Research On Technology Of Human Activity Recognition Under Co-Channel Interference Based On Channel State Information

With the development of communication and sensing technology,human activity recognition technology attracts more and more researchers.The continuous emergence of sensing devices has promoted the development of sensing technology.Applications such as identity authentication,fall detection and gesture recognition make people’s life more intelligent through human-computer interaction.However,traditional human activity identification methods are mainly based on cameras and wearable sensors,which are not only inconvenient to carry,but also seriously violate users’ privacy.In recent years,with the popularity and application of Wi-Fi,human activity recognition technology based on Wi-Fi has achieved good results.Despite the benefits of Wi-Fi recognition technology,such as its non-contact and non-line-of-sight nature,many researchers have not taken into account real-world conditions.Due to the existence of a large number of Wi-Fi devices,the co-channel interference caused by spectrum overlap will inevitably have a serious impact on the recognition performance.To solve this problem,this paper aims to enhance the resilience of the human activity recognition system against co-channel interference by taking a dual-pronged approach.Firstly,the influence of co-channel interference on human activity recognition system is analyzed by experiments,including reducing the signal sampling rate and weakening the correlation between subcarriers.In view of the above analysis,this scheme proposes two sub-carrier selection algorithms from the point of view of signal processing:Spearman’s correlation coefficient based subcarrier selection algorithm and Weighted Dynamic Time Warp based subcarrier selection algorithm,which are respectively applied to the scenarios where the same channel interference is constant and the same channel interference is changing.This study involves the extraction of common time-frequency domain features,followed by the application of several traditional machine learning methods to recognize human activities.Extensive experimental results conducted under various interference scenarios demonstrate that this approach achieves an average recognition rate of 94.7%in the presence of co-channel interference,thereby effectively enhancing the performance of the human activity recognition system.Secondly,from the perspective of network model,a new human activity recognition system IAI-MTL based on multi-task learning is proposed.The proposed system simultaneously learns interference recognition,action recognition,and identity recognition by sharing a common representation layer,thereby allowing for the exchange and integration of the knowledge gained during the learning process.The proposed system leverages the relevant information from interference recognition and action recognition to enhance the performance of identity recognition,thereby improving the overall accuracy of action recognition.The experimental results show that the action recognition accuracy of the system is 96.9%,and the training cost and model complexity are reduced in different degrees compared with the single task learning and transfer learning.