Multi-cue exemplar-based nonparametric model for gesture recognition

This paper presents an approach for a multi-cue, view-based recognition of gestures. We describe an exemplar-
based technique that combines two different forms of exem-
plars - shape exemplars and motion exemplars - in a uni-
fied probabilistic framework. Each gesture is represented
as a sequence of learned body poses as well as a sequence
of learned motion parameters. The shape exemplars are
comprised of pose contours, and the motion exemplars are
represented as affine motion parameters extracted using a
robust estimation approach. The probabilistic framework
learns by employing a nonparametric estimation technique
to model the exemplar distributions. It imposes temporal
constraints between different exemplars through a learned
Hidden Markov Model (HMM) for each gesture. We use
the proposed multi-cue approach to recognize a set of four-
teen gestures and contrast it against a shape only, single-
cue based system.