A gait recognition method and system based on FMCW millimeter wave radar and a storage medium

Abstract

The application discloses a gait recognition method and system based on an FMCW millimeter wave radar and a storage medium, and the method comprises the following steps: constructing a millimeter wave radar echo data matrix; extracting a micro-Doppler time-frequency spectrum image feature of the data matrix; preprocessing the micro-Doppler time-frequency spectrum image; performing image segmentation on the micro-Doppler time-frequency spectrum; training data using a clustering algorithm; and performing gait recognition according to a clustering result. The application solves the problems of low spectrum image accuracy caused by millimeter wave radar signal noise and deep learning overfitting in related technologies.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of intelligent security and defense, and particularly relates to a gait recognition method and system based on FMCW millimeter wave radar and a storage medium. BACKGROUND

[0002] Gait recognition is a new biometric recognition technology, which aims to find and extract the changing features between individuals from the same walking behavior to achieve automatic identity recognition of personnel. At present, the gait recognition technology based on optical images has a wide application in the fields of intelligent security and defense, video monitoring, intelligent police work, etc. However, this technology has many limitations, such as the change of the visual angle of the monitoring camera leading to the obvious change of the representation of the same object captured, and being easily affected by light, weather, etc. in real scenes.

[0003] Compared with the detection methods of camera, laser radar, infrared sensor, etc., the millimeter wave radar has the advantages of wide detection range, not being easily affected by light, weather, heat source, etc. and greatly improving the privacy protection level of the observed object. At present, the millimeter wave radar has been widely applied to human motion recognition. In the micro-Doppler spectrum of the millimeter wave radar echo, there are rich human motion feature information, such as physical characteristics of limb motion frequency, trunk micro-Doppler frequency, etc., which can help to improve the human motion feature recognition accuracy. However, in the application of millimeter wave radar human detection, when only the features are extracted from the micro-Doppler time-frequency spectrum for motion recognition, the recognition result accuracy cannot reach the expected effect, and when deep learning is applied to human motion recognition, although the recognition accuracy is improved, deep learning requires a large amount of data and a long training time, and is prone to overfitting when facing insufficient training samples.

[0004] In order to solve the problems of low spectrum graph accuracy caused by millimeter wave radar signal noise and deep learning overfitting, a gait recognition method and system based on FMCW millimeter wave radar and a storage medium are proposed. SUMMARY

[0005] The embodiment of the present application proposes a gait recognition method and system based on FMCW millimeter wave radar and a storage medium, which at least solves the problems of low spectrum graph accuracy caused by millimeter wave radar signal noise and deep learning overfitting in the related art.

[0006] According to one embodiment of the present application, a gait recognition method based on FMCW millimeter wave radar is provided, comprising:

[0007] constructing a millimeter wave radar echo data matrix;

[0008] extracting micro-Doppler time-frequency spectrum image features of the data matrix;

[0009] Preprocess the micro-Doppler time-frequency spectrum image;

[0010] Image segmentation of the micro-Doppler time-frequency spectrum;

[0011] Training data using a clustering algorithm;

[0012] Gait recognition according to the clustering results.

[0013] In an example embodiment, the construction of the millimeter wave radar echo data matrix comprises the steps of:

[0014] Using a millimeter wave radar to collect radar echo raw data of different gaits of different test subjects;

[0015] Sampling the radar echo raw data and representing the FMCW radar echo data as a two-dimensional matrix S(n, m), where n = 0, 1, …, N-1, m = 0, 1, …, M-1, N is the number of sampling points of the radar in one frequency modulation period, and M is the total number of frequency modulation periods.

[0016] In an example embodiment, the extraction of the micro-Doppler time-frequency spectrum image features of the data matrix comprises the steps of:

[0017] Discrete Fourier transform is performed on each column of the matrix S(n, m) to obtain a matrix R(k, m), and the matrix R(k, m) is preprocessed by moving target display filtering and taking the modulus to obtain a range-time image;

[0018] Superimposing the range cell signals in the range-time image to obtain a matrix D(m)

[0019] Short-time Fourier transform is performed on the matrix D(m) to obtain a matrix T(p, q), thereby obtaining a micro-Doppler time-frequency spectrum image;

[0020] Discrete Fourier transform is performed on each row of the matrix R(k, m) to obtain a matrix V(k, l), and taking the modulus of the matrix V(k, l) can obtain a range-Doppler image.

