CN112765550A - Target behavior segmentation method based on Wi-Fi channel state information - Google Patents

Abstract

The invention provides a target behavior segmentation method based on Wi-Fi Channel State Information (CSI). Firstly, the invention processes the time sequence by using a method of piecewise fitting to remove the trend, and the data fluctuation of the time period with action and the data fluctuation of the time period without action are still kept with larger difference while the whole trend of the sequence is eliminated; secondly, a sliding variance is obtained for the time series after trend removal, and a sliding window variance comparison method is provided according to the change of the sliding variance, so that the CSI data section only containing the action information is extracted. The actual measurement result shows that compared with the existing behavior segmentation method, the target behavior segmentation method designed by the invention can more accurately extract the data segment when the action occurs from the original CSI data. In addition, this method enables efficient partitioning even if the original CSI data includes a plurality of discontinuous operations.

Description

Target behavior segmentation method based on Wi-Fi channel state information

Technical Field

The invention belongs to a data processing method, and particularly relates to a target behavior segmentation method based on channel state information in a Wi-Fi system.

Background

In recent years, a human activity recognition system based on Wi-Fi wireless sensing technology is favored by more and more people because of its characteristics such as non-invasiveness and insensitivity to light. Researchers realize behavior recognition by monitoring Channel State Information (CSI) Information of echo signals and extracting signal features of different actions. Therefore, in activity recognition, it is extremely important to extract data segments from the entire data stream that only contain the time when the action occurred.

For CSI data, the change of surrounding environment information can be sensitively reflected, and the variance of the amplitude of the CSI data in the active state is far larger than that in the inactive state. Based on this, some researchers have segmented the motion by setting an optimal threshold or adaptive threshold. For example, in the TW-see system, researchers first propose a normalized variance sliding window algorithm using the characteristic that the CSI amplitude has large fluctuation variation during the occurrence of the activity; then, by setting an optimal threshold value, the detection of the starting time point and the ending time point of the action is realized. In the Wi-Multi system, researchers have designed a two-step algorithm to detect motion. Firstly, dividing the whole signal into a plurality of subdata segments with the same length; then calculating the variance of each subdata segment and sequencing the variances in an ascending order; and finally, calculating the difference value of the sum of the larger half and the sum of the smaller half in all the variances, comparing the difference value with an adaptive threshold value, and if the value is larger than the threshold value, indicating that actions exist in the subdata segment. And secondly, respectively calculating the maximum eigenvalue of the CSI amplitude autocorrelation matrix and the maximum eigenvalue of the phase autocorrelation matrix, and performing secondary verification on the action data section determined in the first step by a machine learning method.

In addition to the threshold-based motion segmentation algorithm, researchers divide a complete motion waveform into four states with the same length, which are respectively defined as a static state, a starting state, a motion state and an ending state. The off-line stage generates an action state model through a convolutional neural network, and the action state model is used for judging the activity state of the sliding window data in the action segmentation stage, so that the starting time and the ending time of a single activity are determined.

The above methods can achieve segmentation to some extent, but various disadvantages are unavoidable. In the TW-see system, the threshold or other parameters are changed, and the segmentation performance of the system is changed accordingly. The Wi-Multi system firstly divides the whole CSI data into a plurality of data sections with the length equivalent to the packet sending rate, and then performs activity judgment on each data section, and if the length of each data section is too long, the action division accuracy is reduced. The segmentation method based on deep learning requires a large amount of time for model training, and the judgment performance of the model greatly affects the subsequent action segmentation. In addition, the effectiveness of these algorithms may be reduced if the scene changes. Therefore, the invention provides a segmentation fitting trend removing and window variance comparing method to realize the segmentation of the action, and improves the instantaneity and the effectiveness of a behavior segmentation algorithm.

Disclosure of Invention

The invention aims to provide a target behavior segmentation method based on Wi-Fi channel state information in a Wi-Fi system, which can accurately and effectively extract a data segment when an action occurs from original data.

