欢迎访问《图学学报》 分享到：
图学学报

图学学报 ›› 2021, Vol. 42 ›› Issue (2): 174-181.DOI: 10.11996/JG.j.2095-302X.2021020174

• 图像处理与计算机视觉 • 上一篇    下一篇

基于 LSTM 神经网络的人体动作识别

  

  1. 西安理工大学机械与精密仪器工程学院，陕西 西安 710048
  • 出版日期:2021-04-30 发布日期:2021-04-30
  • 基金资助:
    国家自然科学基金项目(51475365)；陕西省自然科学基础研究计划项目(2017JM5088) 

Human action recognition based on LSTM neural network

  1. School of Mechanical and Instrumental Engineering, Xi’an University of Technology, Xi’an Shaanxi 710048, China
  • Online:2021-04-30 Published:2021-04-30
  • Supported by:
    National Natural Science Foundation of China (51475365); Natural Science Basic Research Program of Shaanxi Province (2017JM5088) 

PDF (PC)

378

摘要： 人体动作识别为人机合作提供了基础支撑，机器人通过对操作者动作进行识别和理解，可以提 高制造系统的柔性和生产效率。针对人体动作识别问题，在三维骨架数据的基础上，对原始三维骨架数据进行 平滑去噪处理以符合人体关节点运动的平滑规律；构建了由静态特征和动态特征组成的融合特征用来表征人体 动作；引入了关键帧提取模型来提取人体动作序列中的关键帧以减少计算量；建立了以 LSTM 神经网络为基础 的 Bi-LSTM 神经网络的人体动作分类模型，引入注意力机制以及 Dropout 进行人体动作分类识别，并对神经 网络的主要参数采用正交试验法进行了参数优化；最后利用公开数据集进行动作识别实验。结果表明，该模型 算法对人体动作具有较高的识别率。

关键词: 动作识别, 融合特征, LSTM 神经网络, 注意力机制, Dropout 

Abstract: Human action recognition provides the basic support for human-computer cooperation. Robots can enhance the flexibility and production efficiency of manufacturing system by recognizing and understanding the operator’s action. To resolve the problem of human motion recognition, the original 3D skeleton data was smoothed and denoised to conform to the smooth rule of human joint-point motion based on 3D skeleton data. The fusion feature composed of static and dynamic features was constructed to represent human action. The key frame extraction model was introduced to extract the key frames in human action sequences to reduce the computing load. A Bi-LSTM neural network model based on LSTM neural network was established to classify human actions, and the attention mechanism and Dropout were utilized to classify and recognize human actions, with the main parameters of the neural network optimized by the orthogonal test method. Finally, the open data set was employed for the action recognition experiment. The results show that the proposed model algorithm has a high recognition rate for human actions. 

Key words:  , action recognition, fusion features, LSTM neural network, attention mechanism, Dropout 

中图分类号: