Due to the high-dimensionality of motion captured data which resulted in the complexity in motion analysis, a method of motion data processing based on manifold learning was proposed. Isomap, a classical manifold learning algorithm, was necessary to be improved and extended in this paper. A framework of motion data processing based on manifold learning was built to embed high-dimensionality data into low-dimensionality space. It simplified the motion analysis, and in the same time preserved the original motion features. In order to solve the inefficiency of processing large-scale motion data, Sample Isomap (S-Isomap) algorithm was proposed. Experiments proved that approximate embeddings of motion data computed by S-Isomap were average 10 times faster than by Isomap, while 10% frame samples were selected.