Joint Recognition Of Human Behavior And Activity Location Based On Wi-Fi-CSI

With the rapid development of modern communication technology and the wide popularity of wireless network equipment,wireless perception technology is gradually applied to human’s daily life.Among them,the key technology of wireless sensing through Wi-Fi Channel State Information(CSI)is highly emphasized.Compared with sensing technology based on sensors and video images,it has advantages such as wide deployment,privacy security and strong robustness,so it is favored by more and more researchers and has been used in many fields such as indoor positioning,intrusion detection,breath monitoring and behavior recognition.Even Wi-Fi wireless sensing technology of human behavior recognition has made certain research results,but most of the existing system based on CSI human awareness is a single study scene,low accuracy of perception model,is used to identify the position location or behavior and only one,but the ubiquitous wireless smart scenarios of human-computer interaction,It will be more valuable to study the joint recognition of human behavior and activity location in multiple scenes.In view of the above problems,this thesis conducts a lot of research on wireless perception joint recognition task,carries out data collection and relevant experiments,and proposes a joint recognition method of human behavior and activity location using Wi-Fi combined with deep residual shrinkage network.First,CSI data sets related to human behavior and activity location in three scenarios(darkroom,conference room and corridor)were obtained using common commercial Wi-Fi devices.Then in data pretreatment phase Hampel codes filter is adopted to improve the abnormal points out,adopting Butterworth low-pass filter to noise data,then use PCA to data dimension reduction,got contains most of the characteristic information of low dimension feature subset,subsequent linear interpolation,to obtain the length of the consistent CSI data into neural network to classify.Finally,the pre-processed data were sent into the deep residual shrinkage network with soft threshold,and the classification results associated with 12 positions and 7 kinds of actions were obtained.The experimental results show that the average recognition accuracy of position classification and behavior recognition in the three scenarios is 96.19% and 89.48%respectively.It can realize high precision joint recognition of position and behavior.In addition,compared with other methods,several parameters of the model,including model structure,network layers,activation function and optimizer,are analyzed experimentally to verify the effectiveness of the proposed method.