CN113408476A - Human body posture identification method based on wireless network - Google Patents

Abstract

The invention relates to a human body posture identification method based on a wireless network. Because the sensor is complex to wear, and the camera is easily influenced by problems such as illumination, obstacles, monitoring dead angles, privacy disclosure and the like, an effective supplementary technology for human body Posture identification based on a Channel State Information-Segment position orientation value (CSI-SPE) is provided. The method comprises the steps of collecting posture information (running, jumping, sitting, standing, going up and down stairs) of a human body through WIFI routing equipment, conducting denoising and abnormal value processing through a Butterworth filter and a Hampel algorithm, then constructing a segment posture characteristic value, then transmitting the segment posture characteristic value into a Support Vector Machine (SVM) for identification, and experiments prove that the comprehensive identification precision of the CSI-SPE-based method is 98%, and the experiment results show that the CSI-SPE method has high robustness and identification precision.

Description

Human body posture identification method based on wireless network

The technical field is as follows:

the invention belongs to the field of wireless networks, human body posture recognition or machine learning.

Background art:

at present, the research on the problems related to human body posture recognition in China is closer and closer to daily life, and the research result of human body posture recognition is widely applied to the fields of action recognition, human-computer interaction, clothing analysis and the like. In the existing human posture recognition research, a person perception and intrusion detection system based on a WLAN is designed in the document [1], and similarity differences of different CSI phase matrices are utilized, but the method using eigenvalue mutation has the defect of instability due to being easily influenced by the surrounding environment. Document [2] constructs a four-dimensional eigenvector by extracting the largest and second largest eigenvalues of the correlation matrix of CSI amplitude and phase, but it remains to be verified whether the recognition rate will increase with the increase of the eigenvector. Document [3] proposes to extract subcarrier feature variances corresponding to different postures, and then classify the human behavior by using an SRC algorithm, but the method has poor stability and has an unsatisfactory detection effect under the condition of non-severe human activity. Document [4] omits a data preprocessing process to reduce cost, extracts a mean value and a variance of the amplitude of the original CSI data as features, and performs human perception by using an SVM, however, the original CSI data has interference information, and thus the method has an undesirable effect. The problems and challenges existing in human body posture recognition at present are as follows:

(1) with the continuous improvement of the requirements of people on motion recognition, the requirements of low cost, fine granularity and high precision are realized.

(2) In an indoor environment, how to reduce multipath effects and how to process various environmental noises brought by WIFI signals, and a complex application environment is processed.

Disclosure of Invention

1. The technical problems needed and solved by the invention are as follows:

the human body posture identification method based on the wireless network can overcome the influence of environmental noise, has the advantages of simple feature extraction and accurate and obvious feature identification rate compared with other identification methods, has fine granularity, can reduce the multipath effect, further improves the accuracy of human body posture identification, and ensures that the human body posture estimation efficiency is higher and the cost is lower.

2. The specific technical scheme of the invention is as follows:

1. collecting raw data

2. The acquired raw data set is matrix transformed.

3. And cleaning and processing the data set after matrix transformation.

4. And performing feature extraction on the cleaned and processed data set.

5. And carrying out human body posture recognition on the extracted features to obtain a posture recognition result.

A human body posture identification method based on a wireless network is characterized in that:

1) collecting an original data set:

(2) carrying out matrix transformation on the acquired original data set;

extracting sub-carriers through MATLAB software, wherein data _ f, data _ s and data _ t represent sub-carrier data of a first antenna, a second antenna and a third antenna;

