Research On Indoor Human Activity Detection Based On Channel State Information

Human activity detection is widely used in security monitoring,health monitoring,smart home and other fields.The recognition scheme based on wearable devices requires people to carry special sensors,and the recognition scheme based on computer vision has the limitation of blind area recognition,while the activity recognition technology based on wireless signals does not need to carry devices and has low cost,which is a new research hotspot in recent years,with good research value and significance.Firstly,based on the sensitivity of channel state information subcarriers to different activities,a coarse-grained activity recognition algorithm based on Gaussian mixture model is proposed.After collecting CSI,the algorithm calculates the CSI amplitude according to the subcarrier data format and normalizes it.The CSI amplitude is mapped to a color scale and divided into continuous pixel frames with a fixed time window size.Then,the Gaussian mixture model is used to detect the foreground of the pixel frame.Each pixel is represented by a fixed number of Gaussian distributions,each pixel is modeled according to the Gaussian distribution of multiple different weight values.Each Gaussian distribution corresponds to a state that may produce the color of the pixel.The weight and distribution parameters of each Gaussian distribution update with time.According to the deviation distance between the current pixel value and the Gaussian model,the foreground and background are distinguished,and the CSI changes caused by human activities are separated.Finally,the output of foreground detection is expressed as a result matrix,and coarse-grained activity is identified according to the index in the result matrix of multiple antennas.Secondly,on the basis of coarse-grained activity recognition,this paper proposes a fine-grained respiration detection algorithm based on sequence correlation.The algorithm takes the result matrix of the above algorithm as the input,According to the time correlation caused by respiratory activity changing the wireless signal propagation path and influencing multiple subcarriers in a certain period of time,motion extraction is carried out for each column of the result matrix to filter out the false foreground which has less influence on the number of subcarriers in the foreground detection results.Then,the outliers are removed by using the Hampel function,and the background noise of CSI focused on high frequency is filtered by Butterworth low-pass filter.The correlation function is constructed by the linear relationship of CSI changes in time delay,in which the number of function peaks is respiratory frequency,and the distance between adjacent peaks is respiratory period.At last,experiments are carried out in different CSI sampling period,distance between target and receiving device,number of antennas,sampling times,sight condition and other parameters and different environments..The results show that the average accuracy of coarse-grained activity recognition algorithm based on Gaussian mixture model can reach 98%,and the accuracy of fine-grained respiration detection algorithm based on sequence correlation can reach 96%.There are 41 pictures,4 tables and 77 references in this paper.