the left,
feasible by our methods (benchmark YALE
database). We are also
pioneers to offer several anti-spoofing measures, able to avert advanced attacks by portable video devices.