Real-Time Face Recognition from Surveillance Video - Intelligent Video Event Analysis and Understanding - page 188

Fig. 30 Discriminative capacity of the feature vector: comparison of two features for two
subjects
them into a file. A matrix was initialised with each feature vector as a row and we
specified how many rows represented training data for each subject. The result of
training was saved in an XML database.
The SVM was trained for two subjects using one video for each subject. In the
training videos, the subjects were looking straight ahead. The video for Adrian was
13 seconds long and 67 faces were extracted and passed to the SVM for training.
The video for Darryl was 9 seconds long, and 83 faces were extracted. The number
of frames extracted for Adrian is lower because in many frames his eyes are closed,
so those frames are discarded. Also, Adrian's nose ring made accurate feature ex-
traction more challenging for the front-end! The feature vector in our experimental
system is of low dimensionality, with only six features. Figure 30 illustrates the dis-
criminative capacity of our system, showing a comparison of features 3 and 6 for the
two subjects. Features 1, 2, 4 and 5 measured ratios of distances between features
(relative to the reference length), whereas features 3 and 6 measured the differences
of tangents between the reference points and feature key points. Features 3 and 6
were therefore the most robust to rotations out-of-plane and provided the greatest
discrimination between the subjects in our experiments.
We demonstrated that it is possible to calculate more lengths and tangents be-
tween the feature points we have already extracted. A production system would add
additional feature points to provide a feature vector of larger dimensionality and
therefore capable of greater discrimination.
Testing was conducted using a different video from the training set for each sub-
ject. The video for Darryl was 9 seconds long and he was rotated 10% from frontal
view. The video for Adrian was 11 seconds long and was facing straight ahead. The