Incremental learning of full body motions via adaptive Factorial Hidden Markov Models

This paper describes a novel approach for incremental learning of motion pattern primitives through long-term observation of human motion. Human motion patterns are abstracted into a factorial hidden Markov model representation, which can be used for both subsequent motion recognition and generation. The model size is adaptable based on the discrimination requirements in the associated region of the current knowledge base. As new motion patterns are observed, they are incrementally grouped together based on their relative distance in the model space. A new algorithm for sequentially training the Markov chains is developed, to reduce the computation cost during model adaptation. The resulting representation of the knowledge domain is a tree structure, with specialized motions at the tree leaves, and generalized motions closer to the root. Tests with motion capture data for a variety of motion primitives demonstrate the efficacy of the algorithm.