CN113609977B

CN113609977B - Pedestrian gait recognition method based on channel state information quotient distance

Info

Publication number: CN113609977B
Application number: CN202110892189.XA
Authority: CN
Inventors: 王林; 张德安; 刘文远; 王新雨; 厉斌斌
Original assignee: Yanshan University
Current assignee: Yanshan University
Priority date: 2021-08-04
Filing date: 2021-08-04
Publication date: 2023-10-17
Anticipated expiration: 2041-08-04
Also published as: CN113609977A

Abstract

The invention relates to the technical field of identity recognition, in particular to a pedestrian gait recognition method based on channel state information quotient distance, which comprises the following steps of data acquisition: acquiring CSI data of walking of a person in a sensing area by using one WiFi transmitting end and two WiFi receiving ends; signal pretreatment: carrying out relevant denoising processing on the acquired CSI data, and eliminating irrelevant noise information contained in the CSI data and related to walking; CSID extraction: the change information of the channel state information quotient in the complex plane contains gait characteristics of walking, and CSID is provided to represent the change of the channel state information quotient in the complex plane; identification: and combining the advantages of deep learning in the aspect of feature extraction, and performing feature extraction by using an LSTM network model so as to realize identification. The method can realize the passive sensing of the identity of the person in the indoor environment by using only commercial WiFi equipment, and realizes a user identity recognition system by utilizing the unique gait characteristics of the walking of the person and combining a deep learning model.

Description

Pedestrian gait recognition method based on channel state information quotient distance

Technical Field

The invention relates to the technical field of identity recognition, in particular to a pedestrian gait recognition method based on channel state information quotient distance.

Background

With the rise of the Internet of things, the related technology based on WiFi signal perception is layered endlessly in the world of the tide mats of the Internet of things, and the perceived advantage of no equipment in the environment of the Internet of things becomes a great hotspot in the field of general computing. The advent of these critical application technologies also provides research background and meaning for studies based on WiFi-aware application security, and for many applications, it is necessary to perceive human identity information in advance.

Identification is always the research focus in the fields of pervasive computing and man-machine interaction, and has extremely high economic value and wide application scenes in the fields of smart home, safety monitoring and the like. In terms of identification, numerous solutions are emerging, and deep learning networks in recent research are increasingly integrated into positioning systems by using various modes including cameras, wi-Fi, RFID, voice signals and the like. Among them, the Wi-Fi signal-based identification technology is receiving a great deal of attention due to wide deployment and strong availability.

Identity recognition refers to the process of predefining identity information of a person, and matching the identity information of the current person with a database of known identity information. Some conventional identity recognition methods, such as authentication based on password input, mobile phone screen gesture unlocking and the like, are not high in safety, and the password and the gesture can be stolen by others. At present, identification based on the unique biological characteristics of human beings is a mode with higher safety, and the biological characteristics of human beings can be roughly divided into two types in the biological characteristics, wherein one type is the physiological characteristics of human beings, and the physiological characteristics mainly comprise human faces, irises, fingerprints, palmprints, DNA (deoxyribonucleic acid) and the like. The other is a human behavioral feature, which mainly includes gait, sound, keystroke action, hand signature, and the like. Identity authentication based on human biological features has been widely used in the current society, and various devices (fingerprint check-in machines, DNA detectors, etc.) are designed, which tend to be expensive. In addition, the camera-based method is used for authenticating the human identity through the human face and gait, has higher accuracy, but can involve privacy problems in some environments and has dependence on illumination conditions. By combining the authentication modes, each identity recognition method has respective advantages and disadvantages due to different application scenes. In intelligent space sensing applications, people hope to get rid of the defects, and the people in the environment are hoped to be passively sensed by a sensing technology with the advantages of universality, easiness in deployment and the like.

Disclosure of Invention

In order to solve the problems, the pedestrian gait recognition method based on the channel state information quotient distance provided by the invention can realize passive sensing of the identity of a person in an indoor environment by using only commercial WiFi equipment, does not need to use a specific sensor, and does not need active interaction between the person and the equipment. The method utilizes the unique gait characteristics of human walking and combines a deep learning model to realize a user identity recognition system.

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

the pedestrian gait recognition method based on the channel state information quotient distance comprises the following steps:

step 1, data acquisition: the method comprises the steps that one WiFi transmitting end and two WiFi receiving ends are utilized to collect CSI data of walking of a person in a sensing area and are used for model training and testing;

step 2, signal preprocessing: carrying out relevant denoising processing on the acquired CSI data, and eliminating irrelevant noise information contained in the CSI data and related to walking;

step 3, CSID extraction: the change information of the channel state information quotient in the complex plane contains gait characteristics of walking, and CSID is provided to represent the change of the channel state information quotient in the complex plane;

step 4, identity recognition: and combining the advantages of deep learning in the aspect of feature extraction, and performing feature extraction by using an LSTM network model so as to realize identification.

