| In recent years,Wi-Fi signal sensing has attracted the attention of researchers both domestically and abroad and has been widely applied in fields such as elderly care,healthcare,and smart homes.As Wi-Fi is not a traditional identification technology,its perception function is only a by-product of network communication.Therefore,this type of sensing technology is called Cross-domain Sensing,which uses the Channel State Information(CSI)in Wi-Fi signals to deduce the location and state of the sensing domain,explores the intrinsic relationship between the CSI and sensing domain state in Wi-Fi signals,realizes Cross-domain sensing of Wi-Fi technology,and provides intelligent services to users.Compared with other Cross-domain Sensing technologies,Wi-Fi Crossdomain Sensing has the advantages of wide deployment,no dead spots,and non-contact.Currently,based on Wi-Fi Cross-domain Sensing,the main sensing domains are target trajectory,human behavior,and physiological features,etc.This thesis mainly focuses on human behavior sensing domain research.Based on the different human behaviors in various application scenarios,the optimal solution is designed to achieve low cost,high accuracy,and high robustness in human behavior sensing.This thesis conducts in-depth research on human behavior sensing from coarsegrained to fine-grained.The main research work is as follows:(1)In order to meet the high-robustness detection problem of falling actions,a channel state information-based falling action detection method named Wi-FD is proposed.In the offline stage,a large number of various daily action data is collected,the sensitive antennas are selected using the sliding window variance method,and they are denoised using combined filters.Then,the principal component analysis algorithm is used to select the best subcarrier and convert it into a spectrum graph.The fuzzy C-means algorithm is used to segment the spectrum graph.The complete local binary pattern algorithm is used to extract features,and the features are input into SVM to obtain fingerprint data and build an offline fingerprint library.In the online stage,the real-time collected data is processed and input into SVM to obtain action fingerprints.The fingerprints are matched with the fingerprint library constructed in the offline stage to obtain the recognition result.If the result is walking,sitting,and other actions,it is classified as a normal daily activity,otherwise it belongs to the falling danger action.Experimental results in multiple environments show that the average action recognition rate of Wi-FD reaches 92.3%.It has achieved practicality and high robustness for indoor falling action detection.(2)In order to meet the low-cost and practicality requirements of rehabilitation movement sensing,a commercial wireless device-based indoor rehabilitation exercise recognition method named Wi-KF is proposed.This method first collects standard rehabilitation actions and processes their data.Then,the proposed segmentation algorithm is used to segment and count the processed data,which can accurately segment single actions from continuous movements and lay the foundation for the feature extraction algorithm.The segmented action data is then converted into a spectrum graph,and the convolutional neural network and feature bag combination algorithm are used to extract the features of the spectrum graph.It can process images of arbitrary sizes and solve the problem of spectrum graph size differences.Finally,the obtained features are input into extreme learning algorithm for classification.A large number of experimental tests were conducted in two real environments,and the average action recognition rate of Wi-KF was about 94.9%.It has achieved low-cost and high-precision indoor rehabilitation exercise recognition.(3)In response to the problem of insufficient and inaccurate feature extraction in gesture recognition,leading to low accuracy,a deep spatio-temporal gesture recognition method based on CSI signals named Wi-GC is proposed.This method first collects gesture sample data and selects the antennas sensitive to gestures.Then,a combined filter is used to denoise and smooth the selected antennas.Next,the time-series difference algorithm is used to segment continuous gestures,and the segmented gesture data is input into the RAGRU model.An algorithm combining bidirectional gated recurrent units with self-attention mechanism is used to extract the time features of the CSI sequence,and an algorithm combining residual networks with attention mechanism modules is used to extract the spatial features of the CSI amplitude.Finally,the extracted spatial and time features are fused and input into Softmax for classification.Wi-GC method is extensively and sufficiently verified in real environments,with an average gesture recognition rate between 92%-95.6%,achieving high accuracy and robustness in gesture recognition. |
PDF Full Text Request