F(x)＝x^T (1)

wherein x represents the incoming subcarrier data, T represents transposition, and data _ f, data _ s and data _ T are incoming for matrix transposition; carrying out flattening dimensionality reduction on the three groups of data obtained after transposition to obtain data with (1, n) dimensionality, wherein n represents the number of columns of the data; splicing the null library into a (3, n) matrix through a stack function of the numpy library, wherein n represents the column number of the data; splicing a user number, a user posture and a time stamp into a matrix to construct (6, n), and then constructing an original data set (r, 6) by using a formula 1, wherein r is a line number;

normalizing the data to change the data into data between 0 and 1, and applying a formula (2), wherein x is the incoming matrix data, MIN is the minimum value of each line of data, and MAX is the maximum value of each line of data;

(3) and (3) cleaning and processing the data set after matrix transformation:

a low-pass filter is designed, in the acquisition process, the transmission frequency of the filter designed in the text is 50Hz, and the environmental noise is 250Hz to 500Hz after being consulted; the low-pass filter is therefore designed to cut-off the frequency f_p30Hz, pass band maximum attenuation alpha_p3dB stop band starting frequency f_s250Hz, stop band minimum attenuation alpha_s＝30dB；

By the formula(3) Calculating normalized frequency lambda_pIs the cut-off frequency, λ_sIs the stop band start frequency, where Ω_p＝2πf_p，Ω_pPassband cut-off frequency, Ω_s＝2πf_s，Ω_sA passband start frequency;

of the order N and the parameter C of the Butterworth filter

Calculating by a formula (4) to obtain a parameter C of 1 and alpha as an intermediate variable;

calculating to obtain the order N of the filter as 5 orders;

calculating a parameter B and a parameter A through a button transfer function formula (7) in software

[B,A]＝butter(N,λ_s,λ_p,C) (7)

y is the output signal sequence, where x is the input signal, i.e., the three subcarrier sequences of the data set, where B is the numerator of the Butterworth filter, A is the denominator of the Butterworth filter, and the filter is the filter function in the software;

y＝filter(B,A,x) (8)

outputting the filtered signal in formula (8), and removing abnormal values of the input vector (three subcarriers); calculating the median value of a window consisting of the samples and six samples around the samples by using a Hampel function, wherein each side is provided with three samples; estimating standard deviation of each sample to the median by using the median absolute value to replace an abnormal value, and obtaining a cleaned and processed data set; the data set enabled by the filters and Hampel functions designed herein for this class of tasks improves the completeness of the data set relative to other methods;

(4) and (3) carrying out feature extraction on the cleaned and processed data set:

constructing an input characteristic value, creating a time window, advancing 45 records every time by 90-50 ms (namely 4.5 seconds), and adopting a half-overlapping mode; then creating input data, wherein each group comprises three continuous 90 pieces of data of subcarriers, and is called a segment attitude characteristic value, user activities in the 90 pieces of data are counted, and the label name with the largest occurrence frequency is the group of data;

(5) carrying out human body posture recognition on the extracted features to obtain a posture recognition result

Dividing a data set into a training set, a verification set and a test set; the specific SVM algorithm used herein is:

the training set data T is represented as:

T＝{(x1,y1),(x2,y2),...,(xn,yn)} (9)

xi belongs to a multidimensional space and y has a value of one of 0,1,2,3,4, and n is 1,2, …, n; n is an integer of 1 or more; xi is the ith feature vector, yi is the label name (run 0, sit 1, stand 2, go up and down stairs 3, walk 4);

constructing and solving a convex quadratic programming problem, where α_i,α_jIndicates correspondence to y_iy_jCoefficient of (a)^*Expressing the optimal solution, wherein the physical meaning of alpha is to find an optimal solution to satisfy the formula (10), and the optimal hyperplane is found through continuous optimization of alpha, so that the limit of the classification problem faced by the SVM is obtained:

0＜α_i＜M,i＝1,2,3...n (12)

equations 10,11 and 12 find the optimal solution

Wherein

Representing the function y_iMiddle corresponding independent variable x_iHas n in total, formula 13 calculates coefficient b^*，b^*Representing the function y_iThe number of the optimal solutions corresponding to the constant is n in total;

calculated alpha^*，b^*Substituting formula (15)