In step 1, selecting two pieces of CSI data on a vertical WiFi link for system test, firstly carrying out data preprocessing on each piece of CSI data, carrying out preliminary denoising through two antenna operators to obtain a CSI quotient, then further extracting amplitude quotient and phase difference information of the CSI quotient, further denoising the amplitude quotient and the phase difference by utilizing hampel filtering and discrete wavelet transformation, and reconstructing the denoised amplitude quotient and phase difference to obtain a denoised CSI quotient; and extracting the CSID information containing the subcarriers from the motion of the CSI quotient in the complex plane, and packaging the CSID information and the label representing the true identity to form one sample in the data set.

Preferably, in step 2, the amplitude quotient and the phase difference information of the CSI quotient are extracted and are further denoised respectively, and then the denoised amplitude quotient and phase difference are subjected to signal inverse transformation to obtain denoised CSI quotient information.

Preferably, in step 2, it comprises

Step 21, denoising the CSI quotient: the CSI obtained by the two antennas is divided, the CSI is complex information, the complex information comprises amplitude and phase information of signals, the complex division can obtain that the amplitude of a CSI quotient is changed into the ratio of the amplitudes of the original two antennas, the phase of the CSI quotient represents the phase difference of the original two antennas, and the CSI quotient information obtained by the two antenna manufacturers comprises signal change information caused by walking;

step 22, hampel filtering denoising: removing an abnormal value obviously existing in the signal waveform by using a Hampel filter;

step 23, denoising through discrete wavelet transform: the system further denoises the amplitude quotient and phase difference information after the Hampel filtering by adopting Discrete Wavelet Transform (DWT), and focuses the signal waveform to the maximum extent on human walking behaviors; when decomposing an original signal X with frequency F by using discrete wavelet transformation, the original signal X is filtered by a low-pass filter and a high-pass filter firstly; the low-pass filter is decomposed into a 1 st-level approximation coefficient ACs, the high-pass filter is decomposed into a 1 st-level detail coefficient DCs, the approximation coefficient ACs and the detail coefficient DCs respectively correspond to a low-frequency part of an original signal 0-F/2 and a high-frequency part of F/2 to F, and then the approximation coefficient obtained by decomposing each level is subjected to iterative decomposition through the same decomposition process, so that more related wavelet coefficients with different frequency information and time scales can be obtained.

Preferably, in step 23, db4 wavelet is selected to perform 3-level wavelet decomposition on the amplitude quotient and phase difference of the selected subcarrier; the sampling rate of the CSI collection is 1000Hz, the highest frequency of signals which can be contained in the CSI data is 500Hz, the signals with the frequency of 0-62.5Hz are obtained by reconstructing the 3 rd-level approximation coefficients, and the discrete wavelet transformation can better remove signal noise which is irrelevant to walking movement in the waveform, so that the denoised signal waveform is better focused on the signal waveform change caused by walking.

Preferably, in step 4, a double-layer long-short-term memory neural network is adopted, and the gait features of each user to be identified are extracted by using an LSTM unit structure so as to realize robust identification, and the neural network model extracts gait features contained in the CSID through the LSTM unit in the neural network hidden layer for the input CSID sequence.

Preferably, the CSID obtained by signal preprocessing of the acquired CSI data is used as an input of a network model, and the CSID sequence of 30 subcarriers in the CSI information is represented as the following formula

Wherein i represents a subcarrier number, j represents time-series information; CSID represents information of the change of the euclidean distance from a point to an initial point on the complex plane of CSI quotient in time series, and one sample data in the data set can be represented as d= { (X) _i Y), i=1, 2,3,..30, wherein the CSID sequence extracted from 30 subcarriers in the CSI data and its corresponding one identity tag are included in D; x is X _i Representing 30 CSID sequences obtained by processing 30 sub-carrier information, Y represents a tag value and represents identity information; the identification question is used for creating a model which can be based on the input X _i The matrix predicts the identity tag Y;

for each hidden layer input, LSTM may compare time t and i= [1,2]Layer inputMapping to feature f _t As shown in the formula:

where g represents the activation function and,representing an input, comprising the current information of the sequence +.>Short-term information->Long-term information->Wherein the dimension is d=30, p represents the weight matrix, and b is the bias;

the output of LSTM is subjected to matrix transformation through the full connection layer, and the probability of identifying each user is obtained through the following formula:

y＝W _o *o+b _o ；

wherein O is E R ^m×n Represents the output of the LSTM model, m represents the input batch size, n represents the output dimension size of the LSTM model, W _o ∈R ^n×7 Parameter matrix representing fully connected network, b _o Representing the bias of the network, y ε R ^m×7 Representing a final probability matrix;

normalizing the probability of each user by using a Softmax function; the probability of each output category is calculated by using the following formula, the probability of each output category is expressed, the probability is selected to be the largest, namely the classification result:

wherein y is _i Representing the output value of the ith node of the LSTM network, k representing the number of output nodes, namely the number of classified categories, and outputting the multi-classified problems through a Softmax functionValue conversion to [0,1 ]]Probability distribution problems between;

the loss function used by the model is a cross entropy loss function, which is expressed as the following equation:

where m is the number of samples involved in the loss function calculation, typically the number of samples loaded in a batch by the data set loader, Θ is a parameter of the model, Y _i For the true value of the data, i.e. the tag value, y _i And (3) carrying out gradient regression on the network model according to the condition of the loss function as the predicted value of the network until the training is ended when the loss function is converged.

