A Novel Dimensionality Reduction Method Based on Subspace Learning for 3D Human Motion Data

Original 3D motion sequences lie in high dimensional subspace and on a high-dimensional manifold which is highly contorted, so it is difficult to cluster the similar poses together to form distinct movements. Here we use a non-linear learning dimensionality reduction technique (ISOMAP) based on radius bias function (RBF) generalized to map original motion sequences into low dimensional subspace. Experimental results show that motion intrinsic structures are discovered by this method in low dimensional subspace.