提出了基于三维人体运动数据和等距特征映射(ISOMAP)降维机制的子空间人体运动风格生成和编辑的方法.该方法拓展了传统ISOMAP难以处理非训练(outofsample)数据的局限性,在具有非线性内在属性的高维运动数据空间和低维非线性风格化子空间之间建立映射,使之能够直接应用在非训练数据集.对于映射到低维子空间的运动数据利用分解生成模型分离出运动的内容参数和风格参数,通过在子空间中调整这两种参数并逆向映射到原始运动数据空间,实现运动数据的编辑和新风格的生成.实验结果表明,该方法能够在虚拟现实场景中自动生成各种复杂新风格的人体运动,并具有精度高的特点.
A new framework for automatic synthesis and editing of human motion style based on 3D human motion data and isometric feature mapping (ISOMAP) dimension reduction was proposed. In this frame- work, the generalized ISOMAP was extended to process out of-samples data by building an optimal map- ping function from the input high dimensional space to the embedding low dimensional nonlinear stylized space. The decomposable generative model was used to learn separate style parameters and content param- eters of human motions. New motions with new style could be edited and reconstructed by adjusting and mapping these parameters from subspace to original motion data space. The experimental results show that the proposed method can generate new complex motion styles automatically and accurately in virtual reality scene.