Solving for alpha using a gaussian kernel with a penalty factor M of 8^*Using a gradient descent method, a Gaussian kernel function K (x)_i,x_j) As shown in the formula; the physical meaning of K is a function representative of a following equation, and two independent variables are shared; wherein sigma²Is x_i,x_jStandard deviation of (d);

and f (x) is solved, the result is compared with the true value label, when f (x) is 0, the matching training result is proved to be running, when f (x) is 1, the matching training result is proved to be sitting, when f (x) is 2, the matching training result is proved to be standing, when f (x) is 3, the matching training result is proved to be ascending and descending stairs, and when f (x) is 4, the matching training result is proved to be running.

3. The invention can achieve the following effects:

from the economic point of view: the method provides a significant cost reduction over conventional sensor systems and computer vision systems. Especially, with the popularization of wireless network and 5G technology, the technology can be deep into each family. The building is simple, and the required equipment is basically possessed by common families.

From a social perspective: compared with the traditional sensor system and the computer vision system, the method greatly protects the privacy of the individual. In a computer vision system, a video stream can carry facial features, body features and environmental features of an individual, and if the video stream is leaked, all personal feature information can be exposed to the outside, so that great hidden danger is generated. The method only needs to acquire the personal posture through a wireless network, and does not need to acquire the facial features, the environmental features and the like of the human body. The safety is strong.

From the technical point of view: compared with the traditional sensor system and the traditional calculator vision system, the method is simple to build, and can be built by users familiar with using a Linux system and a Windows system.

Description of the drawings:

FIG. 1 Experimental Environment

FIG. 2 Experimental flow

FIG. 3 confusion matrix

Detailed Description

The experimental environment is a common empty laboratory, wherein the distance between the transmitting end and the receiving end is 1.5 meters, and the experimenter only needs to express the posture between the transmitting end and the receiving end.

Fig. 3 shows the recognition conditions corresponding to the five postures of running 0, sitting 1, standing 2, going up and down stairs 3 and walking 4, which embodies a very accurate recognition condition.

TABLE 1 accuracy of different classifiers

Table 1 shows the accuracy results obtained by using different classifiers, wherein the classification effect of SVC is the best. Also as used herein.

TABLE 2 comparison of different methods

Table 2 shows the accuracy of different experiments, with the highest accuracy of the CSI-SPE method based on the segment attitude eigenvalues herein. Obviously superior to other methods.

Table 3 data set sample form after cleaning and processing

Table 3 shows the data set format of the present invention after the embodiment (2).

1) Collecting an original data set:

1. a host with a Linux Ubuntu system installed, an Intel 5300 network card, three receiving antennas and a router supporting 802.11n protocols are prepared. The CSI TOOLS are installed on the host and the router is then configured in AP mode. Then, the host sends out a ping command, receives data through the antenna and stores the data into a file.

(2) The acquired raw data set is matrix transformed.

And extracting the subcarriers through MATLAB software, wherein the data _ f, the data _ s and the data _ t represent subcarrier data of the first antenna, the second antenna and the third antenna.

F(x)＝x^T (1)

Wherein x represents the incoming subcarrier data, T represents transposition, and data _ f, data _ s, and data _ T are incoming for matrix transposition. And carrying out flattening and dimension reduction on the three groups of data obtained after transposition to obtain data with (1, n) dimensions, wherein n represents the number of columns of the data. They are then spliced into a (3, n) matrix by the stack function of the numpy library, where n represents the number of columns of data. Further, a user number (UserId), a user gesture (Activity) and a Timestamp (Timestamp) are spliced into the matrix to construct (6, n), and then an original data set (n, 6) is constructed by applying formula 1, wherein n is the number of lines, and n is 30000 in the data set.

Further, the data is normalized to change the data to data between 0 and 1, using equation 2, where x is the incoming matrix data, MIN is the minimum value of each column of data, and MAX is the maximum value of each column of data.

