Real-Time Face Recognition from Surveillance Video - Intelligent Video Event Analysis and Understanding - page 169

estimate the most likely eyes by considering their location and size relative to each
other and the face bounding box.
The reference points are located at the centre of the iris/pupil within the two eye
bounding boxes. There are a number of possible approaches to iris detection. If suffi-
cient training data were available, it would be possible to add a third nested Haar cas-
cade to the O PEN CV face detector to detect irises within the eye region. [6] proposes
an integrodifferential operator which can detect iris and pupil boundaries. Another
possibility would be to match curves to the eye region using a Hough Transform.
Circles of approximately the right size could be extracted as irises and pupils. [14]
takes this approach, improving robustness by using a separability filter.
Our approach was a geometric approach using thresholding and blob detection
on the image. This has the advantage of being simple and fast to compute.
Detection uses the following algorithm 7 :

5 averaging filter to remove noise and improve blob detection 8 .
2. As the surface of the eye is highly reflective, there is usually a point of light on
or near the pupil/iris. To locate this point, a gradient is calculated over the image.
Small gradients are discarded. Large gradients are weighted by the distance from
the centre of the bounding box. The highest weighted gradient is taken as a point
assumed to be on or close to the pupil/iris (see figure 7).
3. Histogram equalise to improve blob detection.
4. Threshold image at 0

1. Use a 5 ×

. 25 × max ( image ) .

5. Erode thresholded image with a 3 3 structuring element (see figure 8)

6. Find the blob closest to the gradient point.
7. Use region selection to find the maximum and minimum points within the iris
blob.
8. Calculate the centre of the blob as the average of the maxima/minima. (figure 9)
9. return centre of blob

×

Note that this method works reasonably well on people who are wearing glasses.
[3] solves the problem of wearing glasses by dividing images of an individual into
two classes, one which is “wearing glasses” and one “not wearing glasses”. The
O PEN CV face detector includes eyes wearing glasses in its cascades for face and
eye detection.
Once the two reference points have been determined, the baseline is calculated
as the line joining the two points (see figure 6).

3.3.3 Nostril Detection

Having detected the two reference points, the nostrils are detected next, as the
first feature key points. The nostrils are important, as they are rarely occluded and

7

Detection algorithms were developed experimentally in M ATLAB (see detect iris.m )
and subsequently implemented in C++/O PEN CV.

8

Later experimentation with a larger test set showed that the averaging filter did not improve
detection accuracy, so this step was removed from the C++ implementation.