Device-free Human Activity Recognition Based On Channel State Information

Activity recognition is a technology that recognizes target activity by collecting and analyzing data in the perception environment.It has broad application prospects in special personnel monitoring,smart home,human-computer interaction and other fields.Activity recognition based on wireless signals collects signals reflected from a target through a receiver,and does not require any sensors to be carried by the target.It has the advantages of high security,simple deployment,and privacy protection.With the popularity of Wireless Fidelity(WiFi)devices in ordinary households,WiFi-based activity recognition has gradually developed,but the existing activity recognition algorithms have many problems that need to be further solved: First,WiFi devices receive signals contain a lot of errors and noise.When processing the signal,the existing algorithm mainly achieves the purpose of noise reduction by removing the high-frequency noise,instead of eliminating the influence of invalid paths based on differences between multipath delays;Second,the existing segmentation algorithm generally divides the channel state information(CSI)sequences by threshold,but when the propagation channel environment changes or the target changes,the threshold needs to be readjusted,and activity recognition based on machine learning often only extracts mathematical statistical features as input features of classifiers,and does not fully use the features between different sequences and different subcarriers.Aiming at above-mentioned issues,the main content of this thesis is as follow:First of all,by analyzing delay differences of multipaths,an error elimination algorithm is proposed,which not only eliminates the delay error caused by the transceiver hardware device,but also eliminates the multiple reflections.On the basis of eliminating multipath errors,the algorithm fully retains the effective information and ensures the accuracy of activity recognition.Secondly,a classifier model based on the attention mechanism is designed.This model uses the attention mechanism for data segmentation,which can not only segment the effective sequence adaptively,but also give higher weight to the channel change-related features caused by human activities and improve the accuracy of activity recognition and the training speed of the classifier model.Thirdly,when performing activity recognition,the different position of transceivers will change the path.In this thesis,two different scenarios are studied:transceiver on the same side and transceiver on different sides.In the former case,this thesis introduces an interference suppression algorithm to suppress the reflected signal and direct path signal of the non-identifying target,so as to ensure the robustness of the activity recognition algorithm.Finally,this thesis designs actual measurements on multiple scenarios and different targets,verifies the recognition performance of the system in different scenarios,and compares the performance with several existing algorithms.The experimental results show that the human target activity recognition algorithm proposed in this thesis effectively realizes a higher precision human activity recognition in different test environments.