(3) And (3) cleaning and processing the data set after matrix transformation:

further, the data obtained in the step (2) is cleaned and processed, wherein a low-pass filter is designed, the transmission frequency of the filter designed in the step (2) is 50Hz during the acquisition process, and the environmental noise is 250Hz to 500Hz after being consulted. The low-pass filter is therefore designed to cut-off the frequency f_p30Hz, pass band maximum attenuation alpha_p3dB stop band starting frequency f_s250Hz, stop band minimum attenuation alpha_s＝30dB。

Further, the normalized frequency λ is calculated by formula 3_pIs the cut-off frequency, λ_sIs the stop band start frequency, where Ω_p＝2πf_p，Ω_pPassband cut-off frequency, Ω_s＝2πf_s，Ω_sThe passband start frequency.

Further calculating the order N and the parameter C of the Butterworth filter

The parameter C is 1 and alpha is an intermediate variable obtained by calculation of formula 4.

The order N of the filter is 5 order calculated by formula 5 and formula 6

Further, through a button transfer function (formula 7) in the software, a parameter B and a parameter A are calculated

[B,A]＝butter(N,λ_s,λ_p,C) (7)

The further y is the output signal sequence, where x is the input signal, i.e. the three subcarrier sequences of the data set, where B is the numerator of the butterworth filter, a is the denominator of the butterworth filter, and the filter is the filter function in the software.

y＝filter(B,A,x) (8)

Further, the filtered signal is output in equation 8, and the outlier removing operation is performed on the input vector (three subcarriers). The Hampel function computes the median of a window of six samples around and samples, three on each side. And replacing the abnormal value by the standard deviation of the median value of each sample pair estimated by the median absolute value to obtain a cleaned and processed data set. The data set enabled by the filters and Hampel functions designed herein for this type of task improves the completeness of the data set relative to other methods.

(4) And (3) carrying out feature extraction on the cleaned and processed data set:

further, on the basis of (3), the input characteristic value is constructed, a time window is created, and each time 45 records are advanced by 90 times 50ms, namely 4.5 seconds, and a half-overlapping mode is adopted. Then, input data is created, each group comprises 90 continuous pieces of data of three subcarriers, and the data is called a Segment attitude characteristic value (CSI-SPE), wherein user activities in the 90 pieces of data are counted, and one with the largest occurrence frequency is the tag name of the group of data.

(5) Carrying out human body posture recognition on the extracted features to obtain a posture recognition result

Further, on the basis of (4), the data set is divided into a training set, a verification set and a test set according to the ratio of 7:2: 1. The specific SVM algorithm used herein is:

the training set data T is represented as:

T＝{(x1,y1),(x2,y2),...,(xn,yn)} (9)

in formula 9, xi belongs to an n-dimensional space, and y has a value of one of 0,1,2,3, and 4, and n is 1,2, …, n; n is an integer of 1 or more; xi is the ith feature vector, yi is the label name (run 0, sit 1, stand 2, go up and down stairs 3, walk 4). In the present invention, the feature vector is composed of a Segment attitude feature value (CSI-SPE)

2. Selecting appropriate kernel function and penalty parameter M>0, constructing and solving a convex quadratic programming problem, where α_i,α_jIndicates correspondence to y_iy_jCoefficient of (a)^*Expressing the optimal solution, wherein the physical meaning of alpha is to find an optimal solution to satisfy the formula 10, and the optimal hyperplane is found through continuous optimization of alpha, so that the limit of the classification problem faced by the SVM is obtained:

0＜α_i＜M,i＝1,2,3...n (12)

equations 10,11,12 find the optimal solution (typically using a gradient descent method)

Wherein

Representing the function y_iMiddle corresponding independent variable x_iHas n in total, formula 13 calculates coefficient b^*，b^*Representing the function y_iThe number of the optimal solutions corresponding to the constants is n in total.

