CN113901863A - Human activity classification method based on weighted group sparse Bayesian learning - Google Patents

Abstract

The invention relates to a human activity classification method based on weighted group sparse Bayesian learning, and belongs to the technical field of radar and pattern recognition. The method comprises the steps of preprocessing received human body activity radar echo signals by adopting short-time Fourier transform; extracting features by a principal component analysis method; and carrying out sparse coding on the human activity test sample by using a weighted group sparse Bayesian learning algorithm, and then carrying out classification and identification on the human activity based on a residual minimum criterion. The method takes label information of training samples into consideration, so that sparse representation coefficients have the characteristic of a group structure, and the classification accuracy is improved; the Bayesian model considers the influence of noise, has good adaptability to the actual environment and can steadily realize human activity classification; compared with the traditional method, the method has better classification performance under the noisy condition.

Description

Human activity classification method based on weighted group sparse Bayesian learning

Technical Field

The invention relates to a human activity classification method based on weighted group sparse Bayesian learning, and belongs to the technical field of radar and pattern recognition.

Background

With the research and application of the identification and classification of human activities in various fields such as security monitoring, remote health monitoring and the like, the radar-based human activity classification has the advantages of not invading human privacy, effectively overcoming some inherent technical defects existing in audio, video, infrared and other life detection, along with low cost and easy deployment, and becomes a hotspot of research and development at home and abroad.

Human activities can be viewed as complex targets, as a wide variety of activities produce complex limb movements, and the time-varying trajectories of various parts of the body will be reflected in the radar returns. The radar micro-Doppler characteristics of human body movement are obtained by performing time-frequency transformation on the received radar echo signals, and actually are Doppler time-frequency spectrums obtained by overlapping the movement echoes of a plurality of human body parts in a given observation time. Because the radar Doppler history data only carries radial velocity information, although a human body motion model cannot be reconstructed from the micro Doppler features, because the motion modes of all parts of a human body have differences when the human body moves, the Doppler frequencies generated by the motion of all parts of the human body are different, different human body activities can generate different rich and unique micro Doppler features, namely the radar micro Doppler features carry the feature information of the motion of the human body, and the human body and the motion thereof can be classified and identified by utilizing the micro Doppler features.

In recent years, the sparse representation theory has attracted the interest of researchers in the field of signal processing, and has been effectively applied to a plurality of related problems. John Wright and the like apply Sparse Representation to facial image recognition, and propose a facial image recognition method based on Sparse Representation-based Classification (SRC). The method obtains better recognition effect under the condition that the face image is shielded or polluted by noise and other interferences. The SRC method is considered to be used for human activity recognition, but in practical situations, data acquisition of human activity is affected by the environment where the data acquisition is located, and has more clutter and noise, and the anti-noise performance of the existing algorithm needs to be improved.

Disclosure of Invention

The invention aims to solve the problem that the classification recognition rate is reduced due to the fact that data collected in radar human body activity classification is influenced by noise in the actual situation, and provides a human body activity classification method based on weighted-group sparse Bayesian learning.

In order to achieve the purpose, the invention adopts the following technical scheme:

the human activity classification method based on weighted group sparse Bayesian learning comprises the following steps:

step 1: collecting radar echo signals of human body activities through a radar antenna;

step 2: clutter suppression and short-time Fourier transform are carried out on the radar echo signals to obtain a time-frequency diagram of the radar echo signals, and the method specifically comprises the following steps: clutter suppression is performed firstly, and then the window length is N_wObtaining a time-frequency diagram of the radar echo signal by the short-time Fourier transform;

and step 3: compressing the time-frequency image and converting the time-frequency image into a gray image, reading data of the gray image, elongating the data into one-dimensional vectors according to columns, and extracting features by using a principal component analysis method to obtain feature data of all human activities; dividing all the characteristic data of human body activity into two groups according to a proportion, wherein one group is used as training data and verification data, and the other group is used as test data, namely a test sample; randomly extracting part of data samples in the first group as training samples, placing the training samples of the same class according to columns to form a sub-dictionary, and forming a dictionary A by all the sub-dictionaries;

