Real-Time Face Recognition from Surveillance Video - Intelligent Video Event Analysis and Understanding - page 178

4.2 Feature Vector and Matching

The set of reference points and feature key points is transmitted to the back-end over
TCP/IP. The back-end calculates the feature vector from this set of points following
this algorithm:

1. Calculate the gradient of the baseline (the line connecting the irises)
2. Use the gradient to calculate a rotation matrix to normalise the coordinate space,
so that the baseline is parallel to the horizontal axis. This ensures that the feature
vector is invariant to rotations in-plane
3. Rotate all the points about the left iris
4. Find the face centre line by taking the midpoint between the nostrils
5. Find the intersection of the face centre line and the baseline
6. Normalise the coordinate space so that the intersection of the baseline and face
centre line is . A graphical representation of the normalised feature points

is shown in figure 19.
7. Calculate a set of lengths connecting the intersection point and the irises to the
key feature points. These lengths are expressed as ratios of the baseline. This
ensures invariance to scale
8. Calculate a set of tangents of angles between the intersection point and the key
feature points. The feature vector stores only differences between these tangents.
This ensures invariance to rotations out-of-plane 11

( 0 , 0 )

Fig. 19 Normalised feature points for four feature key points (two nostrils and two ear tips).
The baseline is the horizontal line at the bottom and represents the line that joins the irises.

11

See [33] for proof.