Research On WiFi Human Activity Recognition In Different Environments

Under the current trend of the Internet of Everything,the technology of utilizing artificial intelligence to realize human activity recognition has developed rapidly.Many human activity recognition technologies available on the market are quite mature,including cameras and some wearable sensor devices.But due to the use environment limitations and the high cost of related equipment,practical applications of these technologies have been hampered.On the other hand,in view of the fact that WiFi signals are widely available in daily life and the low-cost WiFi devices,scholars have begun to study how to use Channel State Information(CSI)in WiFi signals to identify human activities.However,although using WiFi to recognize human activities works fine,in general,WiFi signals are mixed with information about the environment,leading to the performance of the model trained in one environment would not up to expectations when used in different environments.In this thesis,in order to solve the above-mentioned problems,WiFi CSI signal data is used as the basis to distinguish different human activities,and a WiFi human activity recognition model suitable for any different environment is proposed.The main work completed in this thesis are as follows:(1)The CSI signal is used to identify human activities,and a CSI acquisition device is built.All the data needed for the experiment are collected within a period of more than 40 days,and the data cleaning and feature extraction are completed.(2)In order to explore the model with strong generalization function,in this thesis,the domain adaptation technology in image field is applied to CSI data,and a human activities recognition model based on the domain adversarial algorithm is proposed.The model can automatically ignore the environmental interference factors in CSI signals and use the features favorable to the activity recognition for classification.In this thesis,the data is divided into source domain with labeled samples and target domain with unlabeled samples.According to the needs of the actual scene,the single-source domain adversarial model WiFi-DANN and the multi-source domain adversarial model WiFi-MDAN are designed.The experimental results show that the activity recognition rate of the single-source WiFi-DANN model is 73.8%,which is 7.8% higher than when the domain adversarial algorithm is not used for learning,while the multi-source WiFi-MDAN model achieves a recognition accuracy of 90.6% in different environments.(3)An environment-adaptive WiFi-EAN model is proposed.In this method,based on the general training model to deal with the activity recognition work in most environments,the middleware model is used to modify the CSI data generated in the special environment so as to adapt it to the general training model.Then the modified CSI data are classified by the general model to obtain the relevant recognition results.The experimental results show that the activity recognition accuracy of the WiFi-EAN model output reaches more than 90%.