wherein, the dictionary A is recorded as: a ═ A₁,A₂,…,A_K) (ii) a The c-th sub-dictionary is noted as: a. the_cAnd c is 1,2, …, K is the number of sample categories in dictionary a;

and 4, step 4: carrying out sparse coding on the human activity test sample in a dictionary A to obtain a sparse representation coefficient, and specifically adopting a weighted group sparse Bayesian learning algorithm to calculate the sparse representation coefficient, wherein the method comprises the following substeps:

step 4.1: establishing a noisy sparse representation model y as Ax + w;

wherein x is a sparse representation coefficient of a human activity test sample y to be solved, A is a dictionary A, w is a mean value of 0 and a variance of beta^-1White noise subject to a gaussian distribution,

where m is the dimension of the test sample y; x obeys gaussian prior distribution

Wherein x is (x)₁,x₂,…,x_n)^T，α_i ^-1Is an element x_iVariance of (a) { α ═ α_iThe coefficient x is a non-negative over-parameter for controlling the sparsity of the sparse representation coefficient x; x is the number of_iAnd α ═ α_iThe value range of the subscript i is 1 to n, n is the column number of the dictionary A, and the superscript T represents the transposition operation of the matrix;

step 4.2: setting initial values of alpha and beta, setting the initialization iteration number iter to be 1, setting the maximum iteration number Max _ iter, and initializing mu⁰＝A^Hy；

Wherein, mu⁰The initial value of the posterior mean value of the sparse representation coefficient x is used, and the superscript H represents the conjugate transpose operation of the matrix;

step 4.3: calculating the covariance Σ, mean μ of the posterior probability distribution of the sparse representation coefficient x:

Σ＝(βA^HA+D)^-1，μ＝βΣA^Hy

wherein, the posterior probability distribution of the sparse representation coefficient x obeys the Gaussian distribution when the human activity test sample y is given, and D ═ diag (alpha)₁,α₂,…,α_n)；

Step 4.4: assuming that the corresponding sparse coefficient variances of the same type of samples are the same, carrying out weighted average on posterior information of the same type of samples, and estimating an element x_iThen the obtained inverse variance estimation of the same type sample is carried outWeighted average, updating element x_iThe specific steps are as follows:

step 4.4 a: estimating an element x_iThe inverse variance of (c):

wherein the content of the first and second substances,

is an element x_iEstimate of the inverse of the variance, n_iIs x_iThe column number of the corresponding sub-dictionary;

the weighting factor lambda belongs to [0,1 ]]G (i) represents and x_iAn element index belonging to the same class of tags;

step 4.4 b: carrying out weighted average on the obtained inverse variance estimation of the same type sample, and updating the element x_iThe inverse variance of (c):

step 4.5: let mu let^iterμ and iter + 1;

step 4.6: if iter is less than or equal to Max _ iter and the current mean value mu^iterAnd the last mean value mu^iter-1With | | | mu between^iter-μ^iter-1||₂>E, repeating steps 4.3 to 4.5 until iter>Max _ iter or current mean μ^iterAnd the last mean value mu^iter-1Satisfies | | mu between^iter-μ^iter-1||₂Less than or equal to epsilon, the current posterior mean being the sparse representation coefficient obtained, i.e.

Wherein | · | purple sweet₂Represents a 2-norm;

from step 4.1 to step 4.6, the sparse coding of the human activity test sample is completed, and a sparse representation coefficient for carrying out the sparse coding on the human activity test sample is obtained;

and 5: based on the minimum residual classification, specifically: calculating the residual r between the human activity test sample y and the linear weighted sum of each type of sub-dictionary_c(y)：

Wherein the content of the first and second substances,

to represent

Keeping the coefficient corresponding to the middle-class-c sub-dictionary unchanged, and setting other coefficients to be zero; when the residual value of a certain class is minimum, y is considered to belong to the class.

