Human motion interpolation. New Method.

Abstract

This thesis shows a new method to perform video interpolation between two di erent human motions in a natural way. The main objective of the method is to provide a natural interpolation in terms of speed, similarity and geometry. To achieve it we use Local Linear embedding (LLE), performing dimensionality reduction to nd in the low dimensional space the transition frames and the number of frames to be interpolated. Then we apply K-means clustering in a 9 dimensional space to distinguish between corresponding body parts in the two transition frames. Finally we will interpolate the corresponding regions with di erent morphing methods. The method show great results when applied in motions that go through similar poses. The thesis also provides a robust method to perform foreground extraction.