1. A through-wall passive people number prediction method based on channel state information is characterized by comprising the following steps: the method comprises the following steps:

step 1: respectively arranging a receiver and a transmitter at two sides of a wall body, and acquiring channel state information by the receiver;

step 2: preprocessing the channel state information acquired by the receiver;

step 2.1: removing outliers of the sub-carriers by adopting a Hampel filter and interpolating;

step 2.2: eliminating errors by adopting a linear fitting method to carry out phase correction;

step 2.3: removing high-frequency noise by using a discrete wavelet threshold on the sub-carriers;

and step 3: amplitude time domain correlation coefficient matrix A of channel state information is equal to (a)_m,n)_M×NPerforming characteristic decomposition, and arranging the decomposed characteristics in descending order according to the size of the characteristic value to obtain Z₁＝[λ₁ λ₂ … λ_M]；

And 4, step 4: time domain correlation coefficient matrix for phase C ═ C_m,n)_M×NPerforming characteristic decomposition, and arranging the decomposed characteristics in descending order according to the size of the characteristic value to obtain Z₂＝[γ₁ γ₂ … γ_M]；

And 5: selecting the first two large characteristic values lambda containing main information₁、λ₂And gamma₁、γ₂As a correlation characteristic, the subcarrier correlation coefficient matrix S is (S)_m,n)_K×KPerforming characteristic decomposition, and arranging the decomposed characteristics according to the size of the characteristic value in a descending order to obtain [ e ]₁ e₂ … e_K]；

Step 6: calculating a first order difference meanSelecting phi₁，φ₂，φ₃As a wall-through number prediction feature;

and 7: standardizing the preprocessed channel state information to be recorded as Y, and obtaining each principal component p through projection_i＝Y×e_i(ii) a Calculating each principal component p_iVariance of (1), noted as beta_i(ii) a Selection of beta₁，β₂，β₃As detection features, a feature space F is constructed through features selected by a time domain and a frequency domain;

F＝[λ₁ λ₂ γ₁ γ₂ φ₁ φ₂ φ₃ β₁ β₂ β₃]

and 8: and inputting the feature space F into a BP neural network for training to obtain a trained classifier for predicting the number of human bodies.

2. The method for predicting the number of the passive people passing through the wall based on the channel state information as claimed in claim 1, wherein: the method for removing and interpolating the outliers of the subcarriers by using the Hampel filter in the step 2.1 specifically comprises the following steps:

step 2.1.1: setting a threshold value alpha and a sliding window length v;

step 2.1.2: calculating median Mid of sequence data X_u,v；

Step 2.1.3: calculating respective data X in the sequence data X_uAnd median Mid_u,vThe absolute difference MAD of;

MAD＝abs(x_u-Mid_u,v)

step 2.1.4: if data X in sequence data X_uIs not in [ (Mid)_u,v-α*MAD),(Mid_u,v+α*MAD)]Within the range, x is determined_uFor outliers in sequence data X, use the median Mid of sequence data X_u,vInstead of the value of the outlier.

3. The method for predicting the number of the passive people passing through the wall based on the channel state information as claimed in claim 1 or 2, wherein: the method for eliminating errors and performing phase correction by using a linear fitting method in the step 2.2 specifically comprises the following steps:

since the phase in the collected channel state information is not the true phase, the measured phase has a certain deviation from the true phase, and therefore the phase needs to be calibrated; the true phase is represented as: denotes the measured phase, θ, of the b-th sub-carrier_bRepresenting the true phase of the b-th sub-carrier,denotes the phase offset, epsilon denotes the constant phase deviation; eliminating error by linear fitting method, and assuming that the sequence number b of each subcarrier is gradually increased in sequenceThe phase after random noise removal is expressed as: the representative number is b_jThe true phase of the sub-carrier.

Background

The ability to estimate the population in an area can be used for a variety of applications. For example, smart buildings can optimize energy consumption based on the number of people in the building; retailers can better plan their business by assessing which portions of the store attract more guests; smart cities can better plan resources by estimating which areas of the city are more congested. Moreover, researchers in the field of computer vision, wireless networks, and environmental science have studied the population counting problem. For example, in computer vision, a photograph of an area is used to identify the number of people in the area. Researchers in the environmental sciences use characteristics of a region of interest, such as temperature, carbon dioxide concentration, etc., to determine the number of people in the region. However, these methods 1) require installation of a network of cameras in the area of interest and are therefore costly to deploy, 2) do not work in the dark, 3) do not work behind the wall, and 4) have privacy concerns. And the Wi-Fi-based wireless sensing technology research can overcome the difficulties and be applied to different environmental conditions.

