Human motion primitive discovery and recognition.

Abstract

We present a novel framework for the automatic discovery and recognition of
\nhuman motion primitives from motion capture data. Human motion primitives are
\ndiscovered by optimizing the 'motion flux', a quantity which depends on the
\nmotion of a group of skeletal joints. Models of each primitive category are
\ncomputed via non-parametric Bayes methods and recognition is performed based on
\ntheir geometric properties. A normalization of the primitives is proposed in
\norder to make them invariant with respect to anatomical variations and data
\nsampling rate. Using our framework we build a publicly available dataset of
\nhuman motion primitives based on motion capture sequences taken from well-known
\ndatasets. We expect that our framework, by providing an objective way for
\ndiscovering and categorizing human motion, will be a useful tool in numerous
\nresearch fields related to Robotics including human inspired motion generation,
\nlearning by demonstration, and intuitive human-robot interaction.