计算机应用

• 人工智能与仿真 •    下一篇

基于渐进式神经网络架构搜索的人体运动识别

王震宇,张雷,高文彬,权威铭   

  1. 南京师范大学 电气与自动化工程学院,南京 210046
  • 收稿日期:2021-05-17 修回日期:2021-09-13 发布日期:2021-09-13 出版日期:2021-09-22
  • 通讯作者: 张雷

Human activity recognition based on progressive neural architecture search

  • Received:2021-05-17 Revised:2021-09-13 Online:2021-09-13 Published:2021-09-22

摘要: 为了解决基于传感器数据的运动识别问题,利用深度卷积神经网络在公开的OPPORTUNITY传感器数据集进行运动识别。首先,神经网络模型设计过程中不再依赖于手动选择合适的拓扑结构,而是通过渐进式神经网络架构搜索算法设计最优拓扑结构以最大化F1得分;其次,该算法使用了基于序列模型优化(SMBO)策略,在该策略中将按照复杂度从低到高的顺序搜索结构空间,同时学习一个代理函数引导对结构空间的搜索;最后,将搜索过程中表现最好的20个模型在OPPORTUNIT数据集上进行完全训练,从中选出表现最好的模型作为搜索到的最优架构。通过这种方式搜索到的最优架构在OPPORTUNITY数据集上的F1得分达到了93.08%,与进化算法搜索到的最优架构及DeepConvLSTM相比分别提升了1.34%和1.73%,证明该方法能够改进以前手工设计的模型结构,且是可行有效的。

关键词: 人体运动识别, 深度学习, 神经网络架构搜索, 卷积神经网络, 基于序列模型优化

Abstract: Concerning the sensor-based activity recognition problem, deep convolutional neural network was proposed to perform activity recognition using the public OPPORTUNITY data set. Firstly, in the process of neural network model design, it was no longer dependent on manual selection of suitable topology, and the progressive neural network architecture search algorithm was used to design the optimal topology in order to maximize the classification F1 score; Secondly, a Sequential Model-Based Optimization(SMBO) strategy was used in the algorithm, in which the structure space was searched in order of increasing complexity, while learning a surrogate function to guide the search through the structure space; Finally, the 20 models with the best performance in the search process were fully trained on the OPPORTUNIT dataset, and the best performing model was selected as the optimal architecture searched. The F1 score of the optimal structure searched in this way reaches 93.08% on the OPPORTUNITY data set, which is increased by 1.34% and 1.73% respectively compared with the optimal structure searched by evolutionary algorithm and DeepConvlSTM, which indicates that the proposed method can improve previously manually-designed architectures and is effective.

Key words: human activity recognition, deep learning, neural architecture search, convolutional neural network, sequential model-based optimization

中图分类号: