Student Action Recognition Based On CSI Signal Using Improved Residual Network

The development of artificial intelligence technology has greatly improved the quality of human life.Among them,behavior recognition technology has been applied in many fields,and it is also sought after by researchers in the field of education,and has gradually become a research hotspot of teaching intelligence.The research on video recognition teaching action has achieved certain results,but there are still problems such as visual blind spots,privacy leakage,and insensitivity to the subtle movements of characters.At the same time,it is difficult for traditional residual networks to effectively extract the channel features and spatiotemporal feature information of Wi-Fi data.For this kind of problem,this dissertation mainly uses Wi-Fi channel state information(Channel State Information,CSI)combined with the improved residual network to solve this kind of problem.The main research work is as follows:1.There are not many cases of student behavior monitoring research in the past,and the use of Wi-Fi signal monitoring and related student motion data sets is even less common.In the early stage of the thesis,an experimental platform was built by modifying the computer driver and replacing the adapter network card.Then,the data of students’ behavior is collected,and the CSI data set of students’ behaviors is established.The experimental design actions are “Hands Up”,“Wave”,“Bow”,“Applaud”,“Walk”,“Sit Down”,and collect data in three different environments.The data collection of students completing these actions in the presence of obstacles was supplemented in the later stage.Then,useing MATLAB tools to process the collected data.2.For the problem of data loss in the deep network,in order to ensure that the data in the deep network is not lost.The main network of this dissertation uses Res Net.Since each CSI signal contains multiple subcarriers,this dissertation uses the threshold feature to improve the attention SE module,so that the module can enhance the feature extraction of the CSI signal by the network while maintaining the original extraction channel characteristics.The threshold module can further filter out the data that is irrelevant or weakly related to the action,and enhance the action signal features.The improved module can be applied to the residual network as a new module to improve the network performance,and the improved network can recognize 94.22% of actions.3.Considering that the CSI signal has certain spatial characteristics,the improved module can extract the channel information of the CSI signal but ignore the spatial information.On the basis of the improved module,this dissertation further adds the function of extracting data space features,so that the network can extract the features of the signal from different dimensions.After that,the architecture of the improved network is introduced,and experiments are carried out on different networks.At the same time,the data in different environments are verified and analyzed.The accuracy rate of the best effect of the network can reach 98.52%.