WiFi CSI-based human sensing has risen high research attention in current days,and three main contents are fine-grained activity sensing,coarse-grained activity sensing,and localization.Existing work has already achieved some promising results,but WiFi CSI-based human sensing still has space to move on as WiFi has limitations in low bandwidth and time resolution.There are three main drawbacks:(1)as for finegrained activity sensing,there exists a gap in the respiration model built on the mobile client;(2)as for coarse-grained activity sensing,the model has poor capabilities in cross-domain sensing because the signals are not independent of the background environment;(3)as for localization,the contactless relative position model is not yet perfect.To solve the above problems,this thesis conducts research on WiFi CSI-based fine-grained activity sensing,coarse-grained activity sensing,and localization.Besides,in order to realize multi-person collaborative sensing,this thesis further carries out the research of WiFi CSI-based collaborative framework.The main contributions of this thesis are shown as follows.(1)In terms of WiFi-based fine-grained activity sensing,we construct a theoretical model concerning the sensing position based on the Fresnel theory to fill the gap of WiFi-based human sensing through mobile clients.We propose a novel method for finegrained activity sensing based on CSI theoretical model,which can extract reliable subcarriers in strong interference environments and provide placement strategies for mobile terminals in multi-target scenarios,solving the problem of the impact of sensing device's location on the sensing performance and realizing the respiration sensing by smartphones.(2)In terms of WiFi-based coarse-grained activity sensing,in order to solve the problem of poor capabilities in cross-domain sensing due that the signals are not independent of the background environment,we design a data de-noising and analysis model with CSI adaptive filtering by analyzing the noise distribution characteristics of CSI raw data.On the basis of this model,we propose a cross-domain sensing method based on a multi-label adversarial network,which increases the accuracy of crossdomain gesture recognition,realizes a low-cost passive alarm activity recognition system,and improves the robustness of the fall detection.(3)In terms of WiFi-based localization,we construct a model to localize the targets' relative positions in a contactless way based on WiFi probes in response to the imperfect framework of relative location sensing.We design two indicators to judge the localization similarity: localization similarity coefficient,and close contact distance,to realize a novel range-free judgment scheme to define localization similarity.This method can be utilized for searching and tracking COVID-19 patients' close-contacts.(4)Based on the above researches,we further explore multi-person collaborative sensing.In particular,we propose a federated learning-based collaborative framework to solve the sensing model's under-fitting problem given the lack of WiFi sensing data.This framework mainly contributes to the privacy protection of collaborative training data in multiple end-users,solving environment dependence of the sensing model,and updating users' local models,achieving high robustness and high versatility in the sensing model.In summary,this thesis makes efforts on four aspects,i.e.,WiFi-based fine-grained activity sensing,coarse-grained activity sensing,localization,and collaborative sensing,and we propose the robust and pervasive sensing methods correspondingly.We also carried out a large number of experiments with commercial WiFi devices in different indoor scenes,and the results verify the effectiveness of the methods in this thesis.The thesis has 86 figures,5 tables,and 193 references. |