Advantageous effects

Compared with the prior art, the invention provides a human activity classification method based on weighted group sparse Bayesian learning, which has the following beneficial effects:

1. compared with the existing human activity classification method, the human activity classification method adopts the Bayesian model for modeling when noise influence exists, can better solve the problem of effective identification of human activity in a noisy environment, and has certain noise robustness;

2. compared with the existing human activity classification method, the human activity classification method takes the label information of the training samples into consideration, so that the sparse representation coefficient has the characteristic of group structure and has better classification performance.

Drawings

FIG. 1 is a schematic block diagram of a human activity classification method based on weighted-group sparse Bayesian learning according to the present invention;

FIG. 2 is a time-frequency diagram of radar echo signals of six human activities, namely (2a) walking, (2b) sitting, (2c) standing, (2d) picking up objects, (2e) drinking water, and (2f) falling, respectively;

FIG. 3 is a schematic diagram of a signal parameter model for weighted-group sparse Bayesian learning according to the present invention;

FIG. 4 is a graph of the classification results of each classifier for five random training samples;

FIG. 5 is a classification result confusion matrix of the human activity classification method based on weighted group sparse Bayesian learning.

Detailed Description

The following describes the implementation process of the human activity classification method based on weighted group sparse bayes learning with reference to the accompanying drawings.

Example 1

The fields of public safety monitoring and indoor human body monitoring, etc. can be based on radar monitoring of personnel activity patterns, i.e. collecting human body activity data using radar and detecting human body major activity events such as falls using the present invention. The invention adopts the Bayesian model for modeling, and can better deal with noisy environment; the introduction of group sparsity effectively improves the classification performance.

The method comprises the steps of preprocessing received human body activity radar echo signals by adopting short-time Fourier transform; extracting features by a principal component analysis method; carrying out sparse coding on the human activity test sample by using a weighted group sparse Bayesian learning algorithm; and classifying and identifying the human body activity based on the residual minimum criterion. The classification method takes the label information of the training samples into consideration, so that the sparse representation coefficient has the characteristic of a group structure, and the classification accuracy is improved; moreover, the Bayesian model takes the influence of noise into consideration, has better adaptability to the actual environment and can steadily realize human activity classification; compared with the traditional method, the method has better classification performance under the noisy condition.

Fig. 1 is a flowchart of a specific implementation of the human activity classification method based on weighted group sparse bayes learning, which comprises the following steps:

step 1: collecting radar echo signals of human body activities through a radar antenna;

step 2: clutter suppression and short-time Fourier transform are carried out on the radar echo signals to obtain a time-frequency diagram of the radar echo signals, and the method specifically comprises the following steps: firstly, a Butterworth high-pass filter is used for suppressing clutter, and then the window length is N_wObtaining a time-frequency diagram of the radar echo signal by the short-time Fourier transform;

the time-frequency diagram is marked as STFT (f, t), and f and t respectively represent frequency components and time components obtained by short-time Fourier transform;

and step 3: compressing the time-frequency image and converting the time-frequency image into a gray image, reading data of the gray image, elongating the data into one-dimensional vectors according to columns, and extracting features by using a principal component analysis method to obtain feature data of all human activities; dividing all the characteristic data of human body activity into two groups according to a proportion, wherein one group is used as training data and verification data, and the other group is used as test data, namely a test sample; randomly extracting part of data samples in the first group as training samples, placing the training samples of the same class according to columns to form a sub-dictionary, and forming a dictionary A by all the sub-dictionaries;

wherein, the dictionary A is recorded as: a ═ A₁,A₂,…,A_K) (ii) a The c-th sub-dictionary is noted as: a. the_cAnd c is 1,2, …, K is the number of sample categories in dictionary a;

and 4, step 4: carrying out sparse coding on the human activity test sample y under the dictionary A to obtain a sparse representation coefficient, and specifically adopting a weighted group sparse Bayesian learning algorithm to calculate the sparse representation coefficient, wherein the method comprises the following substeps:

step 4.1: establishing a noisy sparse representation model y as Ax + w;

wherein x is a sparse representation coefficient of a human activity test sample y to be solved, A is a dictionary A, w is a mean value of 0 and a variance of beta^-1White noise subject to a gaussian distribution,

where m is the dimension of the test sample y; x obeys gaussian prior distribution