[0021] In an example embodiment, the preprocessing of the micro-Doppler time-frequency spectrum image comprises the steps of:

[0022] Obtaining range-Doppler images under two conditions of an empty room and a non-empty room;

[0023] After normalizing the range-Doppler images of the empty room and the non-empty room, a normalized histogram is obtained;

[0024] Comparing the two normalized histograms to obtain noise with a positively skewed distribution and low relative power and calculating a noise threshold;

[0025] The lower threshold of the micro-Doppler time-frequency spectrogram is set as a noise threshold.

[0026] In an example embodiment, the image segmentation of the micro-Doppler time-frequency spectrogram comprises the steps of:

[0027] selecting a window centered on the frequency of the emitted ultrasound;

[0028] scaling the obtained time-frequency spectrogram to a logarithmic scale and normalizing it;

[0029] segmenting the normalized time-frequency spectrogram into different events;

[0030] calculating the frequency component of each event as the frequency amplitude of the time-frequency spectrogram superimposed over time during the event.

[0031] In an example embodiment, the segmentation of the normalized time-frequency spectrogram into different events comprises any one or more combinations of segmenting the time-frequency spectrogram into different events in order according to a preset event length, calculating the similarity of the time-frequency spectrogram, and segmenting the time-frequency spectrogram into the same event when the similarity is higher than a set threshold.

[0032] In an example embodiment, the training of the data using a clustering algorithm comprises the steps of:

[0033] dividing all the data equally into a training set and a test set;

[0034] representing the events as X=(x1, x2…x e ), where e represents the number of equal divisions of the time-frequency spectrogram;

[0035] taking all the events of one class in the training set as the training data set;

[0036] for each event X, calculating the Euclidean distance of the event to each cluster center, and dividing the event X into the cluster with the shortest Euclidean distance;

[0037] recalculating the center of each cluster according to the number of events in the cluster;

[0038] taking the adjusted center point as the cluster center, and repeating the above steps until the clusters no longer change.

[0039] In an example embodiment, the gait recognition according to the clustering result comprises the steps of:

[0040] setting the threshold of the nearest neighbor algorithm as the optimal cluster number n;

[0041] for all the test events of each data in the test set, calculating the Euclidean distance of the test events to the n cluster center points of the m classes, respectively, so that for each event, m x n values are obtained;

[0042] For the same test data, the average distance of each event of the data to each category prototype is calculated respectively, and then m average values are obtained;

[0043] For each category, the sum of the distances of all events of the test data to the category prototype is calculated respectively, and the category represented by the minimum value is the classification result of the test data;

[0044] The above steps are repeated to obtain the classification result of each test data, which is the result of gait recognition.

[0045] According to another embodiment of the present application, a computer readable storage medium storing a computer program for electronic data exchange is provided, wherein the computer program causes a computer to execute the above method.

[0046] According to another embodiment of the present application, a gait recognition system based on FMCW millimeter wave radar is provided, comprising:

[0047] a data acquisition unit;

[0048] a processor;

[0049] a memory;

[0050] and

[0051] one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the processor, and the programs cause the computer to execute the above method.

[0052] The gait recognition method, system and storage medium based on FMCW millimeter wave radar of the present application have the following advantages:

[0053] (1) The micro-Doppler time-frequency spectrum image features of the data matrix are extracted, and the micro-Doppler time-frequency spectrum image is preprocessed to reduce noise, which can effectively remove the influence of all static object reflection frequencies and improve the accuracy of gait recognition data.

[0054] (2) The micro-Doppler time-frequency spectrum is segmented to obtain different events and frequency components of each event, and the three-dimensional image is converted into two-dimensional, which is convenient for subsequent application of clustering algorithm to obtain more effective classification results.

