Research On Cross-Domain Gesture Recognition Based On WiFi Signal

The rapid progress of wireless communication and Internet of Things technology has promoted the intelligent construction and popularization of the human-computer interaction environment.Making full use of the radio frequency signal coverage area established by the widely deployed WiFi wireless network,not only to realize the broadband data communication function,but also to detect the signal change pattern to reveal the wireless perception of changes in the environment,and realize the new vision of the integration of wireless communication and perception.In recent years,a large number of researches on gesture recognition based on WiFi signals have emerged,and many methods of modeling using machine learning have been proposed.However,because the collected channel state(CSI)data is incomplete and highly correlated with the domain.If the collected data is not processed in a targeted and systematic manner,the generalization performance of the trained model is actually unsatisfactory.When a new recognition domain appears,in order to ensure the recognition performance,it is often necessary to perform additional learning and training on the data of the new scene,which makes it difficult to promote practical use.Research on cross-domain gesture recognition methods and technologies based on WiFi signals and organically integrate them into a recognition system has theoretical significance and application value.In response to these problems,in order to achieve cross-domain recognition,this paper studies WiFi recognition technology in depth.First,it reviews the research background and significance,as well as the current research status of WiFi identification methods and perception models in China and abroad.Secondly,it introduces the preliminary knowledge related to WiFi recognition.Finally,on this basis,a new cross-domain gesture recognition system is proposed for WiFi signals.In the system,Preprocess the data according to the characteristics of the WiFi device,and from the data,the body-coordinate velocity profile(BVP)which independent of the environment,position,and direction is extracted as the input feature,and then a model combining 3D convolutional neural network(3DCNN)and adversarial network is designed to capture the spatial and temporal characteristics of BVP that are not related to the human body for learning and classification,and realize cross-domain gesture recognition.The system can obtain gesture information from the collected data that has nothing to do with the environment,location,direction,and the human body.It can be applied to gesture recognition in different domain only by learning and training in the data of specific domain.This paper uses the network public data set to experimentally verify the proposed system.The experimental results show that for 6 different gestures,the average crossdomain recognition accuracy reaches 88.83%.For different domain factors,the average cross-domain recognition accuracy is between 87.28%-89.41%,which can realize cross-domain gesture recognition.And compared with the recent cross-domain gesture recognition,The method in this paper has higher recognition accuracy,compared with the best existing methods,the recognition accuracy is 2.7% higher.