Research And Implementation Of Gesture Recognition Method Based On WiFi

In the information age in which artificial intelligence technology is developing rapidly,the emergence of smart device products makes people's lives more and more convenient.Among them,gesture recognition technology has become a very important way in humancomputer interaction.The existing gesture recognition technology based on Wi Fi signals can only recognize a single user or fewer types of gestures,and the traditional machine learning method of manually extracting features is cumbersome and has limited effect on classifying details.Therefore,researching multiple gesture recognition technologies for multiple users,and improving their recognition accuracy,and realizing the intelligentization of humancomputer interaction have become an urgent issue for the application of this technology.Aiming at the problems that the existing gesture recognition technology does not have the ability to adapt to user differences and gesture diversity,and the shallow classification algorithm needs to manually obtain gesture statistical features,a gesture recognition method based on deep learning is proposed;The feature extraction of gestures is not comprehensive in the network,and the recognition accuracy of multiple types of gestures by multiple users is low.A parallel long short-term memory(LSTM)and a fully convolutional neural network(Fully Convolutional Neural,FCN)joint network model.The main research work is as follows:(1)For For the gesture recognition method built using deep learning,first extract the channel state information(Channel State Information,CSI)data from the Wi Fi signal;then for the characteristics that the gesture action is mainly contained in the low-frequency component,study through the removal of abnormal points and low-pass filters preliminary filtering and the use of discrete wavelet transform algorithms to remove noise as much as possible while ensuring the completeness of gesture features;finally,two deep learning methods,Convolutional Neural Network(CNN)and LSTM,are used to extract and identify the CSI data.(2)In the gesture recognition method based on the parallel LSTM-FCN network,LSTM is responsible for learning the dependence of gestures in the time dimension;FCN extracts the rich detailed features in the spatial dimension,and achieves pixel-level gestures by upsampling through transposition convolution Feature,using global average pooling to prevent overfitting.Finally,the Concatenate layer combines the hidden features of the gestures extracted from the space-time dimension,and classifies the gestures through the Softmax function.This paper proposes that deep learning gesture recognition method can extract features better than traditional machine learning methods.For multi-user gesture scenarios,this paper proposes that LSTM-FCN can comprehensively learn the characteristics of the spatial and temporal dimensions of gestures.In the experimental verification,the experimental data set contains 50 types of gestures from 5 users,and the average recognition accuracy rate is about98.6%,which proves the effectiveness of this method.By comparing and analyzing the performance of the four methods of parallel LSTM-FCN,LSCN,CNN and LSTM,it is proved that parallel LSTM-FCN has higher recognition effect.In addition,the influence of non-line-of-sight environment on the recognition rate is discussed.The results show that the recognition rate of the line-of-sight environment is higher than the non-line-of-sight environment.