The beneficial effects of using the invention are as follows:

1. the method uses the denoising method of the combination of the CSI quotient and the discrete wavelet transformation to effectively remove noise information irrelevant to walking activities in signals, and uses the time-frequency analysis of the discrete wavelet transformation to mainly concentrate the frequency of the signals on the frequency caused by walking, thereby being more beneficial to the extraction of CSID information at the back and obviously improving the accuracy of identity recognition.

2. The method provides the CSID as the basis of identity recognition, and the unique walking gait characteristics of the person contained in the CSID are analyzed through experiments. The system realizes the identification of the person on the basis.

3. The method adopts the deep learning technology to extract the characteristics and realize the identity. And selecting an LSTM model to train and test on the existing public data set by combining with the time sequence characteristics of the CSID. Gait characteristic information of the walking is fully extracted, so that the identification accuracy reaches more than 90%.

Drawings

Fig. 1 is a flow chart of a pedestrian gait recognition method based on the channel state information quotient distance of the present invention.

Fig. 2 is a diagram of the original amplitude phase information on the antenna 1 and the antenna 2 in the pedestrian gait recognition method based on the channel state information quotient distance according to the present invention.

Fig. 3 is a diagram of CSI denoising results in the pedestrian gait recognition method based on the channel state information quotient distance according to the present invention.

Fig. 4 is a schematic view of subcarriers with outliers in the pedestrian gait recognition method based on the channel state information quotient distance according to the present invention.

Fig. 5 is a schematic diagram of a subcarrier after denoising by hampel filtering in the pedestrian gait recognition method based on the channel state information quotient distance according to the present invention.

Fig. 6 is a flowchart of a discrete wavelet transform process in the pedestrian gait recognition method based on the channel state information quotient distance of the present invention.

Fig. 7 is a schematic view of a discrete wavelet transform denoising subcarrier of the pedestrian gait recognition method based on the channel state information quotient distance.

Fig. 8 is a schematic diagram of CSID information on a subcarrier of a pedestrian gait recognition method based on a channel state information quotient distance according to the present invention.

Fig. 9 is a schematic diagram of a network structure in the pedestrian gait recognition method based on the distance of the channel state information quotient according to the present invention.

Fig. 10 is a confusion matrix diagram after user data is identified in the verification of the pedestrian gait recognition method based on the channel state information quotient distance according to the present invention.

FIG. 11 is a graph of score, precision and recall of user data identification in verification of the pedestrian gait recognition method based on channel state information quotient distance of the present invention.

Fig. 12 is a graph showing the change of the identification accuracy with the number of people in the verification of the pedestrian gait recognition method based on the distance of the channel state information quotient.

Fig. 13 is a graph showing the influence of discrete wavelet transform denoising on user recognition accuracy in verification of the pedestrian gait recognition method based on the channel state information quotient distance of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present technical solution more apparent, the present technical solution is further described in detail below in conjunction with the specific embodiments. It should be understood that the description is only illustrative and is not intended to limit the scope of the present technical solution.

As shown in fig. 1 to 9, the pedestrian gait recognition method based on the distance between the channel state information providers according to the present embodiment is divided into four parts, and the flow is shown in fig. 1. The whole system framework mainly comprises four components, namely data acquisition, signal preprocessing, CSID extraction and identity recognition. And the data acquisition utilizes one WiFi transmitting end and two WiFi receiving ends to acquire CSI data of walking of people in the sensing area, and the CSI data are used for model training and testing. The signal preprocessing carries out relevant denoising processing on the acquired CSI data, eliminates irrelevant noise information contained in the CSI data and related to walking, and better lays down for the next step. The change information of the channel state information quotient in the complex plane contains gait characteristics of walking, and CSID (Euclidean distance from the initial point to each point after the complex plane) is proposed to represent the change of the channel state information quotient in the complex plane. In the identification, the advantage of deep learning in the aspect of feature extraction is combined, and the LSTM network model is used for feature extraction so as to realize the identification.

In the data acquisition part, the method uses the existing public data set to test and evaluate the identity recognition system proposed in the paper, selects the CSI data on two vertical WiFi links to perform system test, and the data set contains the CSI information acquired by 7 persons when walking in 8 directions of 4 walking paths. In the method, each piece of CSI data is subjected to data preprocessing, preliminary denoising is carried out through two antenna manufacturers to obtain CSI quotient, then amplitude quotient and phase difference information of the CSI quotient are further extracted, the amplitude quotient and the phase difference are further denoised through hampel filtering and discrete wavelet transformation, and the denoised amplitude quotient and phase difference are reconstructed again to obtain denoised CSI quotient. CSID information for 30 sub-carriers is extracted from the motion of the CSI quotient in the complex plane, and finally a matrix with a size of 2300 x 30 is formed. Therein 2300 represents a sampling sequence (data of a sampling rate 1000,2.3 seconds), 30 represents the number of subcarriers, and each column of the matrix represents CSID information of one subcarrier. The 2300 x 30 matrix is then packaged with a tag representing the true identity to form one sample in the dataset. The resulting dataset contained 5392 samples in total, the dataset was processed to a data set of about 8:2, wherein 4321 training sets and 1071 testing sets are included.