Wherein x ═(x₁,x₂,…,x_n)^T，α_i ^-1Is an element x_iVariance of (a) { α ═ α_iThe coefficient x is a non-negative over-parameter for controlling the sparsity of the sparse representation coefficient x; x is the number of_iAnd α ═ α_iThe value range of the subscript i is 1 to n, n is the column number of the dictionary A, and the superscript T represents the transposition operation of the matrix;

step 4.2: setting initial values of alpha and beta, setting the initialization iteration number iter to be 1, setting the maximum iteration number Max _ iter, and initializing mu⁰＝A^Hy；

Wherein, mu⁰The initial value of the posterior mean value of the sparse representation coefficient x is used, and the superscript H represents the conjugate transpose operation of the matrix;

step 4.3: calculating the covariance Σ, mean μ of the posterior probability distribution of the sparse representation coefficient x:

Σ＝(βA^HA+D)^-1，μ＝βΣA^Hy

wherein, the posterior probability distribution of the sparse representation coefficient x obeys the Gaussian distribution when the human activity test sample y is given, and D ═ diag (alpha)₁,α₂,…,α_n)；

Step 4.4: assuming that the corresponding sparse coefficient variances of the same type of samples are the same, carrying out weighted average on posterior information of the same type of samples, and estimating the sparse coefficient x corresponding to each type of samples_cThen, the weighted average is carried out on the obtained estimation of the inverse variance of the same type of samples, and the sparse coefficient x corresponding to each type of samples is updated_cThe specific steps are as follows:

step 4.4 a: estimating sparse coefficient x corresponding to class c sample_cThe inverse variance of (c):

wherein x is_cFor sparse representation of coefficient x and sub-dictionary A_cThe corresponding sparse coefficient is set to be a sparse coefficient,

is a sub-vector x_cEstimate of the inverse of the variance, N_cIs x_cCorresponding sub-dictionary A_cThe number of columns;

wherein the content of the first and second substances,

size N_c×N_c；

Wherein, the sub-dictionary A_cThe corresponding sparse coefficient is noted as:

μ_cis x_cThe average value of (a) of (b),

x_cthe covariance of (d) is noted as: sigma_c，σ_c ²Is sigma_cA column vector of diagonal elements;

step 4.4 b: carrying out weighted average on the obtained inverse variance estimation of the same type of samples, and updating the sparse coefficient x corresponding to the class c sample_cThe inverse variance of (c):

step 4.5: let mu let^iterμ and iter + 1;

step 4.6: if iter is less than or equal to Max _ iter and the current mean value mu^iterAnd the last mean value mu^iter-1With | | | mu between^iter-μ^iter-1||₂>E, repeating steps 4.3 to 4.5 until iter>Max _ iter or current mean μ^iterAnd the last mean value mu^iter-1Satisfies | | mu between^iter-μ^iter-1||₂Less than or equal to epsilon, the current posterior mean being the sparse representation coefficient obtained, i.e.

Wherein | · | purple sweet₂Represents a 2-norm;

from step 4.1 to step 4.6, the sparse coding of the human activity test sample is completed, and a sparse representation coefficient for carrying out the sparse coding on the human activity test sample is obtained;

and 5: based on the minimum residual classification, specifically: calculating the residual r between the human activity test sample y and the linear weighted sum of each type of sub-dictionary_c(y)：

Wherein the content of the first and second substances,

to represent

Keeping the coefficient corresponding to the middle-class-c sub-dictionary unchanged, and setting other coefficients to be zero;

when the residual value of a certain class is minimum, y is considered to belong to the class.

The present invention will be further described with reference to the following examples.

In this embodiment, a radar is selected to actually measure the human body micro doppler signals for classification and identification, and specifically, human body activity data of project Intelligent RF Sensing for Falls and Health Prediction is collected by an FMCW radar of incotek corporation, and the FMCW radar works in a C-band (5.8GHz) with a bandwidth of 400 MHz. The subject demonstrated six human activities: walk, sit, stand, pick up items, drink water and fall, with the subject demonstrating 2 to 3 times for each activity.