The earliest technology for wireless sensing is based on Received Signal Strength Indication (RSSI), a wireless router is used as a transmitter, receiving equipment supporting Wi-Fi is used as a receiver, and the Received RSSI is used as a parameter for environmental sensing and human body invasion. Therefore, the RSSI is widely used for human body detection and indoor positioning. For example, in indoor positioning, the RSSI can be combined with a wireless signal transmission model to realize positioning of an indoor intrusion target. However, the propagation path of the indoor signal is affected by the shielding object, and the wireless signal has multiple propagation paths indoors, so that the acquired RSSI value fluctuates greatly along with the change of time, the stability of the RSSI research method is deteriorated, and the sensing accuracy of the RSSI is greatly limited. In 2010, scientific researchers complete rewriting of drivers, so that Channel State Information (CSI) is obtained from a commercial Inter5300 network card, and compared with RSSI, CSI has Channel Information with finer granularity, scientific research requirements of the scientific researchers on higher precision can be met, and therefore human body target detection based on the CSI is developed more.

In recent years, with the popularization of wireless devices and the increase of the demand of wireless sensing technology, the research on the application technology of human body detection and human body quantity prediction by people is greatly promoted. The main research scene is in a visual distance scene, in a wall-through scene, due to the fact that Wi-Fi signals are seriously attenuated when passing through a wall body, the analysis difficulty of the signals is increased, a large number of experiments show that the human body number prediction effect of single signal characteristic analysis in the wall-through scene is general, and the human body number prediction effect in the wall-through scene is better by selecting multi-dimensional signal characteristics.

Disclosure of Invention

The invention aims to provide a method for predicting the number of through-wall passive people based on channel state information.

The purpose of the invention is realized by the following technical scheme: the method comprises the following steps:

step 1: respectively arranging a receiver and a transmitter at two sides of a wall body, and acquiring channel state information by the receiver;

step 2: preprocessing the channel state information acquired by the receiver;

step 2.1: removing outliers of the sub-carriers by adopting a Hampel filter and interpolating;

step 2.2: eliminating errors by adopting a linear fitting method to carry out phase correction;

step 2.3: removing high-frequency noise by using a discrete wavelet threshold on the sub-carriers;

and step 3: amplitude time domain correlation coefficient matrix A of channel state information is equal to (a)_m,n)_M×NPerforming characteristic decomposition, and arranging the decomposed characteristics in descending order according to the size of the characteristic value to obtain Z₁＝[λ₁ λ₂ … λ_M]；

And 4, step 4: time domain correlation coefficient matrix for phase C ═ C_m,n)_M×NPerforming characteristic decomposition, and arranging the decomposed characteristics in descending order according to the size of the characteristic value to obtain Z₂＝[γ₁ γ₂ … γ_M]；

And 5: selecting the first two large characteristic values lambda containing main information₁、λ₂And gamma₁、γ₂As a correlation characteristic, the subcarrier correlation coefficient matrix S is (S)_m,n)_K×KPerforming characteristic decomposition, and arranging the decomposed characteristics according to the size of the characteristic value in a descending order to obtain [ e ]₁ e₂ … e_K]；

Step 6: calculating a first order difference meanSelecting phi₁，φ₂，φ₃As a wall-through number prediction feature;

and 7: standardizing the preprocessed channel state information to be recorded as Y, and obtaining each principal component p through projection_i＝Y×e_i(ii) a Calculating each principal component p_iVariance of (1), noted as beta_i(ii) a Selection of beta₁，β₂，β₃As detection features, a feature space F is constructed through features selected by a time domain and a frequency domain;

F＝[λ₁ λ₂ γ₁ γ₂ φ₁ φ₂ φ₃ β₁ β₂ β₃]

and 8: and inputting the feature space F into a BP neural network for training to obtain a trained classifier for predicting the number of human bodies.