In the signal preprocessing part, noise caused by hardware and impulse noise in the signal are primarily removed by means of CSI (channel state information) operators on two antennas. In order to further remove noise which is not related to walking information and is contained in the signal, amplitude quotient and phase difference information of the CSI quotient are extracted, the amplitude quotient and the phase difference information are further denoised respectively, and then signal inverse transformation is carried out on the denoised amplitude quotient and phase difference to obtain denoised CSI quotient information. In the signal preprocessing stage, denoising methods such as a two-antenna CSI (channel state information) manufacturer, hampel filtering, discrete wavelet transform and the like on the same receiver are mainly used.

(1) CSI quotient denoising. As fig. 2 shows the amplitude and phase information of CSI received by antenna 1 and antenna 2, respectively, amplitude fluctuations due to walking behavior can be observed from the amplitude information in the original CSI, but there is a lot of noise information independent of walking behavior in these signal fluctuations due to the influence of the device hardware itself and the environment. Meanwhile, the phase in the original CSI can be seen to be randomly distributed in the interval of [ -pi, pi ], and the related information of the walking behavior can not be directly obtained from the original phase information.

The CSI obtained by the two antennas is divided, the CSI is complex information, which contains the amplitude and phase information of the signal, and the complex division can know that the amplitude of the CSI quotient becomes the ratio of the amplitudes of the two original antennas, and the phase of the CSI quotient represents the difference of the phases of the two original antennas. The extraction of amplitude quotient and phase difference information from CSI quotient is shown in fig. 4. Compared with the original signal diagram, the noise in the amplitude is effectively eliminated, and the phase information which cannot be directly used originally becomes available phase difference information. As can be seen from fig. 3, the walking activity has a significant effect on the signal waveform, that is, CSI quotient information obtained through the two antenna manufacturers includes signal change information caused by walking.

(2) And removing noise by using a hampel filter. Amplitude quotient and phase difference information are extracted from the CSI quotient, and the two-antenna quotient is utilized to see that some impulse noise in the amplitude information is well removed, so that the original unavailable phase information is converted into phase difference information to be available information. However, when signal analysis is performed, as shown in fig. 4, it is found that some abnormal noise still exists in the subcarrier waveforms, and the abnormal noise is often caused by hardware defects and the existence of special noise in the environment.

Here, the system further utilizes a Hampel filter to remove outliers that are apparent in the signal waveform. The main principle of the Hampel filter is to remove abnormal data values through a moving average window, directly take the average value of the data, and remove some points with larger difference to achieve the purpose of removing abnormal noise. From fig. 5, it can be seen that some data outliers present in the amplitude quotient and phase difference are effectively removed.

(3) Discrete wavelet transform denoising. The most immediate benefit of using discrete wavelet transforms is that wavelet decomposition is able to decompose the signal at different scales and different decomposition scales can be chosen according to the actual needs. Often, low frequency information in many signals is particularly important, which implies a major characteristic of the signal, while the high frequency content of the signal gives details or differences in the signal. The signal frequency corresponding to the walking activity of the person is mainly concentrated in low-frequency information, and the signal is decomposed and reconstructed through discrete wavelet transformation, so that the characteristic information generated by the walking activity in the signal can be better reserved.

The system further denoises the amplitude quotient and phase difference information after the Hampel filtering by adopting Discrete Wavelet Transform (DWT), and focuses the signal waveform to the maximum extent on human walking behaviors. The discrete wavelet transform enables multi-resolution analysis of discrete signals in the time and frequency domains. As shown in fig. 6, when decomposing the original signal X with frequency F using discrete wavelet transform, the original X signal is first filtered by low-pass and high-pass filters. Is decomposed into level 1 approximation coefficients (Approximation Coefficients, ACs) by a low pass filter and level 1 detail coefficients (Detail Coefficients, DCs) by a high pass filter. ACs and DCs respectively correspond to the low-frequency part of the original signal 0-F/2 and the high-frequency parts of F/2 to F, and then the approximate coefficients obtained by decomposing each stage are subjected to iterative decomposition through the same decomposition process, so that more related wavelet coefficients with different frequency information and time scales can be obtained.

In the method, db4 wavelet is selected to carry out 3-level wavelet decomposition on the amplitude quotient and the phase difference of the selected subcarrier, so as to achieve the best denoising effect. As the sampling rate of the data set adopted in the method for collecting the CSI is 1000Hz, the highest frequency of the signals which can be contained in the CSI data is 500Hz according to the Nyquist sampling theorem, the signals with the frequency of 0-62.5Hz are obtained by reconstructing the 3 rd-level approximation coefficient, and the discrete wavelet transformation can better remove the signal noise which is irrelevant to the walking movement in the waveform, so that the denoised signal waveform is better focused on the signal waveform change caused by the walking. For the acquired walking CSI quotient data, fig. 7 shows the denoising effect of three-level wavelet approximation coefficient reconstruction for amplitude quotient and phase difference using db4 wavelet.

