A Study Of Cross-Domain Human Sensing Technology Based On Channel State Information

With the rapid development of the Internet of Things,there are higher requirements for sensing technology to conduct human-computer interaction more efficiently.As an emerging sensing technology,wireless sensing technology can sense the movement,identity,and vital signs of a target by analyzing the changes of wireless signals around the target.Due to the advantages of Wi-Fi such as low price and easy deployment,Wi-Fi-based sensing technology has attracted much attention.In this paper,some explorations on human sensing technology based on Wi-Fi CSI are conducted as follows.(1)A Wi-Fi CSI-based indoor human sensing method was proposed for human action and identity recognition.A Wi-Fi CSI acquisition system was built to realize WiFi CSI data acquisition by installing the CSItool open-source software tool on a laptop computer configured with an Intel 5300 wireless network card.Based on this,the more effective CSI amplitude information was extracted and the noise of CSI amplitude was filtered out using Butterworth bandpass filter,followed by reshaping CSI amplitude into a three-dimensional matrix to maximize the preservation of spatial,temporal,and frequency information carried by CSI amplitude,and finally,the recognition of human action and identity was achieved by constructing a recognition model.(2)A multi-task neural network recognition model based on CSI was designed to realize the simultaneous recognition of action and identity.The model used twodimensional convolution to extract features from a three-dimensional matrix,used randomly discarded neurons,batch normalization,early stop,and regularization to reduce the overfitting of the model,and finally used dynamic weighted averaging to dynamically set the weights of model loss.The simultaneous recognition of the user’s action and identity could enhance the generalization performance of the model.After experimental testing,the accuracy of the proposed CSI action and identity recognition model was 95.67% and 87.33% for action and identity recognition,respectively.(3)A CSI-based migration learning method is proposed for realizing cross-domain perception of different users’ actions.To solve the problem that the action recognition accuracy of a new user in a fully trained model decreased significantly,by introducing the idea of cross-domain,the new user and the original user were defined as coming from the target and source domains,respectively.On this basis,a cross-domain perception model based on semi-supervised migration learning was established,and the features and their distribution differences between the new user and the original user after neural network extraction were analyzed,and the COARL distance was used as an index to evaluate the feature distribution differences,to construct the inter-domain distribution loss of the cross-domain perception model.After experimental tests,the proposed cross-domain perception method could recognize the actions of new users with 90.4% accuracy.