Real-Time Face Recognition from Surveillance Video - Intelligent Video Event Analysis and Understanding - page 177

examples, so that there are enough false positives detected by the early stages of
the cascade that can be passed on to train the later stages of the cascade. The data
should be well-separated (for example, not mixing training sets in different poses)
and well-segmented (box boundaries should be consistent across the training set).
These considerations mean that retraining the classifier is a formidable problem.
Note: Feature extraction from a profile image was demonstrated in principle us-
ing M ATLAB code, but due to the issues discussed in this section, we did not develop
a full implementation in O PEN CV.

4Back-End

The feature set extracted for each face is very small — simply an coordinate

for each reference point and feature key point. These points are labelled and sent
to the back-end for analysis (inside a TCP/IP packet). Labels could include camera
number, time-stamp, frame number, face bounding box within frame and an object
tracking ID (if face tracking using particle filtering has been implemented).
In the back-end, a feature vector is calculated from the set of feature points, and
this feature vector is compared to the database of enrolled people using a SVM
(Support Vector Machine).

( x , y )

4.1 Feature Vector

To ensure robustness against changes in scale and rotation, only ratios of lengths
and angles are stored in the feature vector. First, the distances between the reference
points and the other key points are calculated, and expressed as a ratio to the refer-
ence length. As all distances are expressed as ratios, they are invariant to changes in
scale.
Second, the angles between the reference points and other key points are calcu-
lated. The orientation of the baseline is used as the horizontal axis for the coordinate
system of the feature vector. i.e. , the coordinates of the key points must undergo a
matrix transformation into the new coordinate system. Normalising the key points
to the baseline means that the feature vector is invariant to rotations in-plane.
The calculated angles do not themselves form part of the feature vector. Rather,
the feature vector stores the difference between tangents of the angles formed at
each reference point. Taking the difference between tangents means that the feature
vector is invariant to rotations out-of-plane. A proof of this result can be found
in [33].
The feature vector for the profile view is calculated in the same manner as the
frontal view. The key point coordinates are transformed into a coordinate system
using the profile baseline as the horizontal axis. The lengths and angles between
the reference points and the other key points are calculated. These are encoded in
the feature vector as ratios of the profile reference length and difference between
tangents of angles. As with the frontal model, the feature vector is robust to changes
in scale and pose (rotation).