CSID extraction: the preprocessing process of the CSI data is introduced in the previous section, and the change information of the CSI quotient in the complex plane after the CSID is used for representing and denoising is presented in the present section. In the CSI quotient model, there are mainly the following three properties.

(1) The CSI quotient trace in the complex plane is a perfect circle when the reflected path length of a moving object in the environment changes by one wavelength. When the length of the reflection path changes by a plurality of wavelengths, the circle of the CSI quotient on the complex plane rotates correspondingly by the same number of rounds.

(2) When the length change of the reflection path is smaller than one wavelength, the CSI quotient forms an arc whose measure of the arc approximately matches the length change of the reflection path.

(3) The direction of rotation of the CSI quotient arc is closely related to the direction of motion of the target in the fresnel zone.

Firstly, the method carries out related research on the related property of the CSI quotient aiming at the walking activity of the person, carries out signal preprocessing on the acquired CSI data to obtain the CSI quotient, and then observes the change track of the CSI quotient on a complex plane. It was found in experiments that the CSI quotient does not perform a repeated circular motion in the complex plane over the complete circumference, which forms a circular shape, and that the radius size varies. This is because the human body is not an ideal reflection source of the signal, and even if the relevant experimental variables are controlled, the movements of the legs of the person while walking and the minute movements of other parts of the body inevitably have some influence on the signal.

The gait features of human body movement comprise the movement features of all parts of the human body when the human body walks, and mainly refer to the swing of arms, stride frequency information and coordinated movement information of other parts of the human body. Because each person has unique walking behavior, gait characteristics can be used as the basis for person identification. And generally gait behavior is difficult to imitate because it is not unitary, and it is a fusion of characteristic information of various parts of the human body.

The CSI quotient moves in a complex plane because each dynamic part changes the length of a dynamic reflection path in the environment when a human body walks. Because walking gait of each person is different, the walking modes of different persons can be analyzed to generate different influences on the change of the CSI quotient on the complex plane, and the change of the CSI quotient on the complex plane caused by walking contains walking gait information unique to the person, so that the movement information of the CSI quotient on the complex plane can be used as a basis for personal identification.

By analyzing the data collected in the real walking experiment, the motion of the CSI quotient on the complex plane is found to be capable of capturing the beginning and ending of the walking behavior of the human body well. Based on the following experimental observation, when a person is stationary, the environment only comprises a static reflection path, the movement track of the CSI quotient on the complex plane always moves in a smaller range, and when the person starts to walk, a dynamic reflection path exists in the environment, so that the movement range of the CSI quotient on the complex plane can be enlarged.

In order to quantitatively express the analysis result of the motion of the CSI quotient on the complex plane, a CSID is provided, wherein the CSID expresses the distance information from an initial point to each point after the initial point when the CSI quotient moves on the complex plane. The motion of the CSI-quotient in the complex plane is quantified by the point-to-initial point distance information that moves continuously over time.

The CSID information acquired by one subcarrier in the CSI data is shown in fig. 8. As can be seen from the figure, the CSID has the same characteristic of sensing the change of the surrounding environment as the CSI information, and is stable to the static environment and has a certain sensitivity to the dynamic environment. When a person has a static reflection path in an environment, i.e. the environment is stationary, it can be seen from the figure that the fluctuation of the CSID is very smooth, and when a person starts to walk, i.e. a dynamic reflection path in the environment, the fluctuation range of the CSID becomes significantly large.

The CSID well quantifies the motion trail of the CSI quotient in the complex plane. A perfect peak in the CSID represents the process of the CSI quotient making a circular-like motion in the complex plane, because in the case of determining an initial point, when the point starts to make a circular-like motion, it is easy to think that the locus of the point is farther from the initial point until it approaches the initial point after forming a semicircle until it coincides with the initial point. The abscissa of the peak value in the figure represents the moment when the point moves to half on a complete circle, and the change of the ordinate of the peak in the figure is caused by the fact that the human body is not a perfect reflector and each dynamic part of the human body affects the dynamic reflection path in the environment. While in the time dimension the number of peaks contains information about the speed of change of the reflected path length.

From the above analysis, the following conclusions can be drawn: the CSID provided by the method can be used for sensing human body movement in the environment, and the time domain change of the CSID is caused by the dynamic change of the length of a transmitting path caused by each part when the human body walks. The CSID contains a large number of human walking gait features, and can theoretically provide basis for identity recognition. The CSID provided by the system provides a new thought for gait recognition based on the WiFi signal, and in the next work, the powerful feature extraction capability of deep learning is utilized to extract the walking features of the person contained in the CSID, so that the identity recognition is realized. Meanwhile, the favorable sensing capability of the CSID is beneficial to effectively preprocessing and reconstructing the CSI quotient signal by the signal preprocessing of the system.