[0055] (3) A two-dimensional matrix is constructed by FMCW radar signal, k-means clustering algorithm is used to train data, and nearest neighbor algorithm is used for gait recognition, which not only has wide detection range and is not easily affected by light, weather, heat source and other use conditions, but also can effectively realize unsupervised machine learning and avoid the problem of deep learning overfitting. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1is a flow chart of a gait recognition method based on FMCW millimeter wave radar of an embodiment of the present application.

[0057] Figure 2 is a flow chart of sub-step S01 of an embodiment of the present application.

[0058] Figure 3 is a flow chart of sub-step S02 of an embodiment of the present application.

[0059] Figure 4 is a flow chart of sub-step S03 of an embodiment of the present application.

[0060] Figure 5 is a flow chart of sub-step S04 of an embodiment of the present application.

[0061] Figure 6 is a flow chart of sub-step S05 of an embodiment of the present application.

[0062] Figure 7 is a flow chart of sub-step S06 of an embodiment of the present application.

[0063] Figure 8 is a structure schematic diagram of a gait recognition system based on FMCW millimeter wave radar of an embodiment of the present application. DETAILED DESCRIPTION

[0064] The present application will be described in detail below with specific embodiments. The following embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present application. These all belong to the protection scope of the present application.

[0065] A gait recognition method based on FMCW millimeter wave radar of an embodiment of the present application, a flow chart as shown in Figure 1 includes the following steps:

[0066] Step S01, constructing a millimeter wave radar echo data matrix;

[0067] Step S02, extracting the micro-Doppler time-frequency spectrum image features of the data matrix;

[0068] Step S03, preprocessing the micro-Doppler time-frequency spectrum image;

[0069] Step S04, image segmentation of the micro-Doppler time-frequency spectrum;

[0070] Step S05, training data using clustering algorithm;

[0071] Step S06, gait recognition according to the clustering result.

[0072] In one example embodiment, the step S01, the flow chart is as shown in Figure 2 The step S01 includes the following steps:

[0073] The step S011 collects radar echo raw data of different gait of different testers using the millimeter wave radar.

[0074] The step S012 samples the radar echo raw data and represents the FMCW radar echo data as a two-dimensional matrix S(n,m), where n=0,1,…,N-1, m=0,1,…,M-1, N is the number of sampling points of the radar in one frequency modulation period, and M is the total number of frequency modulation periods.

[0075] In this embodiment, the collecting module is placed at a height of 1.5 meters, and the tester walks on the treadmill at a distance of 2.5 meters from the collecting module at a speed of 6 km / h, and the tester faces the collecting module. For each tester, 10 data are collected, each lasting 10 s, and there are m testers in total. It should be noted that the millimeter wave radar is used to collect data of multiple testers, and the running speed of each tester is the same each time through the treadmill, and the relative distance to the radar is approximately the same, so as to ensure that there is no change caused by the motion direction. The collected radar echo raw data is sampled, and it is assumed that the 10 s data contains M frequency sweeping periods, and there are N sampling points in each frequency sweeping period T. The radar echo sampling data is represented as a two-dimensional matrix S(n,m), where n=0,1,…,N-1, m=0,1,…,M-1.

[0076] In one example embodiment, the step S02, the flow chart is as shown in Figure 3 The step S02 includes the following steps:

[0077] The step S021 performs discrete Fourier transform on each column of the matrix S(n,m) to obtain a matrix R(k,m), performs moving target indication filtering preprocessing on the matrix R(k,m), and takes the module to obtain a range-time image.

[0078] The step S022 superimposes the range unit signals in the range-time image to obtain a matrix D(m).

[0079] The step S023 performs short-time Fourier transform on the matrix D(m) to obtain a matrix T(p,q), so as to obtain a micro-Doppler time-frequency spectrum image.

[0080] The step S024 performs discrete Fourier transform on each row of the matrix R(k,m) to obtain a matrix V(k,l), and taking the module of the matrix V(k,l) can obtain a range-Doppler image.

