Design And Implementation Of WiFi Sensing Enhancement Method Based On Image Inpainting

In the past two decades,WiFi technology has achieved great success in data communication.In recent years,WiiFi technology has been further used in the field of sensing due to the attenuation of the channel,which will carry environmental information.Although WiFi sensing applications have made great progress,the strong dependence of WiFi sensing systems on high sampling rates has not been paid enough attention.WiFi sensing usually requires high frequency sampling(200 to 2,000 packets per second)to sense changes in the environment,and WiFi packet transmission rates determined by application requirements cannot consistently meet this requirement.Therefore,the existing WiFi sensing system needs to send special high-frequency data packets for sensing.These "sensing packets"greatly affect the normal communication function of WiFi and cannot realize the multiplexing of communication signals expected by wireless sensing.To solve the above problems,this paper proposes a WiFi sensing enhancement method based on image completion,which uses machine learning method to enhance WiFi sensing ability under low sampling rate,promotes the integration of WiFi sensing and WiFi communication,and pushes WiFi sensing closer to practical application.The core idea of the sensing enhancement method proposed in this paper is to convert lowsampling rate Channel State Information(CSI)into damaged color images.This conversion method can not only deal with traditional three-antenna WiFi data,but also can deal with one-antenna WiFi data or two-antenna WiFi data.Then,based on the characteristics of WiFi perception,an enhanced sensing Generative Adversarial Network(GAN)was desiggned to carry out CSI image completion,thus reducing the need for sampling rate in WiFi perception.In order to avoid the large number of training samples required for GAN training,a lightweight GAN training method is designed in this paper.In this paper,it is found that the original random loss can be discretized based on the characteristics of CSI images,so the corresponding rate matching algorithm is designed.Finally,a lightweight GAN training method is designed and realized,which only needs to train three specific speed models to recover any rate CSI.The experimental results show that the accuracy of gesture recognition and daily activity recognition can be improved from 59.1%and 65.9%to 86.7%and 96.4%respectively by using only 25 data packets per second for the current mainstream sensing systems.Even under the condition of one antenna,the proposed method can improve the accuracy of gesture recognition and daily activity recognition to 80.0%and 92.6%.