Model training: the track of the CSI quotient on the complex plane can reflect the gait characteristics of walking of a person, the track on the complex plane on the CSI is accompanied by time variation, and the CSID information extracted from the complex matrix of the CSI quotient after denoising is a sequence which varies with time. The method hopes to extract gait characteristics from a time sequence CSID sequence and identify the gait characteristics, and in the field of deep learning, RNN (RNN-like network) processing on the time sequence is particularly prominent.

The human walking behavior is a natural continuous behavior, and the signal sequence generated by walking is often more than 1s in the time dimension. The traditional RNN model has the gradient vanishing problem, has limited capability of extracting long-term information, and cannot extract complete gait characteristics under a time sequence. LSTM, as a variant of RNN network, can learn long-term dependencies and is now widely used for timing sequence processing.

In order to be able to capture all previous information in the sequence, including long-term and short-term information of walking behaviour, the method uses a double-layer long-and-short-term memory neural network when extracting gait features. And extracting gait characteristics of each user to be identified by using the LSTM unit structure instead of the nerve unit structure in the traditional RNN model so as to realize robust identification. As fig. 9 shows the basic structure of the network, the neural network model will extract gait features contained in the CSID from the input CSID sequence through LSTM cells in the hidden layer of the neural network.

The CSID obtained by preprocessing the acquired CSI data is used as the input of a network model, and the CSID sequence of 30 subcarriers in the CSI information is represented as the following formula

Where i denotes a subcarrier number, a total of thirty subcarriers are included, j denotes time sequence information, here, a sampling sequence in CSI raw data, and CSID denotes information of a change in euclidean distance from a point to an initial point on a complex plane of CSI quotient in the time sequence. One sample data in a datasetCan be expressed as d= { (X) _i Y), i=1, 2,3,..30, wherein the CSID sequence extracted from 30 subcarriers in the CSI data and its corresponding one identity tag are included in D. X is X _i The 30 CSID sequences obtained by processing the 30 sub-carrier information are represented, Y represents a tag value and represents identity information. In this method, the identification problem is in fact to create a model which can be based on the input X _i The matrix predicts the identity tag Y.

where g represents the activation function and,representing an input, comprising the current information of the sequence +.>Short-term information->Long-term information->Where the dimension is d=30, p represents the weight matrix and b is the bias.

The identification problem is actually a multi-classification problem, so the system performs matrix transformation on the output of the LSTM through the full connection layer, and the probability of identifying each user is obtained through the following formula.

y＝W _o *o+b _o

O∈R ^m×n Represents the output of the LSTM model, m represents the input batch size (batc)h_size), n represents the output dimension size, W, of the lstm model _o ∈R ^n×7 Parameter matrix representing fully connected network, b _o Representing the bias of the network, y ε R ^m×7 Representing the final probability matrix.

The probability for each user is then normalized using the Softmax function. The essence of the Softmax function is that any real vector input maps to another real vector of the same dimension, each number in the mapped vector is between (0, 1) and the sum is 1 (normalized). And calculating the probability of each output category by using the following formula, representing the probability of belonging to each category, and selecting the category with the highest probability as a classification result.

Wherein y is _i The output value of the ith node of the LSTM network is represented, and k represents the number of output nodes, namely the number of classified categories. The output value in the multi-classification problem can be converted into [0,1 ] by the Softmax function]Probability distribution problems between.

where m is the number of samples involved in the loss function calculation, typically the number of samples loaded in a batch by the data set loader, Θ is a parameter of the model, Y _i For the true value of the data, i.e. the tag value, y _i Is a predicted value for the network. And carrying out gradient regression on the network model according to the condition of the loss function until the training is ended when the loss function converges.

It is noted that the deep learning model structure used in the system contains only two layers of LSTM network structure, because the method has performed a relevant preprocessing on the raw data prior to data entry. The CSI quotient preliminarily removes some impulse noise in the CSI original data by using a two-antenna quotient mode, and the signal-to-noise ratio of the original signal is greatly improved. And then extracting amplitude quotient and phase difference in the CSI quotient to perform abnormal noise removal and discrete wavelet denoising on the amplitude quotient and the phase difference respectively, wherein the abnormal value can also cause obvious abnormal value in the CSID due to the distance of the CSI quotient expressed by the CSID at a complex plane point, so that the signal frequency influenced by walking of people is better maintained based on the time-frequency analysis characteristic of discrete wavelet transformation, and the noise of the walking motion of irrelevant people in the signal is further removed. And carrying out signal reconstruction on the amplitude quotient and the phase difference containing the walking information to obtain a denoised CSI quotient, then extracting the CSID from the CSI quotient, and finally, retaining the signal characteristics generated by walking to the greatest extent in the obtained CSID. The two-layer network structure adopted by the method can be used for extracting gait features contained in the CSID, and a more complex network model also often represents more time cost. The distance information represented by the CSID varies with time, and the conventional statistical feature extraction method cannot fully utilize features in a long time sequence and also damages some original feature information in the signal.

Based on the analysis, the method adopts the deep learning model only comprising two layers of LSTM to extract the characteristics and identify the identity, achieves the average 97% of identification accuracy when identifying 4 persons, and achieves the average 90% of identification accuracy when identifying 7 persons.