[0081] In this embodiment, the FMCW radar echo data is expressed as a two-dimensional matrix S(n, m), where n = 0, 1, …, N-1, m = 0, 1, …, M-1, N and M are the number of sampling points and the total number of frequency modulation cycles of the radar in one frequency modulation cycle, respectively. The discrete Fourier transform is performed on each column of the matrix S(n, m) to obtain

[0082]

[0083] wherein the variable k = 0, 1, …, N-1 contains the distance information of the detected target from the radar antenna. The moving target indication filter is preprocessed on the matrix R(k, m), and the modulus is taken to obtain a range-time image. The vertical axis of the image is the range axis, and each column of the image corresponds to a one-dimensional target range image; the horizontal axis of the image is the time axis.

[0084] The range resolution ΔR of the FMCW radar is

[0085]

[0086] wherein B is the signal bandwidth of the FMCW radar, and C0 is the propagation speed of electromagnetic waves in air. The signals of the range units of interest in the range-time image obtained according to formula (1) are added to obtain

[0087]

[0088] wherein k1 and k2 are the minimum and maximum limits of the set range gate. If the range of the human body target from the radar is 0-4.5 m, and the range resolution is ΔR, then k1 and k2 are respectively 0 and 4.5 / ΔR.

[0089] The short-time Fourier transform is performed on D(m) to obtain

[0090]

[0091] wherein g(m) is a Hamming window function, and p is an index value of the window function movement, and the data overlap rate corresponding to each movement of the window function is 95%. The micro-Doppler time-frequency image can be obtained according to T(p, q).

[0092] The discrete Fourier transform is performed on each row of the matrix R(k, m) in formula (1) to obtain

[0093]

[0094] wherein the variable l = 0, 1, …, M-1 is the Doppler frequency information of the target. The modulus is taken on the matrix V(k, l) to obtain a range-Doppler image. The vertical axis of the image is the range axis, and the horizontal axis is the Doppler frequency axis.

[0095] In one example embodiment, the step S03, as shown in the flow chart of Figure 4 includes the steps of:

[0096] Step S031, obtaining the range-Doppler image in the case of empty room and non-empty room;

[0097] Step S032, obtaining the normalized histogram after normalizing the range-Doppler image of empty room and non-empty room;

[0098] Step S033, comparing the two normalized histograms to obtain noise with skewed normal distribution and low relative power and calculating the noise threshold;

[0099] Step S034, setting the lower threshold of the micro-Doppler time-frequency spectrum image as the noise threshold.

[0100] In this embodiment, for the same room, the normalized histogram of the range-Doppler image is obtained in the case of empty room and non-empty room (person activity). By comparing the group of images, it can be found that the image of the empty room case can show noise with skewed normal distribution and low relative power below a certain threshold, which is the noise threshold. The noise threshold is set as the lower threshold to filter the noise with skewed normal distribution, thereby removing the influence of the reflection frequency of all static objects.

[0101] In one example embodiment, the step S04, as shown in the flow chart of Figure 5 includes the steps of:

[0102] Step S041, selecting a window centered on the frequency of the transmitted ultrasound;

[0103] Step S042, scaling the obtained time-frequency spectrum to logarithmic scale and normalizing it;

[0104] Step S043, dividing the normalized time-frequency spectrum into different events;

[0105] Step S044, calculating the frequency component of each event as the frequency amplitude of the time-frequency spectrum superimposed over time during the event.

[0106] In this embodiment, a window is selected centered on the frequency of the transmitted ultrasound, + / - 600 Hz. The obtained spectrum is scaled to logarithmic scale and normalized to keep the amplitude value above half the maximum amplitude of the spectrum. The spectrum is divided into events of interest, and the event length is set to 135 milliseconds. In the present application, each file is 10 seconds long, so there are approximately 74 events per file. The current spectrum is divided into 100 equal parts on the frequency scale, and the amplitude of each frequency is added over time during an event, thereby giving each event a frequency component.

[0107] In one exemplary embodiment, the dividing the normalized time-frequency spectrogram into different events comprises any one or more combinations of sequentially dividing the time-frequency spectrogram into different events according to preset event length (see the above embodiments), calculating the similarity of the time-frequency spectrogram and dividing the time-frequency spectrogram with similarity higher than a set threshold into the same event.

