Real-Time Face Recognition from Surveillance Video - Intelligent Video Event Analysis and Understanding - page 187

Fig. 29 Feature Extraction: Selection of best local features from within the face bounding
box

Ta b l e 1 Effect of coefficient for facial proportions on detection accuracy. The Adrian video
had 342 frames and the Darryl video had 240 frames.

Subject Coefficient No. of faces detected

316 15 Adrian 0.48

0.5 280

0.53 205

0.55 128

0.6 21

Darryl 0.48 170

0.5 170

0.53 170

0.55 168

0.6 117

5.4 Feature Vector and Recognition System

Face recognition in the back-end was implemented using the O PEN CV SVM class.
A set of training videos was processed by the front-end. Feature points from the
training set were sent to the back-end, which generated feature vectors and dumped

15

316 detections includes some false positives — in some cases, eyebrows are detected as
eyes.