The invention relates to a Wi-Fi-based target behavior segmentation method, which specifically comprises the following steps:

the method comprises the following steps: preprocessing CSI data;

assuming that CSI data acquired by a receiver is subjected to data analysis and channel correction to obtain a matrix H, and the matrix is preprocessed to obtain a matrix A

Where N is the number of received packets, N_txAnd N_rxNumber of transmitting and receiving antennas, N, respectively_subIs the number of subcarriers. The specific flow of pretreatment is as follows:

firstly, performing modulus operation on each data in the matrix H, removing phase information in the data, then performing decentralized processing on each row vector, eliminating static components of signals, and calculating the following equation:

wherein, | · | represents a modulo operation;

step two: reducing the dimension of the matrix A by using a principal component analysis method, inhibiting noise and simultaneously enabling the CSI amplitude containing action information to be more prominent, wherein the specific method is as follows;

firstly, the covariance matrix of the matrix A is calculated, then the eigenvalue of the covariance matrix and the corresponding eigenvector are solved by using eigenvalue decomposition, finally the ith principal component is obtained by multiplying the transformed matrix A by the ith eigenvector,

h_i＝A^T×E_i

wherein h is_iIs the ith principal component of the matrix A and has dimension N × 1, E_iThe feature vector corresponding to the ith eigenvalue of the covariance matrix has the dimension of N_sub·N_tx·N_rxX 1. In the present algorithm, the second principal component h of the matrix A is mainly used₂；

Step three: using a piecewise fitting detrending method to the second principal component h obtained in the step three₂The vector T is obtained by processing the vector T,

the method comprises the following specific steps;

firstly, the second principal component h₂Divided into several data sections of length M, i.e.

h₂＝[h₂(W₁),h₂(W₂),…,h₂(W_k),…,h₂(W_N/M)]

Wherein, W_kIs an array [ kM, kM +1, …, (k +1) M-1]And indicates an index value of each data in the k-th data segment. Each piece of data was then fitted separately as follows:

T(W_k)＝h₂(W_k)-fit(h₂(W_k))

wherein fit () is a polynomial fit based on least squares, resulting in:

T＝[T(W₁),T(W₂),…,T(W_k),…,T(W_N/M)]

the following is the implementation of polynomial fitting:

suppose there are m points in the data segment, each data point being (x)_i,y_i) The fitting result is:

wherein, a₀,…,a_nFor the unknowns to be solved, n is the highest power of the polynomial. According to the least-squares method, the sum of squares of deviations is minimized, i.e.

Solving for the above coefficients can be converted to I ═ I (a)₀,a₁,…,a_n) The extreme value problem (2) is that the necessary conditions for obtaining the extreme value according to the multivariate function are as follows:

namely, it is

Using a formal representation of the matrix then:

when the points needing fitting and the maximum times of fitting of the polynomial are given, all coefficients of the fitting polynomial can be solved, and then the points before fitting are used for subtracting the points after fitting to obtain a fitting residual error to form a vector T. Subsequent data processing is based on this data.

Step four: setting a sliding window with the length K, and calculating the sliding variance V of T, wherein V is [ V ═ V [ [ V ]_l]_1×(N-K+1)；

Suppose that the data contained in the ith sliding window is [ T ]_l,T_l+1,…,T_l+K-1]Then, the variance of the ith sliding window data can be expressed as:

step five: calculating a threshold value M, and setting a window length W;

according to the fourth step, we obtain a sliding variance V, and then calculate an adaptive threshold M according to the sliding variance for motion segmentation. The specific calculation idea is as follows:

firstly, carrying out ascending arrangement on V, and searching an original index value set c of a smaller half of numerical values after the ordering before the ordering; then searching the longest continuous index value sequence segment in the set c, wherein the sliding variance data segment corresponding to the sequence segment represents the most stable data segment which does not contain human body activity information in the original signal; and finally, calculating probability density distribution and a probability accumulation integral function of the sliding variance data segment by utilizing kernel density estimation, wherein when the value of the probability accumulation integral reaches 0.975, the corresponding integral upper limit is the optimal threshold M. The window length W is set to the packet transmission rate R_SHalf of (1);

step six: and (5) segmenting the action according to the sliding variance V, the adaptive threshold M and the window length W obtained in the step five. The segmentation algorithm comprises the following steps:

step 1: searching a maximum value MAX in the sliding variance, comparing the maximum value MAX with a threshold value M, if MAX is smaller than M, indicating that no action occurs in the data segment, otherwise, indicating that the action exists in the data segment, and performing Step 2;

step 2: the index indexMax of MAX is recorded, while indexRt ═ indexLt ═ indexMax is defined. At the interval [ indexRt, indexRt+N]In search for V_lIndex value set L corresponding to the index value set larger than threshold value M₁If set L₁If the set is not empty, find the largest element in the set, assign the value of iMax to indexRt, and add indexRt to the new [ indexRt, indexRt + N [ ]]Interval finding V_lIndex value set L corresponding to the index value set larger than threshold value M₁. After the step, the ending time point indexRt of the action can be found;

