Ensemble HMM Learning for Motion Retrieval with Non-linear PCA Dimensionality Reduction

Abstract

As commercial motion capture systems are widely used , more and more 3D motion database become available. In this paper, we presented a motion retrieval system based on ensemble HMM learning. First, 3D features are extracted. Due to high dimensionality of motion's features, then non-linear PCA and radial basis function (RBF) neural network for dimensionality reduction are used. At last each action class is learned with one HMM for motion analysis. Since ensemble learning can effectively enhance supervised learners, ensembles of weak HMM learners are built. Some experimental examples are given to demonstrate the effectiveness and efficiency of our methods.