Real-Time Face Recognition from Surveillance Video - Intelligent Video Event Analysis and Understanding - page 159

This has the effect of projecting the high dimensionality of the face image space
into a feature space of lower dimensionality. Features are classified on a nearest-
neighbour basis. The technique can be applied to local features as well as whole
faces (in this case the Principal Components are referred to as Eigeneyes, Eigen-
nose, etc. ) Eigenfaces produces very reliable results under laboratory conditions,
but is not robust to changes in pose, illumination and expression. It also has high
computational requirements, making it unsuitable for real-time facial recognition
from a surveillance video stream.
Turk discusses improvements to and extensions of his Eigenfaces technique
in [34]. PCA builds its model by minimising the pairwise relationships between
pixels in the image training set. The approach can be generalised to minimise the
second-order and higher-order dependencies in the image space, using Independent
Component Analysis (ICA). ICA attempts to find the basis along which the data are
statistically independent. It was first applied to face recognition in [2], where ICA
was shown to be more robust than Eigenfaces to minor changes in illumination,
expression and appearance (hair, make-up).
Another well-known appearance-based approach is Fisherfaces [3]. The Fish-
erface approach attempts to improve on Eigenfaces by creating a model which is
invariant to changes in illumination and expression (but not pose). Fisherfaces uses
Linear Discriminant Analysis (LDA), which is similar to PCA in that the high di-
mensionality of the image space is reduced to a lower-dimensional feature space.
Fisherfaces attempts to choose the direction of the projection such that variations
in lighting and facial expression are projected away but the features used for recog-
nition are clustered. PCA chooses the projection which maximises the total scatter,
preserving unwanted variations in lighting and facial expression. The computational
requirements of the two approaches are similar. Fisherfaces does not solve the prob-
lem of variations in pose. The image in the test set must be similar to the image
in the training set to be detected. The test set must contain a comprehensive set of
images under different lighting conditions.
PCA, LDA and ICA are compared in [28]. Another similar approach which de-
serves a mention is Tensorfaces [1]. Tensorfaces uses multi-linear analysis, a varia-
tion of PCA, to combine training sets of face images under different poses, expres-
sions and lighting.
All the holistic approaches require a sizeable training database of face images
under different poses, expressions and lighting to provide accurate results. For a
surveillance application, it is unlikely that an accurate model of all possible fa-
cial variations can be constructed in advance. The computational requirements are
also considerable. These challenges mean that holistic, appearance-based methods
have not been demonstrated to work accurately in real-world situations. Turk con-
cludes [34] that Eigenface or other appearance-based approaches must be combined
with feature- or shape-based approaches to recognition to achieve robust systems
that will work in real-world environments.
Structural ,or feature-based , approaches to face recognition begin by locating
local features such as the eyes, nose and mouth. The locations and characteristics
of the features are then used to classify the face. As discussed in [35] and [34],