Research On WiFi CSI Activity Recognition Based On Bidirectional GRU And Hierarchical Attention Mechanism

In recent years,with the increase in the number of wireless devices deployed,wireless signals are widely used for research in indoor positioning,target detection and activity recognition,which have strong practical value for medical and health,smart home,life and entertainment,etc.It is the key research content of human-computer interaction.During transmission,WiFi signals are reflected,scattered,and diffracted on the moving human body.Fine-grained channel state information(CSI)can capture changes in these signals,and different activities can be recognized by analyzing the change patterns of CSI.Compared with sensor-based and computer vision-based activity recognition,activity recognition based on WiFi CSI signals has a series of advantages,such as low cost,easy deployment,non-line-of-sight,no light limitation,and non-invasive perception.The accuracy of the existing activity recognition methods is not satisfactory.By analyzing the activity data,it can be seen that the CSI signal is a time series,and the CSI activity samples have two levels of sub-channel and sub-carrier,the variation pattern of each sub-channel and sub-carrier affected by the activity is different,the contribution to the final activity recognition is different.Based on this,this thesis proposes a method that can match the composition characteristics of CSI activity data to improve the accuracy.The main contributions of this paper are as follows:(1)This thesis proposes a WiFi CSI activity recognition method based on Bidirectional Gate Recurrent Unit-Hierarchical Attention(Bi GRU-HA),which can well match the hierarchy level of WiFi CSI signals and timing characteristics.(2)This thesis collects four datasets including daily activities and classroom activities for two typical indoor environments and two frequency bands of WiFi.The effectiveness of the Bi GRU-HA method is verified by comparing the recognition results of public datasets and self-collected datasets with other methods,and the influence of each module of Bi GRU-HA on the recognition performance is explored by control variable experiments.The average recognition accuracy of Bi GRU-HA on the public datasets is 98.4%,and the average recognition accuracy of daily activities and classroom activities on the self-collected datasets is 92.2% and 89.2%,respectively.(3)In view of the time-consuming and labor-intensive collection of CSI activity data,the datasets size is small,which leads to network overfitting.This thesis proposes three methods for data augmentation,including time stretching,frequency shifting,and different scenarios activity sample data mixing,and the effectiveness of the above data augmentation method is verified by comparative experiments on self-collected datasets.The accuracy rates after time stretching and frequency shifting are increased by 5.9% and6.4%,respectively,and the optimal ratio of data mixing is obtained.