基于CRNN的CSI动作识别  

CSI Action Recognition Based on CRNN

作　　者：丁文超 张俊宝[1] 阴庚雷 DING Wen-chao;ZHANG Jun-bao;YIN Geng-lei(School of Computer Science,Zhongyuan University of Technology,Zhengzhou 450007,China)

机构地区：[1]中原工学院计算机学院,河南郑州450007

出　　处：《计算机技术与发展》2021年第6期7-12,共6页Computer Technology and Development

基　　金：国家自然科学基金(U1504614);河南省教育厅科学技术研究重点项目(14B520055)。

摘　　要：随着Wi-Fi感知技术的发展,出现了大量使用Wi-Fi信道状态信息(channel state information, CSI)进行动作识别的应用。然而大多数的方法在数据预处理和训练阶段都依赖于人工构建特征,构建过程耗时耗力并且需要专家的领域知识。针对上述问题,提出一种基于CRNN(convolutional recurrent neural network)的CSI动作识别方法。将不同手势的CSI数据做低通滤波处理后,通过自组织映射(self organizing maps, SOM)聚类的结果选择最佳子载波,并对该子载波上的CSI数据进行扩增。然后,使用格拉姆角求和场(Gramian angular summation fields, GASF)方法将一维CSI数据转换成二维GASF图像,作为CNN、LSTM构成的CRNN网络的输入数据,训练过程中使用链接时序分类(connectionist temporal classification, CTC)作为损失函数。实验结果表明,该方法能在训练数据较少的情况下达到较高的识别精度,且无需手动构建特征。With the development of Wi-Fi sensing technology, a large number of applications have emerged for motion recognition using channel state information(CSI) from Wi-Fi. However, most methods rely on manual construction of features during the data pre-processing and training stages, the construction process is often time-consuming and labor-intensive and requires expert domain knowledge. Aiming at these problems, we propose a CSI action recognition method based on convolutional recurrent neural network(CRNN). After the CSI data of different gestures are processed by low-pass filter, the best subcarrier is selected by the result of self-organizing maps(SOM) clustering, and the CSI data on the subcarrier is amplified. Then, one-dimensional CSI data is transformed into two-dimensional GASF images by using the method of gamma angular summation fields(GASF),which is used as the input data of CRNN composed of CNN and LSTM. In the training process, the connectionist temporal classification(CTC) is used as the loss function. The experiment indicates that the proposed method can achieve high recognition accuracy with less training data, and there is no need to manually construct features.

关 键 词：信道状态信息 CRNN 动作识别 自组织映射 格拉姆角场 链接时序分类 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]