And performing time-frequency analysis on the radar echo signals of the human body activity by using short-time Fourier transform to obtain a time-frequency graph of the radar echo signals. Fig. 2 shows a time-frequency diagram of radar echo signals acquired under six activities of the same subject. The data set comprises 1700 time-frequency graphs, wherein 75% of data are used as training data and verification data, including 225 time-frequency graphs and 150 time-frequency graphs of falling of people walking, sitting, standing, picking up objects and drinking water, 75 time-frequency graphs are randomly selected as training samples and verification samples according to the ratio of 1:1 for each activity in the data set, and optimal parameters are selected for identification of test data through multiple experiments; and 25% of data are used as test data, including 75 time frequency graphs of walking, sitting, standing, picking up objects and drinking water and 50 time frequency graphs of falling respectively. The invention compares the human activity recognition accuracy with a Support Vector Machine (SVM) and a sparse representation classifier (SRC _ OMP) adopting an OMP algorithm by taking the human activity recognition accuracy as an experimental index. Support vector machine SVM is a commonly used classification algorithm. The SRC _ OMP adopts an OMP algorithm to realize the solution of the sparse coefficient, but the sparse coefficient in the class and the sparse coefficient between the classes are not subjected to distinguishing constraint. In order to obtain better classification recognition rate, the invention adopts a weighted group sparse Bayesian learning algorithm (SRC _ WGSBL), considers the label information of training samples in the process of calculating the sparse coefficient, and estimates the sparse coefficient by using the property of group sparsity, thereby improving the classification performance. Fig. 3 shows the relationship between the signal parameters of the weighted-group sparse bayesian learning proposed by the present invention. In this embodiment, the sparsity in SRC _ OMP is set to 5, the maximum iteration number in SRC _ WGSBL is set to 100, and the following optimal parameters are selected for multiple experiments: the weighting parameter λ is 0.005 and the inverse noise variance β is 100. Fig. 4 shows the comparison of the results of five experiments of the inventive algorithm and the compared classification algorithm, which vividly shows the improvement of the recognition accuracy of the inventive algorithm and the compared two classification algorithms.

TABLE 1 human activity recognition rates of the algorithm of the present invention and two comparative classification algorithms

Table 1 lists the human activity classification recognition rates of the inventive algorithm and the comparative two classification algorithms, and the data in the table are the average values of five experiments. Fig. 5 shows a confusion matrix of the classification results of the algorithm of the present invention, and the recognition accuracy under specific six human activities can be observed.

This specification presents a specific embodiment for the purpose of illustrating the context and method of practicing the invention. The details introduced in the examples are not intended to limit the scope of the claims but to aid in the understanding of the process described herein. Those skilled in the art will understand that: various modifications, changes or substitutions to the preferred embodiment steps are possible without departing from the spirit and scope of the invention and its appended claims. Therefore, the present invention should not be limited to the disclosure of the preferred embodiments and the accompanying drawings.

Claims (5)

1. The human activity classification method based on weighted group sparse Bayesian learning is characterized by comprising the following steps: the method comprises the following steps:

step 1: collecting radar echo signals of human body activities through a radar antenna;

step 2: performing clutter suppression and short-time Fourier transform on the radar echo signal to obtain a time-frequency diagram of the radar echo signal;

and step 3: compressing the time-frequency image and converting the time-frequency image into a gray image, reading data of the gray image, elongating the data into one-dimensional vectors according to columns, and extracting features by using a principal component analysis method to obtain feature data of all human activities; dividing all the characteristic data of human body activity into two groups according to a proportion, wherein one group is used as training data and verification data, and the other group is used as test data, namely a test sample; randomly extracting part of data samples in the first group as training samples, placing the training samples of the same class according to columns to form a sub-dictionary, and forming a dictionary A by all the sub-dictionaries;

and 4, step 4: carrying out sparse coding on the human activity test sample in a dictionary A to obtain a sparse representation coefficient, and specifically calculating the sparse representation coefficient by adopting a weighted group sparse Bayesian learning algorithm;

and step 4, comprising the following substeps:

step 4.1: establishing a noisy sparse representation model y as Ax + w;

wherein x is a sparse representation coefficient of a human activity test sample y to be solved, A is a dictionary A, w is a mean value of 0 and a variance of beta^-1White noise subject to a gaussian distribution,

where m is the dimension of the test sample y; x obeys gaussian prior distribution