The invention provides a pedestrian gait recognition method based on channel state information quotient distance, which mainly comprises the following steps:

(1) Denoising method combining CSI quotient with discrete wavelet transformation

The CSI data directly acquired from the WiFi equipment contains a large amount of noise and environmental noise caused by the equipment, so that the CSI cannot be directly applied to identity sensing work, and the denoising method of the CSI quotient combined with discrete wavelet transformation is provided by combining the denoising method of the CSI quotient with the traditional method of the CSI quotient of the same receiving different antennas. And extracting amplitude quotient and phase difference information of the CSI quotient, and determining wavelet decomposition and reconstruction layers by utilizing time-frequency analysis characteristics of discrete wavelet transformation and sampling frequency of data acquisition, so that waveforms of signals are better focused on human walking behaviors.

(2) Extraction of CSID information

In the prior art, in the work of identifying identity by using WiFi signals, sensing work is mostly directly performed based on the amplitude or phase information of the CSI, in the method, euclidean distance (CSID) from the initial point of a complex plane to each point after the initial point of the complex plane of the CSI is provided as the basis of identity sensing, and the reason that the CSID contains walking gait characteristics is explained in detail.

(3) Feature extraction and identity identification for CSID based on LSTM neural network

The method uses LSTM neural network to extract gait characteristics contained in the CSID so as to realize identity recognition. As CSID is distance information changing along with time, is time sequence information, and is well known, RNN model has natural advantage in processing time sequence signals, but problems such as gradient disappearance and gradient explosion are easy to generate when long-term information is extracted, continuous signals generated by walking behavior always last for more than 1s, long-term gait features of people walking are extracted when gait features are extracted, LSTM is used as a variant of RNN network, and the problems of gradient elimination and gradient explosion can be effectively avoided.

Example 1

The feasibility and effect of the method are described in detail below by way of examples.

1. Identification accuracy

Seven data are selected from the data set to be tested and analyzed, figure 10 shows a confusion matrix after identifying 7 users, and figure 11 shows the corresponding accuracy, recall and F1 score.

According to the method, the identity of 7 users can be identified through the confusion matrix, each row in the confusion matrix represents the true identity label of the data, each column represents the predicted identity label of the data, and the average accuracy rate of the identity identification of seven users reaches 90.66%. As can be seen from fig. 1, the recognition confusion degree of the users 5, 6 and 7 is high, and when the user 7 is recognized, 13% of the samples of the user 7 are recognized as the user 6, which indicates that the user 7 has a similar body shape to the user 6 and has small difference in walking posture, so that the false recognition phenomenon exists in the recognition. The recognition accuracy of the user 3 reaches 100%, because the walking gait characteristics of the user 3 are obviously different from those of the rest 6, and the feasibility of the gait recognition scheme principle adopted by the method is further verified.

2. Influence of the number of training persons on the experimental results

In this section, it is mainly analyzed how the accuracy of the method used by the system changes along with the change of the number of people, and the method uses the data of 7 users in the public data set to test, and the potential intruder is not considered in the test, i.e. no user not belonging to the 7 users exists.

In the experimental process, relevant test analysis is carried out on the recognition accuracy when the target number is 2 and the target number is 7, as shown in fig. 12, the experiment shows that the accuracy of the identification can be slightly reduced along with the increase of the recognition number, and similar gait characteristics can exist during feature clustering in the gait feature space along with the increase of the classification number.

As can be seen from fig. 12, the recognition accuracy of the user identification model proposed by the system is maintained above 97% when the number of recognized people is not more than 4, and the sides of the model reflect that the extracted gait characteristics are different in recognition of different people. When the number of the identification target persons is 5, the identification accuracy rate is reduced to 94%, and when the number of the identification persons is increased to 7, the identification accuracy rate can still reach 90%. The recognition accuracy rate reaches 97% when the number of the recognized people is four, which shows the application potential of the method in intelligent home scenes with only a few users.

3. Influence of denoising method on experimental result

CSID is proposed for gait recognition of people in the environment based on the relevant properties of CSI quotient in complex plane motion. The CSID is closely related to the CSI quotient, and noise information which is contained in the CSI quotient and is irrelevant to walking activities is further eliminated by combining time-frequency analysis characteristics of discrete wavelet transformation in a signal preprocessing stage. In order to verify the effectiveness of the denoising method provided by the system, CSID is extracted from the CSI quotient which is not subjected to discrete wavelet denoising and is used as an original data set for testing, and other experimental parameters are kept unchanged.

As shown in fig. 13, it can be seen from the figure that, when the CSID is directly extracted from the original CSI quotient to identify the CSID, compared with the CSID extracted from the denoised CSI quotient, the identification accuracy of the user 4 is not affected, and the identification accuracy of the other six persons is affected to different degrees, which indicates that the method proposes denoising the amplitude quotient and the phase difference in the CSI quotient by using discrete wavelet transform, and then the method of reconstructing the CSI quotient effectively maintains walking gait characteristics while removing noise, and improves the accuracy of the identification.

