Research On RF-based Key Technologies Of Contactless Target Sensing

Wireless sensing technology has become a hot research topic,such as indoor location,activity recognition,vital signs monitoring,and so on.Compared with the traditional camera-based and sound-based sensing methods,RF-based sensing method has attracted a lot of academia and industry attention due to the advantages of not infringing on privacy,not relying on light intensity,being able to work in non-line-of-sight scenes and having a long sensing range.In addition,RF-based sensing methods have achieved significant progress,such as lip recognition,material recognition,identity recognition and eye movement monitoring.However,there still have four issues when applying RF-based sensing methods in practical environments.First,the sensing accuracy is low.Due to hardware noise and environmental multipath,cheap commercial wireless devices only provide noisy sensing data,which limits the sensing accuracy in the actual environment.Second,the sensing robustness is poor.Users’ action deviation will cause the mismatch between test features and training features,leading to the decline of sensing robustness.Third,the sensing human effort is large.When the environment changes,to ensure high sensing accuracy,existing methods require collecting a large amount of data in a new environment,resulting in significant sample collection costs.Fourth,the sensing range is limited.The nature of existing contactless sensing methods is to employ weak target reflected signals for sensing,which are easily drowned by noise,resulting in limited sensing range and poor robustness based on commercial devices.Based on different application scenarios such as material recognition,identity recognition and gesture recognition,etc.,this dissertation explores the internal relationship between signal amplitude and phase distortion and target material,shape and motion,and explores line-of-sight and multipath signal’s amplitude and phase distribution in the spatial domain.Then,it combines signal processing,wireless channel modeling,deep learning and hardware design to seek effective schemes for solving the problem of low sensing accuracy,poor robustness,large human effort and limited sensing range.The main work of this dissertation is as follows:1)To solve the low sensing accuracy problem,this paper proposes a contactless target material identification method,namely Wi Mi.By observing that the phase and amplitude errors between the CSI reading obtained from two closely-placed antennas are similar,we propose to utilize two antennas on the same equipment to extract stable phase difference and amplitude ratio.In addition,in order to eliminate the impact of indoor multipath,we further observe some subcarriers may be greatly affected by multipath while the rest may not due to the frequency diversity,we thus select the clean subcarriers for improving the accuracy of target sensing.2)To deal with the problem of poor sensing robustness,a multi-feature fusion recognition method based on deep learning is proposed,namely RF-identity.The method analyzes the influence of the target on the disturbance of the wireless channel,and obtains the target motion characteristics by combining the phase and amplitude information.While,we exploit tag diversity in the spatial domain to obtain body shape features.Finally,multidimensional features are fused with a deep learning framework to achieve high robust target recognition.3)To overcome the issue of huge human cost,this paper introduces Wi-Learner,a solution that relies on just one data sample per gesture to deliver accurate cross-domain gesture recognition.Since wireless signals carry a lot of information irrelevant to the target behavior,environmental changes lead to different signal distributions for the same action.Thus,Wi-Learner introduces a meta-learning scheme,making the neural network-based sensing system learns to learn by itself,and adjust the data distribution.Then,a convolution neural network-based autoencoder is designed to capture the Doppler frequency-shifted signal features caused by behavior,automatically filter out noise interference and effectively reduce the feature dimension.Finally,the model can quickly adapt to the new environment with only one sample of the target domain,which greatly reduces the human cost.4)To address the problem of limited sensing range,a low-cost metasurface-based wireless RF link optimization method is proposed,namely RFlens.The core idea is to design a phase-modulating metasurface,and deploys it in the environment to adjust the phase of wireless electromagnetic waves.Then,RFlens directs the transmitted signal to the target direction for enhancing weak reflected signals,so as to expand the sensing range.Specifically,RFlens explores the relationship between metasurface unit-cell geometry and signal characteristics(i.e.,frequency,phase,and amplitude),and optimizes system key parameters by establishing a physical model of the metasurface element to reduce reflection attenuation.At the same time,the relationship between material cost and transmission attenuation is balanced,thus realizing a metasurface with low-cost and high transmission efficiency,which effectively improves the signal strength of the target and expands the sensing range.Unlike traditional beamforming methods that deploy a large array of antennas at the RF end,RFlens does not require modifications to commercial equipment hardware and protocols,greatly reducing labor costs.