step 3: in the interval [ indexLt-N, indexLt]In search for V_lIndex value set L corresponding to the index value set larger than threshold value M₂If set L₂If the set is not empty, find the smallest element iMin in the set, assign the value of iMin to indexLt, and add the value to the new [ indexLt-N, indexLt]Interval finding V_lIndex value set L corresponding to the index value set larger than threshold value M₂. After the step, the starting time point indexLt of the action can be found;

step 4: segment [ indexLt, indexRt]Corresponding V_lSetting to be 0, and then performing Step1, Step2 and Step 3;

through the steps, the data segment when the action occurs can be effectively extracted from the original data, and meanwhile, the algorithm is still effective for segmenting a plurality of discontinuous actions contained in the original data.

Advantageous effects

Firstly, the invention provides a method for processing a time sequence by utilizing a piecewise fitting trend removing method, and the data fluctuation of an action occurrence period and the data fluctuation of a non-action occurrence period are still greatly different while the overall trend of the sequence is eliminated; and then, obtaining a sliding variance from the time series after trend removal, and extracting the CSI data segment only containing the action information by a window variance comparison method according to the fluctuation change of the sliding variance. The actual measurement result shows that the target behavior segmentation algorithm designed by the invention can accurately and effectively extract the data segment when the action occurs from the original data, and meanwhile, the algorithm is still effective for segmenting a plurality of discontinuous actions contained in the original data.

Drawings

Fig. 1 is a specific implementation flow of target behavior segmentation.

Fig. 2 is a diagram illustrating the segmentation result.

Detailed description of the preferred embodiments

The method comprises the following steps: preprocessing CSI data;

assuming that CSI data acquired by a receiver is subjected to data analysis and channel correction to obtain a matrix H, and preprocessing the matrix to obtain a matrix A;

where N is the number of received packets, N_txAnd N_rxNumber of transmitting and receiving antennas, N, respectively_subIs the number of subcarriers. The specific process is as follows:

firstly, performing modulus operation on each data in the matrix H, removing phase information in the data, then performing decentralized processing on each row vector, eliminating static components of signals, and calculating the following equation:

wherein, | · | represents a modulo operation;

step two: reducing the dimension of the matrix A by using a principal component analysis method, inhibiting noise and simultaneously enabling the CSI amplitude containing action information to be more prominent, wherein the specific method is as follows;

firstly, the covariance matrix of the matrix A is calculated, then the eigenvalue of the covariance matrix and the corresponding eigenvector are solved by using eigenvalue decomposition, finally the ith principal component is obtained by multiplying the transformed matrix A by the ith eigenvector,

h_i＝A^T×E_i

wherein h is_iIs the ith principal component of the matrix A and has dimension N × 1, E_iThe feature vector corresponding to the ith eigenvalue of the covariance matrix has the dimension of N_sub·N_tx·N_rxX 1. In the present algorithm, the second principal component h of the matrix A is mainly used₂；

Step three: using a piecewise fitting detrending method to the second principal component h obtained in the step three₂The vector T is obtained by processing the vector T,

the method comprises the following specific steps;

firstly, the second principal component h₂Divided into several data sections of length M, i.e.

h₂＝[h₂(W₁),h₂(W₂),…,h₂(W_k),…,h₂(W_N/M)]

Wherein, W_kIs an array [ kM, kM +1, …, (k +1) M-1]And indicates an index value of each data in the k-th data segment. Each piece of data was then fitted separately as follows:

T(W_k)＝h₂(W_k)-fit(h₂(W_k))

wherein fit () is a polynomial fit based on least squares, resulting in:

T＝[T(W₁),T(W₂),…,T(W_k),…,T(W_N/M)]

when the points needing fitting and the maximum times of fitting of the polynomial are given, all coefficients of the fitting polynomial can be solved, and then the points before fitting are used for subtracting the points after fitting to obtain a fitting residual error to form a vector T. Subsequent data processing is based on this data.

