@CONFERENCE\{IMM2010-05899,
    author       = "R. Larsen and J. B. Nielsen",
    title        = "Face Detection and Recognition in Video-Streams",
    year         = "2010",
    month        = "may",
    pages        = "36",
    booktitle    = "Eighth French-Danish Workshop on Spatial Statistics and Image Analysis in Biology, {SSIAB,} Book of Abstracts",
    volume       = "",
    series       = "",
    editor       = "",
    publisher    = "",
    organization = "",
    address      = "",
    url          = "http://www2.imm.dtu.dk/dataanalyse/ssiab8/",
    abstract     = "In this work we demonstrate how the Viola-Jones face detector [1] can be combined with a
person specific active appearance model [3] and used for automated annotation of video
streams. Here we understand annotation as identification of the first frame and last frame of a
persons face continuous appearance in a video stream, as well as detection, segmentation, and
identification of the face in each frame. The Viola-Jones face detector is based on simple
sums and differences of image pixels within rectangular areas. A main advantage of these
features is that they can be computed in constant time using the so-called integral image in
which a pixel value is the sum of all pixel values above and to the left of that pixel in the
original image. A boosting classifier is then employed, where each of the weak classifiers in
the sequence of classifiers is constrained to be based on a single of the features described
above. In each step of the sequence of classifiers the number of false negatives is minimized.
In this way a cascade of classifiers is constructed in which at each stage if a sub-windowed is
rejected as not being a face no further processing is done, the positives and false positives are
send to the next stage. This make for a computationally very efficient face detector. When a
face is detected in a sequence of frames within certain motion limits a sub-sequence is defined.
From this sequence a series of frames uniformly sampled over time are selected for
segmentation using a person-specific active appearance model. The active appearance model
uses a face model based on a truncated principal component model of combined variation of
shape as defined by a set of and texture as defined by the sampled intensity values across the
face.
We demonstrate the face detection and recognition scheme on a series of Danmarks Radio
(Danish Broadcasting Corporation) game shows featuring actor and talk show host Jarl Friis
Mikkelsen. In a series of video streams face sub-sequences are successfully detected and
classified as being either the actor in question or not. The computational implementation is
based on the OpenCV implementation of the Viola-Jones face detector and the publically
available active appearance model software from {DTU} Informatics [3].
1. P. Viola and M. Jones. Robust real-time face detection. International Journal of
Computer Vision (IJCV) 57(2) 137-154, 2004.
2. M. B. Stegmann, B. K. Ersb{\o}ll, and R. Larsen. {FAME} – a flexible appearance
modelling environment. {IEEE} Trans. on Medical Imaging, 22(10):1319-1331, 2003
3. T.F.Cootes, G.J. Edwards and C.J.Taylor. {''}Active Appearance Models{'',} {IEEE} {PAMI,} 
Vol.23, No.{6,} pp.681-685, 2001"
}