Spatio-temporal LTSA and Its Application to Motion Decomposition

This paper describes a STLTSA-based framework to analyze and decompose human motion for synthesis. In this work, we mainly intend to extend a manifold learning method, local tangent space alignment, to a spatio–temporal version for manifold analysis and offer an effective method of estimating the intrinsic dimensionality of motion data. Based on an assumption that a long sequence of motion is composed of a number of short motion units, we can decompose a motion into several basic motion units in a low-dimensional manifold space and extract motion cycles from the cyclic unit. The generation of new complex movement using obtained motion units is feasible and promising.