一种基于DenseNet与WGAN-GP的运动迁移方法

A Motion Migration Method Based on DenseNet and WGAN-GP

针对目前人体骨骼模型运动迁移算法计算精确度不高,提出了一种基于DenseNet的骨骼卷积网络与WGAN-GP模型的运动迁移方法。通过对源与目标骨骼模型分别提取静态特征,并对源骨骼静态特征与源运动序列使用动态编码器提取源动态特征,从而能够对目标骨骼静态特征与源动态特征使用解码器生成目标运动序列,完成运动迁移。在网络模型训练时,同时引入了WGAN-GP网络模型机制对生成序列和源运动序列的动、静态特征误差进行约束。实验结果表明:该方法的运动迁移模型各关节点相对于单位身高的运动误差较小,能够生成较好保留源动态特征的目标运动序列。

To address the problems that the current motion migration algorithm of human skeletal model is not computationally accurate,a motion migration method based on DenseNet skeletal convolutional network with WGAN-GP model was proposed.By extracting static features from the source and target bone models separately,and using a dynamic encoder to extract source dynamic features from the source bone static features and source motion sequences,a decoder can be used to generate target motion sequences from the target bone static features and source dynamic features,completing motion transfer.During the network model training,the WGAN-GP network model mechanism was introduced simultaneously to constrain the dynamic and static feature errors of the generated sequences and the source motion sequences.The experimental results show that the motion migration model of this method has a small motion error at each joint point relative to the unit height and can generate the target motion sequence with better preservation of the source dynamic features.