Research And Application Of CSI-based Activity Recognition Technology

In recent years,activity recognition based on Channel State Information(CSI)has become a popular research direction.The basic principle is that different human activities reflect different WiFi signals,and by processing the CSI data in the WiFi signals,it is possible to identify human activities.Compared with the video domain,CSI-based activity recognition does not rely on cameras,which can protect user privacy.Additionally,it can work well under varying lighting conditions.Compared with wearable sensors,CSIbased activity recognition does not require users to wear any special devices,which is more convenient for users.Therefore,CSI-based activity recognition has broad application prospects in human activity monitoring,smart home,and other fields.This thesis proposes a multi-scale convolutional Transformer system for activity recognition based on CSI.The system uses multi-scale convolutional modules to effectively extract local information from CSI data and uses Transformer to extract global information from the data.In addition,this thesis extends the current CSI-based activity recognition task to a CSI-based activity detection task,which requires the model not only to predict the activity category,but also to predict the time boundaries of the activity.A new dataset was collected,and the performance of the model was verified through experiments.Furthermore,since CSI data is not intuitive in form and difficult to label,it requires a lot of manual labor and time.Therefore,semi-supervised learning techniques were used in the study,allowing the model to learn from a small amount of labeled data and a large amount of unlabeled data,and the effectiveness of the semi-supervised technique was verified through experiments.