Indexing, Object Segmentation, and Event Detection in News and Sports Videos - Multimedia Database Retrieval: Technology and Applications - page 188

Fig. 7.16 Segmentation of Tennis sequence, ( a ) Frame no. 1 with the user mark, ( b )the
segmentation result at Frame no. 10
variations of the face distribution to be highly nonlinear and complex. The detection
rate usually drops quickly under this condition.
In this section, a robust and effective method is presented for detecting faces
in video sequences. The key step and the main contribution of this work is the
incorporation of a normalization technique based on normalized local histograms
with an optimal adaptive correlation (OAC) technique. This alleviates the common
problem of illumination.
Face detection techniques have been studied extensively, which include fea-
ture based methods using geometric information, such as skin color, geometric
shapes, motion information, and machine-learning based approaches, such as neural
networks, Gaussian mixtures, Support Vector Machines and statistical modeling
[ 204 - 206 ]. In automatic face detection algorithms, at the initial step, a pyramid
of downscaled copies of the given input image is produced. Then, a sliding window
scans each of the downscaled images and finally a classifier is applied on all possible
window locations to decide whether the region contains a face object or not.
Practically, the number of windows or equivalently the number of times the clas-
sification will be processed is typically tens of thousands depending on the image
size and demagnification factor. AdaBoost algorithm [ 207 ] employs this method in
a fast way and has been widely investigated in video face detection systems. The
key point is that fast, but less discriminating classifiers can reliably reject most of
the windows containing non-face objects while passing the windows containing the
maybe-face objects to a second level classifier, which is slower than the previous
one but has higher discriminating power. This procedure iteratively continues and
can provide high detection performance with much less computational expense.
Since illumination is one of the most important factors that determine success
or failure in face detection, many approaches have been proposed to handle
the illumination problem. Most algorithms for face detection presume that the
illumination variation is uniform or lighting must be controlled. Georghiades