Synthesizing Human Motions

Abstract

In this work different data-driven methods for the synthesis of natural human full body motions are presented. My research in this area was based on the following fundamental questions: Suppose we have all the motion capture data ever recorded, how could we use them? What benefits do they offer us? Which applications can arise?<br /> In fact, most of the motion capture data recorded are used for one specific project only and never reused although all these motion data contain valuable information about how human motions look like. To be able to handle a large amount of motion capture data I developed two basic techniques: A method for fast similarity search of single poses and motion sequences and a method for automatic annotation of motion capture data. Based on these two basic techniques three different methods of motion synthesis have been developed.<br /> In the first approach, tensor based multilinear representations are constructed from annotated motion sequences. As will be shown this representation is especially suitable for motion synthesis.<br /> In the second approach, given motion sequences are enhanced with respect to missing degrees of freedom using a technique for motion texturing. Here, similar motions are retrieved efficiently from the database, using a novel technique for fast similarity search. This fast motion retrieval was identified as the essential step to use the database as prior-knowledge to drive the synthesis process.<br /> Finally a technique for motion synthesis from sparse key frames is introduced. Employing the search algorithm again, a so called motion graph, a structure for motion synthesis is computed on the fly. The result of this synthesis is then refined by the motion texturing approach.<br /> All techniques and algorithms are tested and evaluated on the two largest freely available motion capture databases.