Research On Human Behavior Recognition Method Based On Channel State Information

Human behavior recognition is widely used in smart homes,health care and indoor monitor.Compared with traditional detection methods using wearable devices and cameras,the Wi-Fi-based detection method has certain advantages.Wi-Fi based solutions are device-free and users do not need to wear any sensors.Compared to cameras,Wi-Fi signals have better coverage and can still work in poor lighting conditions.This thesis carries out the following research work on Wi-Fi based indoor human behavior recognition.(1)By analyzing the characteristics of CSI and studying the human perception model based on CSI,the feasibility of CSI for activity recognition is proved.To explain the relationship between the changes in human behavior and CSI,a CSI path decomposition model is derived.On this basis,the concept of Fresnel zone is introduced,and the Fresnel reflection model is constructed to explain the periodic variation of CSI waveform.(2)A series of signal processing methods are designed according to CSI characteristics.CSI is a complex form with two dimensions,amplitude and phase.For the amplitude information,the outlier filtering and median filtering methods are designed,and the approximate phase of CSI is obtained by linear transformation method.In addition,to solve the problem of CSI subcarrier selection,a subcarrier selection algorithm based on principal component analysis is designed.(3)To realize accurate recognition of human behavior,a recognition method based on feature description is proposed.In the process of behavior recognition,the extracted behavior features are used for behavior classification,but the extracted features cannot accurately describe human behavior,and there is a certain similarity between these features.Therefore,this thesis analyzes the sensitivity of different features to behaviors and the similarity between features to select the features with the best performance for feature matrix.The experimental results show that the selected features can realize the accurate recognition of human behavior,and the average recognition accuracy is about 95% in the case of single target.(4)In general,single dimension features are used in behavior classification,which cannot reflect the position relationship between features.Therefore,the convolutional neural network is used to transform the single dimension features into spatial dimension features.In addition,by using residual network module and attention mechanism,the performance of the network is improved,and finally the average recognition accuracy is about 96%.The thesis contains 53 figures,7 tables and 83 references.