The Research On Human Action Recognition Based On Wi-Fi

Computer science is developing rapidly,Intelligent Human–Computer Interaction and Pervasive Intelligent Computing are two important branches of Computer Science.Human-computer interaction requires the interactive mode to meet the characteristics of natural,friendly and humanized.Pervasive computing emphasizes the calculation of the integration of the environment,that is,people can obtain and deal with information at anytime,anywhere,in any way.They are important components in the field of artificial intelligence and in big data applications."Human centered" is their core ideas and goals.The human action recognition technology occupies the core position in these two fields and is an important supporting technology.Therefore,the intensive study on human action recognition has great significance for the development of ubiquitous Intelligent Computing and human-machine interaction technology and plays an important role in virtual reality,sports rehabilitation,smart home and other fields.There are many ways of traditional human action recognition,such as mechanical action capture,acoustic action capture,electromagnetic action capture,action capture based on image / video,and the action capture based on wireless signal in recent years.Among them,human action recognition technology based on video / image and human action recognition technology based on inertial sensor parameters are the two mainstream human action recognition technologies at present.The two technologies are relatively mature at present,and some representative commercial products have been born,such as Microsoft's Kinect,Leap Motion and Noitom's action capture system.But they also have obvious shortcomings.The human motion recognition based on video / image is sensitive to light and can only work in the range of sight distance.Although human motion recognition based on inertial sensors can recognize many human actions at present,it requires users to carry or deploy a large number of sensor nodes,which is expensive and difficult to deploy,and it will also bring additional burden to users' body.In order to avoid the above problems,this paper uses the wireless channel state information(CSI)to recognize human actions.Compared with the traditional method of human action recognition using signal indication strength(RSSI),CSI reflects the information of physical layer during signal transmission,so it can reflect the change information in real environment.In this paper,an ordinary Wi-Fi router and a commercial wireless network card are used to design a non-invasive method of table tennis action identification in the ordinary apartment environment.Table tennis is a complex and systemic movement,which brings great challenges to anomaly detection,anomaly extraction and action classification.The research work of this paper can be divided into three parts:(1)The collection and preprocessing of the original data.This part describes the use of commercial wireless devices to build experimental platform to get the original CSI data and transform the CSI data to the form we need.Due to the existence of a large number of interfering factors in the environment,according to the characteristics of interference factors,this paper adopts Butterworth low-pass filter method and wavelet soft threshold de-noising,and achieved good results.At the same time,we use special hardware devices(DMR*** radio monitoring system)to collect data of remote human activity monitoring.The data processing methods are similar to CSI data processing methods,but the data are somewhat different.(2)Anomaly detection and anomaly extraction.Combined with its own practical application,we designed an anomaly detection method-Abnormity detection algorithm based on PCA,which can accurately detect whether there is action in the human data flow,and accurately identify the continuous interval of action.The detection rate of CSI data stream is 98%,and the detection rate of data stream for special hardware is up to 100%.This undoubtedly proves the effectiveness and versatility of this method.(3)Feature extraction and sample classification.In this paper,spectral analysis and wavelet packet decomposition are used to extract the eigenvalues that can reflect human dynamics exactly.Finally,we use SVM for training and classification.The average accuracy rate of 6 table tennis action recognition is 90.33%,then we added 3 "symmetrical" actions,and the correct recognition rate dropped to 80%,which was confirmed by the theory proposed by previous researchers,the correct recognition rate dropped to 80%.This is in accordance with the predecessors' theory.We conducted an CSI based human action recognition experiment in an ordinary family environment.The results confirm the feasibility and effectiveness of the proposed method,and we have a clear understanding of the effect of CSI on human action recognition.At the same time,remote human activity monitoring experiment based on special hardware also proves that remote monitoring of human activities is feasible and effective.