• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学

• 人工智能与数据挖掘 • 上一篇    

基于CNN-LSTM的活动识别方法

李允,孟凡荣,张磊,邵长兴,崔淑敏,朱少杰   

  1. (中国矿业大学计算机学院,江苏 徐州 221116)
  • 收稿日期:2018-12-12 修回日期:2019-04-08 出版日期:2019-09-25 发布日期:2019-09-25
  • 基金资助:

    中央高校基本科研业务费专项资金(2014XT04);教育部博士点基金(20110095110010);江苏省自然科学基金(BK20130208)

An activity recognition method based on CNN-LSTM

LI Yun,MENG Fan-rong,ZHANG Lei,SHAO Chang-xing,CUI Shu min,ZHU Shao-jie   

  1. (School of Computer Science,China University of Mining and Technology,Xuzhou 221116,China)
     
  • Received:2018-12-12 Revised:2019-04-08 Online:2019-09-25 Published:2019-09-25

摘要:

在识别活动时,传统的循环神经网络RNN识别方法不考虑传感器活动数据之间依赖性强的问题,导致识别准确率降低。为了提高识别准确率,解决活动数据依赖性强的问题,用长短期记忆网络LSTM进行活动识别,LSTM在考虑当前点输入的同时考虑先前点的输出,能够保持数据之间的强依赖性。但是,LSTM在处理传感器活动数据的特征提取方面时间效率不高,而卷积神经网络CNN能共享卷积核,且可以从杂乱无章的数据中提取出明显特征向量。提出一种基于CNN-LSTM的活动识别方法CLAR,利用CNN能够很好地提取出活动序列数据中的特征向量,并将提取出的特征向量作为LSTM的输入,利用LSTM门限之间的相互作用进行活动识别,使得依赖性很强的活动数据成为活动识别的优势,进而提高活动识别的准确率和时间效率。实验表明,CLAR方法的识别准确率比单一神经网络活动识别方法的准确率提高了9%,时间平均缩短了10%。
 

关键词: 活动识别, LSTM, CNN, 特征提取

Abstract:

When identifying activities, the traditional recurrent neural network (RNN) recognition method does not consider the problem of strong dependency among sensor activity data, resulting in a decrease in recognition accuracy. In order to improve the recognition accuracy and solve the problem of strong dependency among sensor activity data, the long-short term memory network (LSTM) is used for activity recognition. The LSTM considers both the input of current points and the output of the previous points to maintain strong dependency among data. However, it has low time efficiency in feature extraction for sensor activity data. The convolutional neural network (CNN) can share convolution kernels and extract obvious feature vectors from disordered data. We present a CNN-LSTM activity recognition (CLAR) method. The CLAR uses the CNN to extract the feature vectors in the activity sequence data, takes the extracted feature vectors as the input of the LSTM, and uses the interaction between thresholds to recognize activities, which makes the highly dependency among activity data become an advantage of activity recognition, thus improving the accuracy and time efficiency of activity recognition. Experimental results show that the CLAR method is 9% more accurate than the single neural network activity recognition method, and it consumes 10% less time on average.

 

 

Key words: activity recognition, LSTM, CNN, feature extraction