Real-Time Face Recognition from Surveillance Video - Intelligent Video Event Analysis and Understanding - page 185

Several general rules were set when filtering the features:

• The right eye should be in upper right part of the face

•

The left eye should be in the upper left part of the face

•

The nose should be in the central region of the face. Sometimes the bounding
boxes for the nose and the bounding boxes for the eyes can briefly overlap.

The proportions of the human face have been studied for many years, in particular
by surgeons specialised in reconstructive and plastic surgery. Thus, medicine has
determined that the height of the nose relative to the frontal face should be around
47% [23]. The frontal face Haar cascade often crops the lower part of the face (see
figures 27 and 28). This cropping is very unpredictable, so the face height is unreli-
able for filtering features. However, the nose is never cropped. Since the nose height
is not affected by cropping, we tried to use it in defining a ratio that would give the
approximate position of the top part of the eyes.
The human frontal face can be divided into three horizontal regions. The aesthetic
ideal is that the three regions are the same height, but in our system we allow some
flexibility. Since we are interested in determining a relation that would give us the
position of the eyes relative to the top of the face, we are only interested in the
upper region. This lies between the trichion 13 and glabella 14 . This is approximately
33% of the entire facial height but since the Haar detection of the frontal face crops
this part, we calculate this region as being between 20-30% of the face height. The
correct position of the eyes should be detected just under the glabella. So the position
of the eyes on y axis should be approximately 0.2-0.3

× face height (1) and the

height of the nose should be approximately 0 face height (2). By substituting

face height from (2) into (1) it follows that the eyes are expected at approximately
0.42-0.63

. 47 ×

nose height .
We conducted experiments to determine the coefficient that works best within
this interval. Table 1 shows the number of faces (containing two eyes and a nose)
found in the video stream for different values of the coefficient.
The value 0.53 was chosen for the facial proportion coefficient, because it pro-
vided good detection rates and no false positives. 0.5 provides even better detection
rates but was considered too close to the value that outputs false positives.
In our test system, the position of the top of the nose bounding box was approx-
imated using a fixed number of pixels. The detection was calibrated for the size
of the faces in the test videos. In a real system, this parameter would have to be
automatically calibrated to work on a wider range of images.
The proportions that were used to choose the best nose and best eyes from all
candidates are shown in figure 29.

×

13

The point where the hairline meets the midpoint of the forehead.

14

Reference point in anthropology representing a smooth elevation in the frontal bone just
above the bridge of the nose.