Real-Time Face Recognition from Surveillance Video - Intelligent Video Event Analysis and Understanding - page 189

videos were processed by the front-end and feature points were sent to the back-end
for identification. The back-end calculated the feature vector for each frame and
returned identification results as shown in table 2.

Ta b l e 2 Results of identification test for two subjects

Faces Faces False False

Subject found identified positive negative

(front end) (back end)

Adrian 20 20 0 0

Darryl 38 38 0 0

These tests can be considered as a proof-of-concept. It would require a larger
set of test data to do an accurate comparison of the discriminative capacity of the
feature vector.
It was not possible to test on rotations larger than 10% as we were using the ear
tips as a feature, and at greater than 10% rotation, one ear is occluded. This could
be compensated for by adding more features and by using missing feature theory to
ignore the effects of occluded features.

6 Conclusions

This project has demonstrated a system for face recognition from surveillance video.
The system was demonstrated to be robust to changes in illumination, scale, facial
expression and reasonably robust to occlusions and changes in pose. Attention was
given to performance considerations, and the system can operate in real time. All of
these features make this approach suitable for a video surveillance application.
One of the notable findings was the improvement in both performance and accu-
racy of the O PEN CV face detector, when the Viola-Jones face detection was com-
bined with an analysis of colour skin-tone information. Colour information could
be exploited further to improve the detection of local features as discussed in [10].
Our feature-detection algorithms were demonstrated to be invariant to changes in
lighting direction.
The local features that we used were irises, nostrils and ear-tips. Irises and nostrils
are reasonably straightforward to extract in frontal view, although reflected light
from glasses or a nose-ring can degrade performance. Ears are more difficult to
locate as Viola-Jones' method is better at detecting “blocky” features rather than
outline features. Ears can also be occluded by hair, earrings or self-occluded by
a rotation of the face. Nonetheless, it was observed that the difference in tangent
angles to the ear-tips was one of the most discriminant features in the feature vector.
In profile view, the key features were the location of the iris and the shape of the
nose profile. The iris may not be available if the person is wearing glasses, as the stem
of the glasses usually occludes the eye. It is easy to detect the ear in profile view (if
it is not occluded) as it becomes a “blocky” feature when viewed from the side.