Alpha calculated by equations 13 and 14^*，b^*Substituting into equation 15

The invention uses a Gaussian kernel function, the penalty coefficient M is 8, and alpha is solved^*Using a gradient descent method, a Gaussian kernel function K (x)_i,x_j) As shown in equation 16. The physical meaning of K is that the function refers to the following equation, sharing two independent variables. Wherein sigma²Is x_i,x_jStandard deviation of (2).

And f (x) is solved, the result is compared with the true value label, when f (x) is 0, the matching training result is proved to be running, when f (x) is 1, the matching training result is proved to be sitting, when f (x) is 2, the matching training result is proved to be standing, when f (x) is 3, the matching training result is proved to be ascending and descending stairs, and when f (x) is 4, the matching training result is proved to be running.

(6) And (5) verifying the result:

the method adopts the confusion matrix and the secondary indexes of the confusion matrix for verification, has obvious advantages compared with other technologies referred by the text, the filter and the characteristic extraction method designed by the text are superior to other methods, and the result accuracy rate reaches about 98% after a plurality of experiments.

Reference documents:

[1].LI L X.Research and Design of Personnel Perception and Intrusion Detection Technology Based on WLAN[D].Beijing:Beijing University of Posts and Telecommunications,2018.

[2].QIAN K,WU C,YANG Z,et al.Enabling Contactless Detection of Moving Humans with Dynamic Speeds Using CSI[J].ACM Transactions on Embedded Computing Systems,2018,17(2):1-18.

[3].XIAO L,PAN H.Human motion recognition system based on Wi-Fi signal[J].Journal of Beijing University of Posts and Telecommunications,2018,41(3):119-124.

[4].WANG T,YANG D D,ZHANG S Q,et al.Wi-Alarm:Low-Cost Passive Intrusion DetectionUsing Wi-Fi[J].Sensors,2019,19(10):2335.

[5] haohanshi, Zhang Dayang, Dang Xiao super, Chun Yu, a non-contact human action recognition method based on CSI [ J/OL ] computer engineering 1-13[2021-03-07], https:// doi.org/10.19678/j.issn.1000-3428.0057612.

[6] Yin heptere, dingwen super, zhang junbao. CSI action recognition based on time convolution network [ J ]. proceedings of the central institute of technology, 2020,31(05):59-65.

[7] Guoargin, xu Zhi Meng, Chen Liang Qin, a human body action recognition method based on WiFi channel state information [ J ] the technical report of sensing, 2019,32(11): 1688-1693.

Claims (1)

1. A human body posture identification method based on a wireless network is characterized in that:

1) collecting an original data set:

(2) carrying out matrix transformation on the acquired original data set;

extracting sub-carriers through MATLAB software, wherein data _ f, data _ s and data _ t represent sub-carrier data of a first antenna, a second antenna and a third antenna;

F(x)＝x^T (1)

wherein x represents the incoming subcarrier data, T represents transposition, and data _ f, data _ s and data _ T are incoming for matrix transposition; carrying out flattening dimensionality reduction on the three groups of data obtained after transposition to obtain data with (1, n) dimensionality, wherein n represents the number of columns of the data; splicing the null library into a (3, n) matrix through a stack function of the numpy library, wherein n represents the column number of the data; splicing a user number, a user posture and a timestamp into a matrix to construct (6, n), and then constructing an original data set (r, 6) by using a formula 1, wherein r is a line number;

normalizing the data to change the data into data between 0 and 1, and applying a formula (2), wherein x is the incoming matrix data, MIN is the minimum value of each line of data, and MAX is the maximum value of each line of data;

(3) and (3) cleaning and processing the data set after matrix transformation:

a low-pass filter is designed, in the acquisition process, the transmission frequency of the filter designed in the text is 50Hz, and the environmental noise is 250Hz to 500Hz after being consulted; the low-pass filter is therefore designed to cut-off the frequency f_p30Hz, pass band maximum attenuation alpha_p3dB stop band starting frequency f_s250Hz, stop band minimum attenuation alpha_s＝30dB；

