Wireless Human Activity Recognition Method With Interference Detection Feedback Mechanism

With the rapid development of artificial intelligence and Internet of Things technology,human activity recognition has become an important research direction in the field of humancomputer interaction.Among them,commercial Wi-Fi-based human activity recognition technology,which is inexpensive to deploy and does not require human-carrying equipment,has broad application prospects.However,the propagation paths of Wi-Fi signal is extremely susceptible to interference,and it is still difficult to apply this technology widely.Existing commercial Wi-Fi-based human activity recognition schemes mostly use artificial features and traditional machine learning technologies.Although they can achieve high recognition accuracy in ideal scenarios,this thesis verifies that they work not well in interference scenarios.In order to improve the practicability of human activity recognition technology based on commercial Wi-Fi,for the problems of low recognition accuracy and poor stability when the existing methods are affected by interference,this thesis proposes a solution for human activity recognition in interference scenarios,using deep learning algorithms to obtain more anti-interference recognition performance and adding an interference detection as feedback mechanism to guarantee stable performance.Therefore,we focuses on two core issues.1.A more anti-interference human activity recognition method is studied.In order to obtain a human activity recognition model suitable for a variety of interference effects,firstly we establish a diverse set of interference activity data sets.Then a Long Short-Term Memory(LSTM)algorithm suitable for CSI data is proposed,and we establish IT-LSTM human activity recognition models under different interferences.Experiments show that,compared with the benchmark machine learning method,this method improves the average accuracy of activity recognition in interference scenarios from 46% to 79%.2.An interference detection method that guarantees stable recognition performance is studied.In order to improve the accuracy of interference detection,we first perform timefrequency domain analysis on the CSI data,constructing interference-rich spectrum fingerprints.Then we use the convolutional neural network to train the interference detection model,and propose a model interpreter-based method to evaluate the performance of the model.By marking the key information that the model depends on,it is observed whether the model has learned stable information to distinguish the type of interference,so as to judge whether the current model is reliable.Experiments show that in 9 types of interference detection tasks,the interference detection model achieves 94.38% detection accuracy.In activity recognition tasks,this method provides feedback for activity recognition methods,thereby improving the average accuracy of activity recognition in interference scenarios 17%,the average false positive rate is reduced by 24%.This thesis evaluates in three indoor scenes with interference,and the results show that the proposed wireless human activity recognition scheme with interference detection feedback mechanism achieves an average accuracy of 90.55% and an average misjudgment of 2.5%.