Step four: setting a sliding window with the length K, and calculating the sliding variance V of T, wherein V is [ V ═ V [ [ V ]_l]_1×(N-K+1)；

Suppose that the data contained in the ith sliding window is [ T ]_l,T_l+1,…,T_l+K-1]Then, thenThe variance of the ith sliding window data can be expressed as:

step five: calculating a threshold value M, and setting a window length W;

according to the fourth step, we obtain a sliding variance V, and then calculate an adaptive threshold M according to the sliding variance for motion segmentation. The specific calculation idea is as follows:

firstly, carrying out ascending arrangement on V, and searching an original index value set c of a smaller half of numerical values after the ordering before the ordering; then searching the longest continuous index value sequence segment in the set c, wherein the sliding variance data segment corresponding to the sequence segment represents the most stable data segment which does not contain human body activity information in the original signal; and finally, calculating probability density distribution and a probability accumulation integral function of the sliding variance data segment by utilizing kernel density estimation, wherein when the value of the probability accumulation integral reaches 0.975, the corresponding integral upper limit is the optimal threshold M. The window length W is set to the packet transmission rate R_SHalf of (1);

step six: and (5) segmenting the action according to the sliding variance V, the adaptive threshold M and the window length W obtained in the step five. The segmentation algorithm comprises the following steps:

step 1: searching a maximum value MAX in the sliding variance, comparing the maximum value MAX with a threshold value M, if MAX is smaller than M, indicating that no action occurs in the data segment, otherwise, indicating that the action exists in the data segment, and performing Step 2;

step 2: the index indexMax of MAX is recorded, while indexRt ═ indexLt ═ indexMax is defined. In the interval [ indexRt, indexRt + N]In search for V_lIndex value set L corresponding to the index value set larger than threshold value M₁If set L₁If the set is not empty, find the largest element in the set, assign the value of iMax to indexRt, and add indexRt to the new [ indexRt, indexRt + N [ ]]Interval finding V_lIndex value set L corresponding to the index value set larger than threshold value M₁. After the step, the ending time point indexRt of the action can be found;

step 3: in the interval [ indexLt-N, indexLt]In search for V_lIndex value set L corresponding to the index value set larger than threshold value M₂If set L₂If the set is not empty, find the smallest element iMin in the set, assign the value of iMin to indexLt, and add the value to the new [ indexLt-N, indexLt]Interval finding V_lIndex value set L corresponding to the index value set larger than threshold value M₂. After the step, the starting time point indexLt of the action can be found;

step 4: segment [ indexLt, indexRt]Corresponding V_lSetting to be 0, and then performing Step1, Step2 and Step 3;

through the steps, the data segment when the action occurs can be effectively extracted from the original data, and meanwhile, the algorithm is still effective for segmenting a plurality of discontinuous actions contained in the original data.

Claims (3)

1. A target behavior segmentation method based on Wi-Fi channel state information comprises the following steps:

the method comprises the following steps: preprocessing CSI data;

assuming that CSI data acquired by a receiver is subjected to data analysis and channel correction to obtain a matrix H, and the matrix is preprocessed to obtain a matrix A

Where N is the number of received packets, N_txAnd N_rxNumber of transmitting and receiving antennas, N, respectively_subIs the number of subcarriers. The specific flow of pretreatment is as follows:

firstly, performing modulus operation on each data in the matrix H, removing phase information in the data, then performing decentralized processing on each row vector, eliminating static components of signals, and calculating the following equation:

wherein, | · | represents a modulo operation;

step two: the principal component analysis method is used for reducing the dimension of the matrix A, noise is suppressed, and meanwhile, the CSI amplitude containing action information is more prominent, and the specific method is as follows:

firstly, the covariance matrix of the matrix A is calculated, then the eigenvalue of the covariance matrix and the corresponding eigenvector are solved by using eigenvalue decomposition, finally the ith principal component is obtained by multiplying the transformed matrix A by the ith eigenvector,

h_i＝A^T×E_i

wherein h is_iIs the ith principal component of the matrix A and has dimension N × 1, E_iThe feature vector corresponding to the ith eigenvalue of the covariance matrix has the dimension of N_sub·N_tx·N_rxX 1. In the present algorithm, the second principal component h of the matrix A is mainly used₂；

Step three: using a piecewise fitting detrending method to the second principal component h obtained in the step two₂The vector T is obtained by processing the vector T,

step four: setting a sliding window with the length K, and calculating the sliding variance V of T, wherein V is [ V ═ V [ [ V ]_l]_1×(N-K+1)；

