结合深度自编码和时空特征约束的运动风格转移方法  

Motion Style Transfer via Deep Autoencoder and Spatio-Temporal Feature Constraint

作　　者：胡东 彭淑娟[1,2] 柳欣 杜吉祥[1,2] Hu Dong;Peng Shujuan;Liu Xin;Du Jixiang(College of Computer Science and Technology,Huaqiao University,Xiamen 361021;Key Laboratory of Pattern Recognition and Computer Vision,Xiamen City,Xiamen 361021)

机构地区：[1]华侨大学计算机科学与技术学院,厦门361021 [2]厦门市模式识别与计算机视觉重点实验室,厦门361021

出　　处：《计算机辅助设计与图形学学报》2018年第5期946-956,共11页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61673185;61673186);福建省自然科学基金(2017J01112);华侨大学科研创新能力培养资助项目(1611414006)

摘　　要：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.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 motion.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 semantics can be well transferred from one style to another one.The experiments have shown that the proposed 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 generalization ability and adaptability of our proposed model in comparison with existing counterparts.

关 键 词：深度自编码 运动风格转移 时空特征约束 格拉姆矩阵 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]