| In recent years,with the rapid development of Io T,the massive data generated by Io T promotes the development of pervasive computing.Human activity recognition(HAR)has been extensively studied in medical health,elderly care service,sports and other fields.The most commonly used equipments within HAR are wearable sensors,which also have shortcomings: Firstly,data collected by wearable sensors donot contain environment.Secondly,the wearable sensor is in direct contact with the body,the sense of foreign body is strong which makes the comfort and acceptance of users are not high.Therefore,non-contact sensing methods based on Wi-Fi and other devices have gradually attracted attention.This paper focuses on the following researches on activity recognition based on Wi-Fi CSI signals:(1)Empirical study for activity recognition based on Wi-Fi signals: This thesis explores the gap on Wi-Fi activity recognition between exisiting research and actual use.This study is based on real scenes and is geared towards the continuously changing user activities.Perform scenarios and equipment layout,data collection and annotation,activity classification and analysis to build a complete practical operation and research process.Conduct empirical research on the impact of acquisition conditions and user location on Wi-Fi signal perception mode to explore the limitations and deficiencies of existing researchs.(2)Transfer method of knowledge between heterogeneous models for Wi-Fi signals: In real life,the users' task scenes are diversified which lead to different mission requirements.This thesis designed heterogeneous transfer learning method for Wi-Fi signals,explored the ability of knowledge transferring between different layers of two heterogeneous models,and draw guiding conclusions about the mode of cross-layer transfer between heterogeneous models.Use the method transferring arcoss datasets to evalutate whether the existing model can help the training of the new task model when the new task with large differences appears.(3)Activity sequence prediction method based on Wi-Fi signals: According to the characteristics of Wi-Fi signals containing environmental information,combined with the data collection of continuously changing activity sequences under real scenes carried out in this thesis,research and design based on Encoder-Decoder mode for activity sequence prediction method were conducted.Based on the given 5-step activity data,the following three-step activities are predicted.According to the comparison of different sequence models and search strategies,the feasibility of activity sequence prediction based on Wi-Fi signal is analyzed.(4)Design and implementation of Wi-Fi activity monitoring prototype system: This thesis aims at the monitoring requirements of indoor single people(elderly people),designed and implemented the system.Realized functions including real-time monitoring,activity alarm,historical activity data analysis,data management,etc.,providing a basis for combining Wi-Fi activity recognition with broader service tasks such as senior care and health. |