[0108] In the present embodiment, the calculating the similarity of the time-frequency spectrogram and dividing the time-frequency spectrogram with similarity higher than a set threshold into the same event is calculating the similarity according to the frequency similarity and / or the amplitude similarity and / or the phase similarity of the time-frequency spectrogram and dividing the time-frequency spectrogram with similarity higher than a set threshold into the same event.

[0109] The frequency similarity is calculated according to any one of the frequency deviation average or the frequency deviation variance of the time-frequency spectrogram and the positive correlation of the frequency similarity, denoted by variable p.

[0110] The amplitude similarity is calculated according to any one of the amplitude deviation average or the amplitude deviation variance of the time-frequency spectrogram and the positive correlation of the amplitude similarity, denoted by variable q.

[0111] The phase similarity is calculated according to any one of the phase deviation average or the phase deviation variance of the time-frequency spectrogram and the positive correlation of the phase similarity, denoted by variable w.

[0112] The calculating the similarity according to the frequency similarity and / or the amplitude similarity and / or the phase similarity of the time-frequency spectrogram is any one of calculating the similarity according to the positive correlation of the frequency similarity and the similarity, calculating the similarity according to the positive correlation of the amplitude similarity and the similarity, calculating the similarity according to the positive correlation of the phase similarity and the similarity, calculating the similarity according to the positive correlation of the frequency similarity and the amplitude similarity and the similarity, calculating the similarity according to the positive correlation of the frequency similarity and the phase similarity and the similarity, calculating the similarity according to the positive correlation of the amplitude similarity and the phase similarity and the similarity, calculating the similarity according to the positive correlation of the frequency similarity and the amplitude similarity and the phase similarity and the similarity, denoted by variable z.

[0113] A1-A7 in Table A represent different embodiments of calculating the similarity, wherein the frequency similarity p of the time-frequency spectrogram, the amplitude similarity q of the time-frequency spectrogram, and the phase similarity w of the time-frequency spectrogram involved in Table A are obtained by the above-mentioned formula.

[0114] Table A Different embodiments of calculating the similarity

[0115]

[0116]

[0117]

[0118]

[0119]

[0120]

[0121] In this embodiment, the threshold Z = 0.7 is set in advance, and if the similarity z (for example, A7) of the two time-frequency spectrograms calculated according to any one of Table A is greater than Z, it is determined that the two time-frequency spectrograms are segmented into the same event.

[0122] In one example embodiment, the step S05, the flow chart is as shown in Figure 6 , comprising:

[0123] Step S051, all data is evenly divided into a training set and a test set;

[0124] Step S052, an event is represented as X = (x1, x2…x e ), where e represents the number of equal divisions of the time-frequency spectrogram, and a i represents the amplitude at the i-th equal division frequency in this event;

[0125] Step S053, all events of one category in the training set are taken as a training data set;

[0126] Step S054, for each event X, the Euclidean distance of the event to each cluster center is calculated respectively, and the event X is divided into the cluster with the shortest Euclidean distance;

[0127] Step S055, the center of each cluster is recalculated according to the number of events in the cluster;

[0128] Step S056, the adjusted center point is taken as the cluster center, and the above steps are repeated until the cluster no longer changes.

[0129] In this embodiment, the number of equal divisions of the time-frequency spectrogram e = 100, and the event is represented as data X = (x1, x2…x 100 ), where a i represents the amplitude at the i-th equal division frequency in this event. All events of one category (files of the same test person) in the training set are taken as a training data set. Let k = 15, that is, the events are divided into 15 clusters.

[0130] First, initialization is performed, and 15 events are randomly extracted as seeds of the 15 clusters. Denoted as Y = (y1, y2…y 100).

[0131] Then for each event X, calculate the Euclidean distance between the event and each cluster center Y

[0132]

[0133] Select the shortest distance among the distances between event X and each cluster, and assign X to the cluster.

[0134] Recalculate the center of each cluster: Let cluster A has n events, event X j = (x 1j , x 2j … x 100j ), then the new center point

[0135]

[0136] Repeat the clustering process using the adjusted center points until the clusters do not change. The resulting 15 center points are the representative points of the class. Concatenate the 15 events to get a prototype of the class.

[0137] Repeat the above steps for other classes to get the prototype of each class.