The present invention may further comprise:

the method for removing and interpolating the outliers of the subcarriers by using the Hampel filter in the step 2.1 specifically comprises the following steps:

step 2.1.1: setting a threshold value alpha and a sliding window length v;

step 2.1.2: calculating median Mid of sequence data X_u,v；

Step 2.1.3: calculating respective data X in the sequence data X_uAnd median Mid_u,vThe absolute difference MAD of;

MAD＝abs(x_u-Mid_u,v)

step 2.1.4: if data X in sequence data X_uIs not in [ (Mid)_u,v-α*MAD),(Mid_u,v+α*MAD)]Within the range, x is determined_uFor outliers in sequence data X, use the median Mid of sequence data X_u,vInstead of the value of the outlier.

The method for eliminating errors and performing phase correction by using a linear fitting method in the step 2.2 specifically comprises the following steps:

since the phase in the collected channel state information is not the true phase, the measured phase has a certain deviation from the true phase, and therefore the phase needs to be calibrated; the true phase is represented as: denotes the measured phase, θ, of the b-th sub-carrier_bRepresenting the true phase of the b-th sub-carrier,denotes the phase offset, epsilon denotes the constant phase deviation; eliminating error by linear fitting method, and assuming that the sequence number b of each subcarrier is gradually increased in sequenceThe phase after random noise removal is expressed as: the representative number is b_jThe true phase of the sub-carrier.

The invention has the beneficial effects that:

the present invention addresses the problem of predicting the number of people when transceivers are on both sides of a wall, and provides a method that maintains high detection performance when Wi-Fi signals pass through different wall materials. The invention extracts multidimensional characteristics from the time domain correlation of the subcarriers and the correlation among the subcarriers respectively, and selects a BP neural network with good capability of processing complex data to complete the mapping of detection characteristics and prediction results.

Drawings

FIG. 1 is a flow chart of the present invention.

Fig. 2 is a graph of the detection accuracy of a person on three different wall materials.

Fig. 3 is a graph of the detection accuracy of two persons in three different wall materials.

Fig. 4 is a graph of the detection accuracy of three persons on three different wall materials. .

Detailed Description

The invention is further described below with reference to the accompanying drawings.

The invention belongs to the field of through-wall passive human body detection based on channel state information, and particularly relates to a through-wall passive people number prediction method based on channel state information. The invention aims to provide a method for predicting the number of through-wall passive people based on channel state information, which can keep higher detection performance when Wi-Fi signals pass through different wall materials.

The purpose of the invention is realized as follows: the method comprises the following steps:

step 1: respectively arranging a receiver and a transmitter at two sides of a wall body, and acquiring channel state information by the receiver;

step 2: extracting the collected channel state information for preprocessing;

step 2.1: removing outliers of the sub-carriers by adopting a Hampel filter and interpolating;

step 2.2: in order to reduce the error of the original phase, a linear fitting method is adopted to eliminate the error for phase correction;

step 2.3: further removing high-frequency noise by using a discrete wavelet threshold on the sub-carrier;

and step 3: and (5) extracting features. The preprocessed matrix can be represented as:

order toIndicating the CSI of the subcarriers in the mth packet,indicating the CSI for the k-th subcarrier.

CSI amplitude time domain correlation coefficient matrix A ═ a_m,n)_M×NCan be expressed as:

similarly, we can also calculate the time domain correlation coefficient matrix C ═ C (C) of the phase by the same method_m,n)_M×N

Wherein c is_m,n＝corr(∠g_m,∠g_n) Respectively performing characteristic decomposition on the matrix A and the matrix C, and arranging Z in descending order according to the size of characteristic values₁＝[λ₁ λ₂ … λ_M],Z₂＝[γ₁ γ₂ … γ_M]According to the property of the matrix, the main information of the matrixes A and C is contained in the first two characteristic values, so that the lambda is selected₁、λ₂And gamma₁、γ₂As a correlation feature.

Then the correlation coefficient matrix S for the sub-carriers is equal to (S)_m,n)_K×KCarry out feature decomposition in descending order [ e ]₁ e₂ … e_K]And calculating the first order difference mean value of the characteristic valueSelecting phi₁，φ₂，φ₃As a feature for predicting the number of people passing through the wall, in addition, theStandardized processing recordFor Y, the respective principal components p are obtained by projection_i＝Y×e_i. And calculates each principal component p_iVariance of (1), noted as beta_iSelecting beta₁，β₂，β₃As detection characteristics, a characteristic space F ═ lambda is constructed through characteristics selected by a time domain and a frequency domain₁ λ₂ γ₁ γ₂ φ₁ φ₂ φ₃ β₁ β₂ β₃]。

And 4, step 4: and selecting a BP neural network to train data and predicting the number of human bodies.

