Real-Time Face Recognition from Surveillance Video - Intelligent Video Event Analysis and Understanding - page 181

Fig. 22 Affect of Skin Detection on Performance: Adrian

The strategy of skin detection would not be of any benefit if large areas of the
background were skin-coloured. In a real system, whether or not to use skin detec-
tion could be configured at each sensor.
For our experimental system, we chose a window of 40 pixels around skin regions
as giving the best compromise between performance and accuracy.

5.2 Face Detection

The most important features for the O PEN CV face detector are the eyes and nose.
During informal experiments, it was observed that occluding the eyes or nose pre-
vents a face from being detected. The mouth is a less important feature. It is often
possible to occlude the mouth without affecting face detection. In any case our sys-
tem does not extract features from the mouth area as we restricted our system to
detecting features in the non-deformable part of the face.
As illustrated in figure 3, if we did not detect a suitable face and local features
within a frame, that frame would be discarded. Figures 23 and 24 show how many
face candidates are rejected at each stage of the detection process.
Note that the O PEN CV face detector returns many false positives. In our tests,
each frame of video contained exactly one face, but figure 23 shows that the number
of faces returned by the Haar cascade is greater than the number of frames. The
results show that skin detection slightly reduced the number of false positives (while
speeding up the overall detection rate).
More face candidates are rejected by the nested Haar cascades which detect local
features (eyes and nose). If the system cannot detect two eyes and a nose, the face
candidate is discarded. Finally, we identify reference points and feature key points