Suppose that the data contained in the ith sliding window is [ T ]_l,T_l+1,L,T_l+K-1]Then, the variance of the ith sliding window data can be expressed as:

step five: calculating a threshold value M, and setting a window length W;

according to the fourth step, we obtain a sliding variance V, and then calculate an adaptive threshold M according to the sliding variance for motion segmentation. The specific calculation idea is as follows:

firstly, carrying out ascending arrangement on V, and searching an original index value set c of a smaller half of numerical values after the ordering before the ordering; then searching the longest continuous index value sequence segment in the set c, wherein the sliding variance data segment corresponding to the sequence segment represents the most stable data segment which does not contain human body activity information in the original signal; and finally, calculating probability density distribution and a probability accumulation integral function of the sliding variance data segment by utilizing kernel density estimation, wherein when the value of the probability accumulation integral reaches 0.975, the corresponding integral upper limit is the optimal threshold M. The window length W is set to the packet transmission rate R_SHalf of (1);

step six: and (5) segmenting the action according to the sliding variance V, the adaptive threshold M and the window length W obtained in the step five.

2. The method as claimed in claim 1, wherein the step three includes applying a piecewise fitting de-trend method to the second principal component h obtained in the step two₂The vector T is obtained by processing the vector T,

firstly, the principal component h₂Dividing the data into a plurality of data segments with equal length, then carrying out polynomial fitting on each data segment by using a least square method, finally carrying out difference on each data segment and corresponding fitting data, and recombining the fitting error data in sequence to form a vector T. The specific process is as follows:

firstly, the second principal component h₂Divided into several data sections of length M, i.e.

h₂＝[h₂(W₁),h₂(W₂),L,h₂(W_k),L,h₂(W_N/M)]

Wherein, W_kIs an array [ kM, kM +1, L, (k +1) M-1]And indicates an index value of each data in the k-th data segment. Each piece of data was then fitted separately as follows:

T(W_k)＝h₂(W_k)-fit(h₂(W_k))

wherein fit (g) is a least squares based polynomial fit, resulting in:

T＝[T(W₁),T(W₂),L,T(W_k),L,T(W_N/M)]

the following is the implementation of polynomial fitting:

suppose there are m points in the data segment, each data point being (x)_i,y_i) The fitting result is:

wherein, a₀,L,a_nFor the unknowns to be solved, n is the highest power of the polynomial. According to the least-squares method, the sum of squares of deviations is minimized, i.e.

Solving for the above coefficients can be converted to I ═ I (a)₀,a₁,L,a_n) The extreme value problem (2) is that the necessary conditions for obtaining the extreme value according to the multivariate function are as follows:

namely, it is

Using a formal representation of the matrix then:

when the points needing fitting and the maximum times of fitting of the polynomial are given, all coefficients of the fitting polynomial can be solved, and then the points before fitting are used for subtracting the points after fitting to obtain a fitting residual error to form a vector T. Subsequent data processing is based on this data.

3. The Wi-Fi channel state information-based target behavior segmentation method according to claim 1, wherein in step six, the motion is segmented according to the sliding variance V, the adaptive threshold M and the window length W obtained in step five, and the method comprises the following steps:

and (4) obtaining a sliding variance V through data processing in the previous stage, and setting a threshold value M and a window length W to segment the motion. The segmentation algorithm comprises the following steps:

step 1: searching a maximum value MAX in the sliding variance, comparing the maximum value MAX with a threshold value M, if MAX is smaller than M, indicating that no action occurs in the data segment, otherwise, indicating that the action exists in the data segment, and performing Step 2;

step 2: the index indexMa of the MAX is recorded, while indexRt ═ indexLt ═ indexMax is defined. In the interval [ indexRt, indexRt + N]In search for V_lIndex value set L corresponding to the index value set larger than threshold value M₁If set L₁If the set is not empty, find the largest element in the set, assign the value of iMax to indexRt, and add indexRt to the new [ indexRt, indexRt + N [ ]]Interval finding V_lIndex value set L corresponding to the index value set larger than threshold value M₁. After the step, the ending time point indexRt of the action can be found;

step 3: in the interval [ indexLt-N, indexLt]In search for V_lIndex value set L corresponding to the index value set larger than threshold value M₂If set L₂If the set is not empty, find the smallest element iMin in the set, assign the value of iMin to indexLt, and add the value to the new [ indexLt-N, indexLt]Interval finding V_lIndex value set L corresponding to the index value set larger than threshold value M₂. After the step, the starting time point indexLt of the action can be found;

step 4: segment [ indexLt, indexRt]Corresponding V_lSetting to be 0, and then performing Step1, Step2 and Step 3;

through the steps, the data segment when the action occurs can be effectively extracted from the original data, and meanwhile, the algorithm is still effective for segmenting a plurality of discontinuous actions contained in the original data.

Images