Research On Human Behavior Recognition Based On CSI

Human motion recognition technology is one of the hotspots of Internet of Things technology,which has strong theoretical research and practical application value.Especially in indoor scenes,due to the influence of multi-path and other factors,the efficiency of identification and equipment cost have attracted much attention.Unlike traditional device-based schemes,the continuous progress of wireless technology has promoted the development of device-free sensing,which can use ordinary WiFi signals to sense human state.Without the user's equipment,it can be realized including indoor ranging,gesture recognition,orientation recognition,motion tracking and human number sensing.Because no device technology does not require users to wear any devices or even participate in the sensing process,it provides users with a new way of intelligent sensing.Since channel state information describing carrier level in physical layer can be measured by commercial WI-FI devices,more and more researchers begin to use it for more reliable and convenient human detection.Firstly,this thesis studies the indoor environment model,uses the optimal Fresnel zone to set up the experimental environment,and finds the most effective signal area.According to the different use environment,PEM(Percentage of non-zero Elements)index is introduced to evaluate the indoor environment,which effectively reduces the adverse signal interference caused by environmental disturbance.Secondly,a large number of experiments are carried out on the collected raw data,and a subcarrier selection method is found to replace the CSI data stream average.The PCA stream is used to select the carrier,and then some high-frequency noise is removed while the maximum interest information of the signal is retained.Thirdly,a lightweight algorithm FallBR based on CSI is designed,and a complete experimental algorithm flow is designed.The data cleaning and feature extraction are studied.Finally,by comparing a large number of filtering data,a band-pass filter is found to achieve the optimal denoising effect.The experimental data after cleaning are analyzed by time-frequency analysis,and the corresponding frequency features are extracted by wavelet analysis and time-frequency analysis to distinguish the different events of human body.The experimental results show that the passive lightweight human behavior recognition method proposed in this thesis reduces the experimental error by adding the environmental index evaluation PEM,and experiments are carried out in two different experimental scenarios.It is found that the experimental effect is better in the environment with low multipath effect,and the sampling frequency has a greater impact on the experiment,and the experiment has a greater impact on human walking behavior.The recognition accuracy is over 90%.The average accuracy is slightly higher than that of CARM and Han,and the cost of experiment is low.It does not need a lot of data training to reduce the computational complexity.For the bending and squatting behavior of fall behavior and its similar characteristics,this method uses the PBC feature based on radar signal from the frequency perspective to identify the fall,thus the error is lower.