Calculating the normalized frequency λ by equation (3)_pIs the cut-off frequency, λ_sIs the stop band start frequency, where Ω_p＝2πf_p，Ω_pPassband cut-off frequency, Ω_s＝2πf_s，Ω_sA passband start frequency;

filter order N and parameter C for calculating butterworth

Calculating by a formula (4) to obtain a parameter C of 1 and alpha as an intermediate variable;

calculating to obtain the order N of the filter as 5 orders;

calculating a parameter B and a parameter A through a button transfer function formula (7) in software

[B，A]＝butter(N，λ_s，λ_p，C) (7)

y is the output signal sequence, where x is the input signal, i.e., the three subcarrier sequences of the data set, where B is the numerator of the Butterworth filter, A is the denominator of the Butterworth filter, and the filter is the filter function in the software;

y＝filter(B，A，x) (8)

outputting the filtered signal in formula (8), and removing abnormal values of the input vector (three subcarriers); calculating the median value of a window consisting of the samples and six samples around the samples by using a Hampel function, wherein each side is provided with three samples; estimating standard deviation of each sample to the median by using the median absolute value to replace an abnormal value, and obtaining a cleaned and processed data set; the data set enabled by the filters and Hampel functions designed herein for this class of tasks improves the completeness of the data set relative to other methods;

(4) and (3) carrying out feature extraction on the cleaned and processed data set:

constructing an input characteristic value, creating a time window, advancing 45 records every time by 90-50 ms (namely 4.5 seconds), and adopting a half-overlapping mode; then creating input data, wherein each group comprises three continuous 90 pieces of data of subcarriers, and is called a segment attitude characteristic value, user activities in the 90 pieces of data are counted, and the label name with the largest occurrence frequency is the group of data;

(5) carrying out human body posture recognition on the extracted features to obtain a posture recognition result

Dividing a data set into a training set, a verification set and a test set; the specific SVM algorithm used herein is:

the training set data T is represented as:

T＝{(x1，y1)，(x2，y2)，...，(xn，yn)} (9)

xi belongs to a multidimensional space and y has a value of one of 0,1,2,3,4, and n is 1,2, …, n; n is an integer of 1 or more; xi is the ith feature vector, yi is the label name (run 0, sit 1, stand 2, go up and down stairs 3, walk 4);

constructing and solving a convex quadratic programming problem, where α_i，α_jIndicates correspondence to y_iy_jCoefficient of (a)^*Expressing the optimal solution, wherein the physical meaning of alpha is to find an optimal solution to satisfy the formula (10), and the optimal hyperplane is found through continuous optimization of alpha, so that the limit of the classification problem faced by the SVM is obtained:

0＜α_i＜M，i＝1，2，3...n (12)

equations 10,11 and 12 find the optimal solution

Wherein

Representing the function y_iMiddle corresponding independent variable x_iHas n in total, formula 13 calculates coefficient b^*，b^*Representing the function y_iThe number of the optimal solutions corresponding to the constant is n in total;

calculated alpha^*，b^*Substituting formula (15)

Solving for alpha using a gaussian kernel with a penalty factor M of 8^*Using a gradient descent method, a Gaussian kernel function K (x)_i，x_j) As shown in the formula; the physical meaning of K is a function representative of a following equation, and two independent variables are shared; wherein sigma²Is x_i，x_jStandard deviation of (d);

and f (x) is solved, the result is compared with the true value label, when f (x) is 0, the matching training result is proved to be running, when f (x) is 1, the matching training result is proved to be sitting, when f (x) is 2, the matching training result is proved to be standing, when f (x) is 3, the matching training result is proved to be ascending and descending stairs, and when f (x) is 4, the matching training result is proved to be running.

Images