The foregoing is merely exemplary of the present invention, and those skilled in the art can make many variations in the specific embodiments and application scope according to the spirit of the present invention, as long as the variations do not depart from the spirit of the invention.

Claims

1. The pedestrian gait recognition method based on the channel state information quotient distance is characterized by comprising the following steps of: the method comprises the following steps:

step 4, identity recognition: combining the advantages of deep learning in the aspect of feature extraction, performing feature extraction by using an LSTM network model so as to realize identity recognition;

in step 2, the amplitude quotient and the phase difference information of the CSI quotient are extracted and are respectively subjected to further denoising, and then the denoised amplitude quotient and phase difference are subjected to signal inverse transformation to obtain denoised CSI quotient information;

in step 2, it includes

Step 21, denoising the CSI quotient: the CSI obtained by the two antennas is divided, the CSI is complex information, the complex information comprises amplitude and phase information of signals, the complex division operation shows that the amplitude of a CSI quotient is changed into the ratio of the amplitudes of the original two antennas, the phase of the CSI quotient represents the difference of the phases of the original two antennas, and the CSI quotient information obtained by the two antenna manufacturers comprises signal change information caused by walking;

step 23, denoising through discrete wavelet transform: the system further denoises the amplitude quotient and phase difference information after the Hampel filtering by adopting Discrete Wavelet Transform (DWT), and focuses the signal waveform to the maximum extent on human walking behaviors; when decomposing an original signal X with frequency F by using discrete wavelet transformation, the original signal X is filtered by a low-pass filter and a high-pass filter firstly; the low-pass filter is decomposed into a 1 st-level approximation coefficient ACs, the high-pass filter is decomposed into a 1 st-level detail coefficient DCs, the approximation coefficient ACs and the detail coefficient DCs respectively correspond to a low-frequency part of an original signal 0-F/2 and a high-frequency part of F/2 to F, and then the approximation coefficient obtained by decomposing each level is subjected to iterative decomposition through the same decomposition process, so that more relevant wavelet coefficients with different frequency information and time scales are obtained.

2. The pedestrian gait recognition method based on the channel state information quotient distance according to claim 1, wherein: in step 1, selecting CSI data on two vertical WiFi links for system test, firstly carrying out data preprocessing on each piece of CSI data, carrying out preliminary denoising through two antenna manufacturers to obtain a CSI quotient, then further extracting amplitude quotient and phase difference information of the CSI quotient, further denoising the amplitude quotient and the phase difference by utilizing hampel filtering and discrete wavelet transformation, and reconstructing the denoised amplitude quotient and phase difference to obtain a denoised CSI quotient; and extracting the CSID information containing the subcarriers from the motion of the CSI quotient in the complex plane, and packaging the CSID information and the label representing the true identity to form one sample in the data set.

3. The pedestrian gait recognition method based on the channel state information quotient distance according to claim 1, wherein: in step 23, db4 wavelet is selected to perform 3-level wavelet decomposition on the amplitude quotient and the phase difference of the selected subcarrier; the sampling rate of the CSI collection is 1000Hz, the highest frequency of signals which can be contained in the CSI data is 500Hz, the signals with the frequency of 0-62.5Hz are obtained by reconstructing the 3 rd-level approximate coefficient, the signal noise which is irrelevant to the walking motion in the waveform is further removed by discrete wavelet transformation, and the denoised signal waveform is better focused on the signal waveform change caused by the walking.

4. The pedestrian gait recognition method based on the channel state information quotient distance according to claim 1, wherein: in step 4, a double-layer long-short-term memory neural network is adopted, an LSTM unit structure is used for extracting gait features of each user to be identified so as to realize robust identification, and the neural network model extracts gait features contained in the CSID through the LSTM unit in the neural network hidden layer on the input CSID sequence.

5. The pedestrian gait recognition method based on the channel state information quotient distance as claimed in claim 4, wherein: the CSID obtained by preprocessing the acquired CSI data is used as the input of a network model, and the CSID sequences of 30 subcarriers in the CSI information are expressed as the following formula

Wherein i represents a subcarrier number, j represents time-series information; CSID represents information of the change of the euclidean distance from a point to an initial point on the complex plane of CSI quotient in a time sequence, and one sample data in the data set is represented as d= { (X) _i Y), i=1, 2,3,..30, wherein the CSID sequence extracted from 30 subcarriers in the CSI data and its corresponding one identity tag are included in D; x is X _i Representing 30 CSID sequences obtained by processing 30 sub-carrier information, Y represents a tag value and represents identity information; the identification question is used for creating a model which can be based on the input X _i The matrix predicts the identity tag Y;

for the input of each hidden layer, LSTM will time t and i= [1,2]Layer inputMapping to feature f _t As shown in the formula:

y＝W _o *o+b _o ；

wherein y is _i Representing the output value of the ith node of the LSTM network, k representing the number of output nodes, i.e. the number of classified categories, and converting the output value in the multi-classified problem into [0,1 ] by the Softmax function]Probability distribution problems between;