Action Recognition Based On Channel State Information And Deep Learning

Human action recognition is one of the important research contents of intelligent applications.The accurate identifi cation of human behavior can improve the quality of human-computer interaction and expand the scope of intelligent application,which has great application prospect and economic value for the study of intelligent home,intelligent teaching video,lip language recognition,keyboard recognition and so on.Traditional action recognition has some restrictive problems such as sensitive lighting,high cost,special equipment to be carried,privacy and security invasion.To avoid these problems,we use the channel state information of wireless signals to recognize human actions.The specific research contents of this paper can be divided into the following parts.(1)The current research status of human body motion recognition is summarized,including current research status based on video,sensors,radio frequency based,etc.Typical research of CSI-based motion recognition is given;and action classification method based on CSI-based motion recognition is summarized.This section describes the principles of RSS and CSI for describing wireless channels,and compares RSS and CSI.(2)This article sets up an experimental environment for acquiring CSI data.Then,a CSI sensing motion signal capability test is performed.Whether or not the person's actions really caused a change in the wireless signal can be judged.Then,the CSI raw packets corresponding to different actions are collected in the experimental environment.(3)This paper analyzes the composition and structure of CSI original data packets and explains the meaning of each parameter in the data packets.All CSI data packets representing different actions are preprocessed,and CSI amplitude is extracted.The CSI value of the same action and the CSI value of different actions are analyzed based on the amplitude.The CSI amplitude of all actions is extracted to build an action database.(4)This paper analyzes the basic principles of the recurrent neural network and the convolutional neural network,and establishes a deep neural network structure based on CNN and RNN.Then,the action database is randomly assigned as a training sample,a validation sample,and a test sample.Training samples and validation samples are used to train and verify the model,and then test samples are used to test the model.Finally,the classification accuracy of different actions is obtained.The contribution of this article is,The CSI data generated by different human actions is collected through experiments.Then,the CSI data is converted into an image data to input CNN;the CSI data is converted into a time series input RNN,and CNN-based and RNN-based indoor human motion recognition are implemented.Through the use of neural networks,the complexity of indoor signal transmission is effectively solved,and the diversity of multiple actions such as movement amplitude and movement speed is eliminated,and high-accuracy indoor human movement recognition is realized.