Research Of Human Activities Recognition Method Based On WiFi CSI

With the rapid development of environmental perception and communication technology,the monitoring of human daily activities has been widely used in the fields of smart home,entertainment,and medical care.Human activity recognition technology has attracted the attention of academia and industry.In human activity recognition schemes based on wearable devices,cameras,and environmental devices,there are limitations such as inconvenient wearing,invasion of privacy,expensive devices,and complex deployment.Device-free Wi Fi sensing technology overcomes the above limitations and enables human action recognition by correlating different human behaviors with channel distortions.Aiming at the problems of low recognition accuracy and few recognition types in the current use of Wi Fi Channel State Information(CSI)for human activity recognition technology,this dissertation studies a parallel convolutional neural network(CNN)and Long Short-term memory network(LSTM)deep learning algorithm for activity recognition method,the main research contents include:Research on the acquisition of human activity CSI data use Intel 5300 network card,and then research on the preprocessing of the CSI data.Since the collected CSI data has noise in all frequency bands,Discrete Wavelet Transform(DWT)is used to filter the noise,which not only obtains the human activity information in the low frequency range,but also retains the high frequency detail components;then,this dissertation designs A CSI subcarrier selection and matrix sparse scheme is proposed.This scheme uses Principal Component Analysis(PCA)method to select subcarriers for CSI human activity data according to the eigenvalues.In this process,the Walsh Hadamard Transform(WHT)compresses the feature vector,which not only selects the sub-carriers closely related to human behavior,but also realizes the sparseness of the feature.A parallel CNN-LSTM deep learning algorithm is proposed for human activity recognition,and the human activity CSI data after data preprocessing is put into the established network for recognition.Since human activities change dynamically over time,on the one hand,the CNN network can be used to extract the spatial behavior features of the target activities,and on the other hand,the LSTM network can extract the time series features of the target activities that change over time.Through the parallel CNN-LSTM deep learning algorithm,the spatial behavior features and time series features are fused,and finally the fused features are classified by the Softmax classification function to identify different human behaviors.Through experimental comparative analysis,when different neural network models perform deep learning on this dataset,the accuracy,loss value and confusion matrix of human action recognition reflect their performance.The parallel CNN-LSTM deep neural network algorithm proposed in this dissertation achieves an average recognition accuracy of 98.0% for 16 activity categories,which is higher than the single CNN network,LSTM network,Bidirectional LSTM network and serial CNN-LSTM network.