Research On Refinement And Generalization Method Of Wireless Sensing Data Samples Based On Video Analysis

Device-free wireless sensing technology,as an essential innovative sensing paradigm,has been widely used in many sensing scenarios,such as action recognition,identity authentication,and fall detection.Due to wireless signals cannot be visualized,it is difficult to distinguish which samples are non-compliant and obtain the distribution of the samples.Hence,the existence of non-compliant samples reduces the quality of the datasets,and the distribution of limited samples results in poor model generalization capability.In this thesis,we focus on two problems: 1)removing non-compliant wireless data and 2)improving the generalization capability of wireless sensing recognition models.To solve the above issues,we propose a set of schemes and evaluate them with two types of wireless sensing datasets(CSI and RFID).The detailed research contents of this thesis are as follows:(1)This thesis proposes a wireless data refinement method based on video recognition.By using video recognition technology to pick out the inconsistent action samples,and then filter the non-compliant wireless data based on the one-to-one correspondence relationship.(2)This thesis proposes a wireless data generalization method based on video marking.We first employ video to mark the angle and position of the action samples,and then analyze the distribution of the samples.When the variation in angle or position distribution exceeds the experimental pre-set threshold,the user will re-collect the wireless data by adjusting the angle and position.This method improves the generalization capability of the wireless sensing recognition model by increasing the samples.(3)This thesis designs and implements a wireless sensing data sample refinement and generalization system,and evaluates proposed schemes on CSI datasets and RFID datasets.Extensive experiments demonstrate that our proposed method can improve the wireless sensing recognition accuracy from 89.5% to 96.5% and 89.1% to 95.8% on CSI datasets and RFID datasets,respectively.