Real-Time Face Recognition from Surveillance Video - Intelligent Video Event Analysis and Understanding - page 190

In a real-world system, it would be necessary to select more features than we
have outlined here to provide sufficient discrimination between people enrolled on
the system. Suitable features could include corners and edges on facial features such
as the eyebrows and mouth. We have already mentioned that the eyebrows are one
of the most important features in human facial recognition [29]. Features in the non-
deformable part of the face should be accorded more weight, to provide robustness
to changes of expression. We have also mentioned that the full set of features may
not be available in every case. It may be necessary to select a different set of features
for different subjects. Missing-feature theory could be applied to this problem [17].
It is unknown how many features are required to provide discrimination over a
larger population. This could be the subject of a future study. We envisage that this
type of system would be used in a controlled environment (such as a Secure Corridor
in an airport) where the population is of limited size and people can be enrolled onto
the system as they enter the Corridor.
It is likely that the best results will be achieved by combining this approach with
other methods. Colour information was used in detection but not as part of the
recognition system. [29] states that pigmentation is an important recognition cue
for humans (at least as important as shape), so it would be natural to include skin
pigmentation, hair colour or eye colour as part of the feature vector.
Our shape-based approach could also be combined with appearance-based ap-
proaches in a hybrid detection system. Just like Bertillon in the 19th century, local
feature recognition could be used to narrow the search space and then global fea-
tures could be used to get an exact match. It is also possible to use appearance-based
methods on local features (Eigeneyes, etc. ). In this case, our method could be used
to perform pose estimation. [19] discusses how to improve face recognition using a
combination of global and local features.
The accuracy of facial recognition can also be improved by combining face de-
tection with other forms of detection using multi-modal fusion. Gait analysis would
be a likely candidate as it does not require any additional sensors.
Finally, it is worth mentioning that this approach to machine vision also has non-
security applications, for example the recognition of human faces by a robot or
computer game.
References
1. Alex, M., Vasilescu, O., Terzopoulos, D.: Multilinear analysis of image ensembles: Ten-
sorFaces. In: Heyden, A., Sparr, G., Nielsen, M., Johansen, P. (eds.) ECCV 2002. LNCS,
vol. 2350, pp. 447-460. Springer, Heidelberg (2002)
2. Bartlett, M., Movellan, J., Sejnowski, T.: Face recognition by Independent Com-
ponent Analysis. IEEE Transactions on Neural Networks 13(6), 1450-1464 (2002),
doi:10.1109/TNN.2002.804287
3. Belhumeur, P., Hespanha, J., Kriegman, D.: Eigenfaces vs. Fisherfaces: Recognition us-
ing class specific linear projection. IEEE Transactions on Pattern Analysis and Machine
Intelligence 19(7), 711-720 (1997)