Research On Sleep Action Recognition Technology Based On WiFi Signal

In recent years,people's awareness of health management has been increasingly enhanced,and their attention to the quality of sleep has been gradually raised.They hope to be able to know their sleep status conveniently and quickly in their daily life.To address this need,a number of researchers have focused on home-based sleep monitoring.However,the traditional sensor-based and computer vision-based perception technology has some problems,such as expensive equipment,easy occlusion of perception range and invasion of privacy.With the popularization of wireless network and intelligent terminal,Wi Fi-based wireless perception technology has been favored by many researchers due to its advantages of low cost,wide coverage and protection of privacy.As one of the index parameters of sleep quality,the activity of turning over in the sleep stage has not been reported publicly due to its various types and difficulties such as the similar influence of some activities on the channel.This thesis explored and studied the recognition technology of 8 typical sleep activities.The specific research work is as follows:(1)Aiming at the problem of low accuracy of Convolutional Neural Network(CNN)recognition of activities,a Recognition Error Correction method,called REC method,based on the correlation between consecutive activities is proposed in this thesis.In the data acquisition stage,different from the existing methods of centrally placing the receiving antennas,according to the characteristics of spatial diversity,the receiving antennas are distributed in this thesis,so as to obtain multidimensional Channel State Information(CSI).Then,the outliers are removed from the original CSI amplitude sequence,and the CSI amplitude sequence of sleep action is extracted by the threshold-based sliding window method.Then,the data from the subcarrier sensitive to the activity is selected as the training sample of the CNN model.In order to improve the recognition performance of the model,this thesis defines the concept of reliable group based on CNN output matrix,takes reliable group as the benchmark and the relevance between continuous actions as the principle,and proposes a REC method to correct the activities of CNN misidentification.Experimental verification shows that this method has a good improvement effect on CNN recognition results,and the recognition accuracy of sleep activity can reach 95%under the condition of sufficient training samples.(2)Aiming at the problem that the REC method does not significantly improve the recognition performance of CNN when the training samples are insufficient,this thesis modifies the threshold condition and improves the error correction steps on the basis of the REC method,and proposes an Improved Recognition Error Correction method,called IREC method.This method solves the problem that the REC method cannot correct when the two preselected activities are difficult to distinguish by setting the interval threshold and the credible threshold between them.Experimental verification shows that when the original CNN recognition accuracy is low,this method can still increase the recognition accuracy of sleep activities to85%,which is 23% higher than the recognition accuracy using the REC method.