Research And Implementation Of WiFi Activity And Gesture Recognition In Response To Changing Environments

With the development of wireless communication technology and artificial intelligence,WiFi-based indoor human sensing has become a hot research topic.Among them,activity recognition and gesture recognition are widely used.WiFi signal has the advantages of wide coverage and high popularity,and its physical layer channel state information(CSI)is more sensitive to people’s activities and gestures,which is suitable for indoor human sensing tasks.However,WiFi signals have the defects of being unstable and greatly affected by environmental changes,which restricts the further development and practical application of wireless sensing.Therefore,this thesis focuses on the impact of the environment on wireless signals,and studies the indoor WiFi human sensing from the perspective of responding to environmental changes.The sensing tasks are mainly activity recognition and gesture recognition.This thesis focuses on the impact of environmental changes on wireless signals,and studies the recognition of human activity and gesture recognition in a changing environment.In response to environmental changes,this thesis proposes a domain alignment method based on pseudo-labels.By using a small number of labeled samples to assign pseudo labels on unlabeled samples,and using dynamic time warping distance to measure the similarity between samples,iteratively assign pseudo labels on the samples with the greatest similarity in each category and convert them into soft labels form.The method uses the maximum mean discrepancy to measure the data distribution distance between different domains,and aligns domains by increasing the distance between different classes and reducing the distance between the same classes.Using convolutionrecurrent neural network to extract the temporal and spatial features of the sample,and classify activities and gestures.According to the degree of change,this thesis divides the environment into changes within a room and changes across rooms.Unsupervised domain adaptation is proposed for the changes in the room,and the samples in the target domain environment are unlabeled.Semi-supervised domain adaptation is proposed for crossroom changes,with a small amount of labeled data in the target domain environment.This thesis has carried out experimental verification on the open source data set and self-collected data set.For the unsupervised domain adaptation scheme,the average accuracy of the method in activity recognition reaches 85.2%,and the accuracy of gesture recognition reaches 82.7%,when compared with source domain model method,the increases were 22.9% and 27.2% respectively.For the semi-supervised domain adaptation scheme,the average accuracy of the method in activity recognition reaches 81.1%,and the accuracy of gesture recognition reaches 78.5%,when compared with source domain model method,the increases were 57.6% and 59.2% respectively.At the same time,this thesis also compares with other domain adaptation methods in the state of art,and the method in this thesis also outperforms them.