Research On Human Behavior Recognition Based On Channel State Information In WiFi Environment

The rapid development and increasing popularity of new applications such as virtual reality,smart home and smart transportation,where WiFi as a sensing tool makes the development of wireless sensing a step further.With the help of WiFi to achieve wireless sensing,in the future smart home and other scenarios,according to the user’s different behavioral action instructions to complete different specified tasks.At the level of human-computer interaction,along with the development of WiFi sensing,different behavioral actions can be performed within the specified WiFi sensing area to realize the control of smart devices over space and make the future life more intelligent.The aim of this thesis is to study the recognition of human behavioral actions in WiFi environment.With wireless sensing as the domain background,a commercial WiFi router is used as the data collection device,and the benchmark experiments are validated by open source datasets.By analyzing the impact of different human behavioral activities on the received signals,the channel state information reflecting the activity state is then inferred.Exploring the denoising algorithm to recover the signals of pure activity fragments,studying the waveforms of the activity fragments,extracting their statistical features and time-frequency domain features,and constructing human behavioral activity datasets,which are subsequently classified and identified by classification algorithms to complete the classification.The main research contents are as follows:(1)For the human activity recognition problem,in order to do a more impartial analysis of the recognition effect,this thesis uses a publicly available open source dataset.In the pre-processing stage,the median filter,Butterworth low-pass filter,mean filter and discrete wavelet transform denoising algorithms are compared in detail,and the discrete wavelet transform denoising algorithm that can retain high-frequency detail components and filter out low-frequency components is selected.After denoising,the activity segments are subjected to Principal Component Analysis(PCA)to downscale the subcarriers,further mining statistical features to characterize human activities,and finally the classification of human activities is completed by a random forest classifier.Several experiments are conducted in laboratory,corridor and non-visual scenes.The recognition of six kinds of human activities in the laboratory environment achieves 92% recognition accuracy.(2)For the gait-based identification problem,the gait information of different human bodies is analyzed,and pre-processing algorithms such as CSI-Ratio,conjugate multiplication,and discrete wavelet transform are compared.The feature extraction part selects the time-domain and frequency-domain features reflecting activity characteristics,and constructs new features by merging feature vectors to form new features.Stock Well time-frequency analysis is introduced to extract the time-frequency mapping features of gait information.Finally,CNN-Bi LSTM network is used to mine the deep features in the air domain and time domain,after which the classification and recognition of gait is completed,and an accurate recognition rate of 91.7% is achieved through simulation experiments.