摘要
对运动数据集进行可视化是实现目标运动资源搜索的有效手段,文中提出了一种基于双层自编码的运动数据集可视化方法。首先在姿态层采用变分自编码器获得每个三维人体姿态在二维散点图中的坐标并使得相似的姿态彼此相邻,从而能够支持用户对目标运动资源的快速定位。其次在运动层采用双向长短时记忆递归网络对姿态的上下文运动进行自编码,获得的低维运动表示能够有效地支持用户在线运动分析。在CMU基准运动数据集上的实验结果验证了本文方法的有效性。
The motion data set visualization is an effective means for target motion search.A method for the motion data set visualization based on two-layer auto-encoders is presented.Firstly,a variational auto-encoder is adopted at the pose layer to obtain the coordination of each 3D human pose in the scatterplot and make similar poses adjacent to each other,supporting rapid position of the target motion resource for uses.Secondly,BiLSTM recurrent network is used to encode the contextual motion of each pose at the motion layer,and the obtained latent motion representation can effectively support online motion analysis on users.Experimental results on CMU motion library verify the effectiveness of the method.
作者
陈松乐
孙知信
CHEN Songle;SUN Zhixin(Engineering Research Center of Post Big Data Technology and Application of Jiangsu Province,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;Research and Development Center of Post Industry Technology of the State Posts Bureau(Internet of Things Technology),Nanjing University of Posts and Telecommunications,Nanjing 210003,China)
出处
《南京邮电大学学报（自然科学版）》
北大核心
2020年第3期22-30,共9页
Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition
基金
国家自然基金(61672299,61972208,61702281)
江苏省高校自然基金(18BC051)资助项目。
关键词
运动数据
可视化
自编码器
motion data
visualization
auto-encoders
