Motion style transfer associated with the motion editing techniques can be well utilized to meet the requirements of different motion processing purpose in computer animation. Often, the data driven based motion style processing may suffer from the unnatural posture and lack of adaptability with continuous mo-tion. To tackle these problems, we present an efficient motion style transferring approach by using deep autoencoder and spatio-temporal feature constraint. According to the user request for style transferring, we propose to decompose the human motion into behavior motion and style motion. By embedding the history motion frames within the deep autocoder model, the discriminative spatio-temporal features corresponding to the behavior motion and style motion can be well extracted. Finally, we exploit the style constraints in feature space to control the motion style transfer by Gram matrix, whereby the motion style of different se-mantics can be well transferred from one style to another one. The experiments have shown that the pro-posed approach can produce a variety of different movement styles, and the transferred motion styles are visually natural and vivid. Meanwhile, the related experiments have also demonstrated the good generaliza-tion ability and adaptability of our proposed model in comparison with existing counterparts.%运动风格化处理综合运用了运动编辑技术, 可以较好地满足不同动画师对运动风格处理的要求. 针对运动数据驱动方式的风格化处理存在未考虑运动时序性而导致姿态不协调和运动不自然等问题, 提出一种结合深度自编码和时空特征约束的自动风格转移模型. 首先根据用户对风格转移的要求把运动分解为行为运动和风格运动; 其次通过在深度自编码网络模型中增加历史运动序列关联信息建立风格特征提取模型, 分别对行为运动和风格运动的时空特征进行提取; 最后使用格拉姆矩阵对提取的特征建立运动风格转移约束, 自动实现把一个运动的风格特征转移到另一个运动的行为内容上. 实验结果表明, 文中方法能够实现多种不同种类运动风格转移, 并且转移后生成的运动序列风格自然和真实; 此外, 相关实验证实了文中提出的模型具有较强的泛化能力和自适应性.
展开▼