Compared with the prior art, the invention has the beneficial effects that: the invention is applied to the field of through-wall passive human body detection based on channel state information, and mainly aims at predicting the number of human bodies when transceivers are arranged on two sides of a wall body. Because the traditional human body quantity prediction based on the channel state information is finished in the same indoor environment of the transceiver, the prediction result in the Wi-Fi signal through-the-wall scene is not ideal. In order to better process the channel state information, the method extracts multidimensional characteristics from the time domain correlation of subcarriers and the correlation among the subcarriers respectively, and selects a BP neural network with good capability of processing complex data to complete the mapping of detection characteristics and prediction results.

As shown in fig. 1, the method includes four modules in total: data acquisition, signal preprocessing, multi-dimensional feature extraction and classifier construction.

(1) A data acquisition module: the transmitter and the receiver are respectively arranged in two adjacent rooms or the transceivers are arranged in a corridor adjacent to the rooms, and the distance between the transceivers is about 3 meters. A TP-Link router with double antennas is used as a transmitter, and a notebook computer provided with an Intel 5300 wireless network card is used as a receiver. The sampling frequency is set to be 200Hz, the sampling time is 60s, and data sampling is respectively carried out on nobody, 1 person, 2 persons and 3 persons existing indoors.

(2) A preprocessing module: in the process of acquiring the CSI data, the sub-carriers are reflected in the wireless transmission process, so that an error exists between the data received at the receiving end and the actual data, and the data needs to be preprocessed to reduce the influence on the prediction result.

(2.1) abnormal value processing: the method adopts a Hampel filter. Setting a sliding window with length k, and calculating the median Mid of sequence data X_i,kAnd calculating the absolute difference MAD ═ abs (x) between each data and the median value_i-Mid_i,k) Where i represents the ith point of X for sequence data, when X_iIs not in [ (Mid)_i,k-γ*MAD),(Mid_i,k+γ*MAD)]When within the range, x is_iWhen an outlier in the sequence data X is found, the median Mid of the sequence X is used_i,kInstead of this point, γ is chosen to be 3.

(2.2) phase correction: since the phase in the collected channel state information is not the true phase, the measured phase has a certain deviation from the true phase, and the phase needs to be calibrated, and the true phase is expressed as: denotes the measured phase, θ, of the k-th sub-carrier_kRepresenting the true phase of the k-th sub-carrier,indicating a phase offset and epsilon a constant phase deviation. The method adopts a linear fitting method to eliminate errors, and assumes that the serial number of each subcarrier is gradually increased in sequenceThe phase after random noise removal is expressed as: the representative number is k_jThe true phase of the sub-carrier.

And (2.3) carrying out discrete wavelet threshold denoising treatment on the processed data.

(3) Multi-dimensional feature extraction: respectively corresponding to CSI amplitude time domain correlation coefficient matrix A ═ a_m,n)_M×NAnd the time domain correlation coefficient matrix C of the phase (C ═ C)_m,n)_M×NPerforming characteristic decomposition Z₁＝[λ₁ λ₂ … λ_M]，Z₂＝[γ₁ γ₂ … γ_M]And selecting the first two large characteristic values lambda containing main information₁、λ₂And gamma₁、γ₂As a correlation feature. The subcarrier correlation coefficient matrix S ═ (S)_m,n)_K×KAnd (3) performing characteristic decomposition to arrange the characteristic values in a descending order, wherein the characteristic vectors corresponding to the characteristic values are as follows: [ e ] a₁ e₂ … e_K]And obtaining a first order difference mean value according to the feature vectorTo prevent the loss of the main content in the selected feature, select phi₁，φ₂，φ₃As a through-wall people number prediction feature, the normalized CSI data after preprocessing is marked as Y, and each principal component p is obtained through projection_i＝Y×e_i. And calculates each principal component p_iVariance of (1), noted as beta_iSelecting beta₁，β₂，β₃As detection characteristics, a characteristic space F ═ lambda is constructed through characteristics selected by a time domain and a frequency domain₁ λ₂ γ₁ γ₂ φ₁ φ₂ φ₃ β₁β₂ β₃]。

(4) And selecting a BP neural network with better mapping capability on nonlinear data to train the extracted features.

(5) And (4) according to the feature extraction in the step (3), performing feature extraction on the residual data, and predicting a data result through a trained classifier.

The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.