A Segmentation Approach for Interaction Analysis Using Non-Parametric Bayesian Methods

摘要

We propose a segmentation method for multiple related time series data using non-parametric Bayesian methods. This method assumes that human motion follows a double articulation structure, where low-level dynamical behaviors are discovered using beta process - autoregressive hidden Markov model (BP-AR-HMM), and high-level semantic behaviors are formed from these low-level behaviors using nested Pitman-Yor language model (NPYLM). Specifically, this work explores how the choice of autoregressive lag order affects the proposed segmentation method.