Wherein x is (x)₁，x₂，...，x_n)^T，α_i ^-1Is an element x_iVariance of (a) { α ═ α_iThe coefficient x is a non-negative over-parameter for controlling the sparsity of the sparse representation coefficient x; x is the number of_iAnd α ═ α_iThe value range of the subscript i is 1 to n, and n is the column number of the dictionary A;

step 4.2: setting initial values of alpha and beta, setting the initialization iteration number iter to be 1, setting the maximum iteration number Max _ iter, and initializing mu⁰＝A^Hy；

Wherein, mu⁰The initial value of the posterior mean value of the sparse representation coefficient x is used, and the superscript H represents the conjugate transpose operation of the matrix;

step 4.3: calculating the covariance Σ, mean μ of the posterior probability distribution of the sparse representation coefficient x:

∑＝(βA^HA+D)^-1，μ＝β∑A^Hy

wherein, the posterior probability distribution of the sparse representation coefficient x obeys the Gaussian distribution when the human activity test sample y is given, and D ═ diag (alpha)₁，α₂，...，α_n)；

Step 4.4: assuming that the corresponding sparse coefficient variances of the same type of samples are the same, carrying out weighted average on posterior information of the same type of samples, and estimating an element x_iThen carrying out weighted average on the obtained inverse variance estimation of the same type sample, and updating the element x_iVariance of (2)Reciprocal;

step 4.5: let mu let^iterμ and iter + 1;

step 4.6: if iter is less than or equal to Max _ iter and the current mean value mu^iterAnd the last mean value mu^iter-1With | | | mu between^iter-μ^iter-1||₂If epsilon, repeat steps 4.3 to 4.5 until iter > Max iter or the current mean value mu^iterAnd the last mean value mu^iter-1Satisfies | | mu between^iter-μ^iter-1||₂Less than or equal to epsilon, the current posterior mean being the sparse representation coefficient obtained, i.e.

Wherein | · | purple sweet₂Represents a 2-norm;

from step 4.1 to step 4.6, the sparse coding of the human activity test sample is completed, and a sparse representation coefficient for carrying out the sparse coding on the human activity test sample is obtained;

and 5: classification based on minimum residuals.

2. The human activity classification method based on weighted group sparse Bayesian learning as recited in claim 1, wherein: step 2, specifically: clutter suppression is performed firstly, and then the window length is N_wThe short-time Fourier transform of the radar echo signal is obtained.

3. The human activity classification method based on weighted group sparse bayes learning according to claim 2, characterized in that: in step 3, the dictionary a is recorded as: a ═ A₁，A₂，...，A_K) (ii) a The c-th sub-dictionary is noted as: a. the_cAnd c is 1,2, K is the number of sample classes in the dictionary a.

4. The human activity classification method based on weighted group sparse Bayesian learning as recited in claim 3, wherein: step 4.4, the concrete steps are as follows:

step 4.4 a: estimating an element x_iThe inverse variance of (c):

wherein the content of the first and second substances,

is an element x_iEstimate of the inverse of the variance, n_iIs x_iThe column number of the corresponding sub-dictionary;

the weighting factor lambda belongs to [0,1 ]]G (i) represents and x_iAn element index belonging to the same class of tags;

step 4.4 b: carrying out weighted average on the inverse variance estimation of the obtained same-class samples, and updating the element x_iThe inverse variance of (c):

5. the human activity classification method based on weighted-group sparse Bayesian learning as recited in claim 4, wherein: step 5, specifically: computing a residual r between the test sample y and a linear weighted sum of each type of sub-dictionary_c(y)：

Wherein the content of the first and second substances,

to represent

Keeping the coefficient corresponding to the middle-class-c sub-dictionary unchanged, and setting other coefficients to be zero; when the residual value of a certain class is minimum, y is considered to belong to the class.

Images