@TechReport {Starner/Pentland:1995,
  author=	{ Thad Starner and Alex Pentland},
  year=		{ 1995},
  keywords=	{ ASL HMM},
  institution=  { MIT - The Media Laboratory},
  title=	{ Real-Time American Sign Language Recognition from Video
  	Using Hidden Markov Models},
  journal=	{ MIT Media Laboratory Perceptual Computing Section
  	Technical Report},
  volume=	{ 375},
  annote=	{
 This paper experimentally demonstrates a real time Hidden Markov Model based
 system for recognizing American Sign Language(ASL) without explicitly modelling
 fingers.It uses a 40 word lexicon.The system uses a color camera to track hand
 in real time.The hands are tracked using their color - first using colored 
 gloves and then by natural skin color.The tracking does not attempt a fine
 description of human hand, instead it only produces a coarse description of
 hand shape,orientation and trajectory.In the tracking process,the subject is 
 made to wear gloves of fixed color (different for both hands) and the images
 are scanned for a pixel of appropriate color.After this pixel is found,it is
 used as a seed and the region is grown by checking eight neighbouring pixels
 for appropriate color.Each pixel checked is considered a part of the hand.
 This in effect helps to prevent edge and lighting aberrations.The centriod
 is stored as seed for next frame.Second Moment analysis is performed on the
 resulting bitmap and the seed.The tracking for natural skin color is done in
 almost same way using the fact that all human hands have approximately the
 same hue and saturation.In both cases the obtained information is input
 to a HMM for recognition of signed words.The HMM in this case is assumed to
 be a first order Markov process.Experimental verification shows the error 
 rates to be low.

 This work is very useful because ASL is the language of choice for most
 deaf people in United States.Moreover changing background and occlusion is
 not a problem for the tracking method used.    -Soumyadeep Paul 2/98},
}

@Article{Rehg/Kanade:1994,
  author=	{ Rehg,James M. and Takeo Kanade},
  year=		{ 1994},
  keywords=	{ HUMAN MOTION NONRIGID GESTURE RECOGNITION VISUAL TRACKING
    MODEL-BASED VISION},
  institution=	{ CMU - CS},
  title=	{ DigitEyes: Vision-Based Human Hand Tracking},
  journal=	{ Proceedings of European Conference on Computer Vision},
  month= 	{ May},
  annote=	{
 This paper describes a model-based hand tracking system, called DigitEyes
 that can recover the state of a 27 DOF hand model from gray scale images at
 speeds of up to 10 Hz.The Denavit-Hartenburg representation is used in which
 each finger link has an attached link coordinate system.The transformation of
 these frames is used to model the kinematics of the hand.The hand is modelled
 as a collection of 16 rigid bodies - each finger having 3 individual finger 
 links and a palm.The four fingers are assumed to have planar motions only 
 with 4 degrees of freedom.The thumb is modelled with a 5 DOF model.Each finger
 has a rigid anchor point with which it is attached to the palm.Finger links
 are modelled as cylinders and the tips as hemispheres.Tracking is done using
 local image-based trackers.The output of the trackers are hand features which
 is input to the state estimation algorithm.In the state estimation algorithm
 the current state is used to predict the feature locations in the next frame.
 The state estimation algorithm uses a residual vector which is the difference 
 between the measured and predicted states.The state correction is obtained
 by minimizing the magnitude of residual vector squared.If the magnitude of
 residual vector is zero then the hand model is perfectly aligned with image
 data.Modified Gauss-Newton algorithm is used to solve the problem of minimizing
 the residual.The paper also extends the work to multiple camera case.
 The above system is tested on a 3D graphical mouse with a single camera.It
 is also tested on whole hand tracking with two cameras.It performs good in
 both the cases.

 The paper is a very comprehensive one and since the system extracts the 
 features from unmarked and unadorned hands, it is convenient too.The paper
 made a very good reading.   - Soumyadeep Paul 2/98},
} 

@InProceedings {OHYA/EBIHARA_et_al:1996,
  author=	{ Jun OHYA and Kazuyuki EBIHARA and Jun KURUMISAWA 
                  and Ryohei NAKATSU},
  year=		{ 1996},
  keywords=	{ VIRTUAL KABUKI THEATER POSTURE ESTIMATION},
  institution=	{ ATR - MIC Research Lab.},
  title=	{ Virtual Kabuki Theater: Towards the Realisation of
                  Human Metamorphosis Systems},
  journal=	{ IEEE International Workshop on Robot and Human Communication},  annote=	{
 This paper describes the Virtual Kabuki Theater the authors have recently
 developed. In such a theater , people at different locations can perform as 
 Kabuki actors in a Kabuki environment. The system used consist of three 
 different modules.In the first module 3D models of Kabuki actors are created
 in advance. The model is created by a wireframe model in which color texture
 is mapped.This model could be deformed according to facial expression data
 obtained.In the second module,facial expressions and body motions are detected 
 in real time using an infrared camera.Facial expression is detected using 
 a camera fit to the helmet.In the third module, these movements are reproduced
 in Kabuki actor's model by deforming the wireframe.The human posture is 
 obtained by first detecting the significant points such as tips of hands and 
 top of the head.The positions of elbow and knee are now expressed as a linear 
 combination of significant points and the coefficients are found using GA.

 Though the method followed is good,but the authors have neglected the 
 movement of the fingers.A good reading.  -Soumyadeep Paul 4/96},
}