期刊文献+

基于GAN的多维时间序列人体步态数据增强

Data Augmentation of Multi-Dimensional Human Gait Data Using Generative Adversarial Networks
下载PDF
导出
摘要 典型的外骨骼机器人控制策略通常需要大量的人体步态数据来计算控制器的参考输入;步态预测或分类任务也需要大量的步态数据用于模型训练.数据增强技术可以合成与真实数据分布相似又保持原始数据动态特性的人工数据,从而降低人工收集步态数据的成本,有助于提高控制器的精度或模型的泛化能力.本研究以健康受试者在光学运动捕捉平台采集的原始步态数据为基础,通过建立多维和一维生成对抗网络模型学习真实数据的潜在分布从而生成相似分布的合成数据以达到一倍至数十倍增强真实数据的效果.对增强后的数据采用T-SNE可视化、JS散度和MAE度量两种数据的相似度,相比于一维时间序列GAN模型,多维时间序列GAN模型无论是从数据分布相似还是统计相似上精度都更高. A typical exoskeleton-robot control strategy usually requires a large amount of human gait data for calculating the reference input of the controller.Also,gait prediction or classification tasks require enormous data for training.Data augmentation technologies can synthesize false data that maintain similar distribution and dynamic characteristics with the real data,thus reducing the cost of manually collecting gait data and improving the controllers'precision or models'generalization ability.Based on the collected human gait data from the subjects,this paper focuses on establishing a multi-dimensional time-series GAN(MTGAN)and a one-dimensional GAN(OTGAN)to learn the potential distribution of the real data and generate similarly distributed data in terms of a similar distribution and a larger amount.T-SNE visualization,JS dispersion,and MAE similarity measurements indicate the MTGAN model gives better performance on both distribution similarity and statistics similarity than the OTGAN model.
作者 李志康 廖伍代 汪鑫 王燕 王双红 LI Zhikang;LIAO Wudai;WANG Xin;WANG Yan;WANG Shuanghong(School of Electronic Information,Zhong Yuan University of Technology,Zhengzhou 450007,China)
出处 《赣南师范大学学报》 2021年第6期62-67,共6页 Journal of Gannan Normal University
基金 河南省科技攻关项目(202102210135) 国家重点研发计划项目(2020YFB1712403)。
关键词 生成对抗网络 多维步态数据 增强 下肢外骨骼机器人 generative adversarial network multi-dimensional gait data augmentation lower-limb exoskeleton robot
  • 相关文献

参考文献4

二级参考文献37

  • 1http://www.ehendrick.org/healthy. (2010-06).
  • 2Bradley D, Marquez C, Hawley M, et al. NeXOS - the design, development, and evaluation of a rehabilitation system for the lower limbs [J]. Mechatronics, 2009, 19: 247-257.
  • 3Malcolm P, Segers V, Caekenberghe IV, et al. Experimental study of the influence of the m. tibialis anterior on the walk-to-run transition by means of a powered ankle-foot exo- skeleton [J]. Gait Posture, 2009, 29: 6-10.
  • 4Akdogan E, Arif M. The design and control of a therapeutic ex- ercise robot for lower limb rehabilitation: Physiotherabot [J]. Mechatronics, 2011, 21:509-522~.
  • 5Bradley D, Marquez C, Hawley M, et al. NeXOS - the design, development, and evaluation of a rehabilitation system for the lower limbs [J]. Mechatronics, 2009, 19:247-257.
  • 6American Heart Association. [2010]. http://www.heart.org.
  • 7Dietz V. Body weight supported gait training: from laboratory to clinical setting [J]. Brain Res Bull, 2009, 78: I-VI.
  • 8Hussain S, Xie SQ, Liu GY. Robot assisted treadmill training: Mechanisms and training strategies [J]. Med Eng Phys, 2010,33: 527-533.
  • 9Klamer T, Henry K, Chan C, et al. Patterns of muscle coordina- tion vary with stride frequency during weight assisted treadmill walking [J]. Gait Posture, 2010, (31): 360-365.
  • 10Hocoma. Lokomat. functional locomotion therapy with aug- mented feedback[OL], http://www.hocoma.ch/en/products/lo- komat/. [2009-12-08].

共引文献49

相关作者

内容加载中请稍等...

相关机构

内容加载中请稍等...

相关主题

内容加载中请稍等...

浏览历史

内容加载中请稍等...
;
使用帮助 返回顶部