[0138] In an exemplary embodiment, repeat the clustering process for different number of clusters (from 5 to 70, step size 5), and the best value is obtained when the prototypes obtained by k-means clustering method are closest to the original time-frequency map, and the prototypes of each class are most different.

[0139] In an exemplary embodiment, the step S06, the flow chart is shown as Figure 7 , including:

[0140] Step S061, set the threshold value of the nearest neighbor algorithm to the number of clusters after clustering;

[0141] Step S062, for each data in the test set, calculate the Euclidean distance between each test event and the n cluster center points of the m classes, then for each event, obtain m x n values;

[0142] Step S063, for the same test data, calculate the average distance between each event of the data and the prototype of each class, then obtain m average values;

[0143] Step S064, for each class, calculate the sum of the distances between all events of the test data and the prototype of the class, then the class represented by the minimum value is the classification result of the test data;

[0144] Step S065, repeat the above steps to get the classification result of each test data, which is the result of gait recognition.

[0145] In the embodiment, the threshold value of the nearest neighbor algorithm is the optimal cluster number n obtained in step 5, and the Euclidean distance between each test event of each data in the test set and the n cluster centers of the m categories is calculated respectively, then m x n values are obtained for each incoming event; for the same test data, the average distance of each event of the data to each category prototype is calculated respectively, then m average values are obtained; for each category, the sum of the distances of all events of the test data to the prototype of the category is calculated respectively, and the category represented by the minimum value is the classification result of the test data; repeat the above steps to obtain the classification result of each test data, which is the result of gait recognition.

[0146] The computer readable storage medium of the embodiment of the application stores a computer program for electronic data exchange, wherein the computer program causes the computer to execute the above method.

[0147] The gait recognition system based on the FMCW millimeter wave radar according to the embodiment of the application has the advantages that the gait recognition system based on the FMCW millimeter wave radar can recognize the gait of a person in a non-contact manner, and the gait recognition system based on the FMCW millimeter wave radar has high recognition accuracy and high recognition speed. Figure 8 The gait recognition system based on the FMCW millimeter wave radar according to the embodiment of the application has the advantages that the gait recognition system based on the FMCW millimeter wave radar can recognize the gait of a person in a non-contact manner, and the gait recognition system based on the FMCW millimeter wave radar has high recognition accuracy and high recognition speed.

[0148] The data acquisition unit;

[0149] The processor;

[0150] The memory;

[0151] and

[0152] One or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the processor, and the programs cause the computer to execute the above method.

[0153] Of course, those skilled in the art should recognize that the above embodiments are only used to illustrate the application, and are not intended to limit the application, and any changes and modifications to the above embodiments within the scope of the application will fall within the protection scope of the application.

Claims

1. A gait recognition method based on FMCW millimeter wave radar, characterized by, The method comprises the following steps: constructing a millimeter wave radar echo data matrix; extracting micro-Doppler time-frequency spectrum image features of the data matrix; preprocessing the micro-Doppler time-frequency spectrum image; the preprocessing of the micro-Doppler time-frequency spectrum image comprises the steps of: obtaining distance-Doppler images in both empty room and non-empty room conditions; obtaining normalized histograms after normalizing the distance-Doppler images of the empty room and the non-empty room; comparing the two normalized histograms to obtain noise with a positively skewed distribution and low relative power and calculating a noise threshold; and setting a lower threshold of the micro-Doppler time-frequency spectrum image as the noise threshold; performing image segmentation on the micro-Doppler time-frequency spectrum; the image segmentation on the micro-Doppler time-frequency spectrum comprises the steps of: selecting a window centered on the frequency of the transmitted millimeter wave; scaling the obtained time-frequency spectrum image to a logarithmic scale and performing normalization; and dividing the normalized time-frequency spectrum image into different events; and calculating a frequency component of each event as a frequency amplitude of the time-frequency spectrum image superimposed over time during the event; the dividing of the normalized time-frequency spectrum image into different events comprises any one or more combinations of sequentially dividing the time-frequency spectrum image into different events according to a preset event length, calculating the similarity of the time-frequency spectrum image, and dividing the time-frequency spectrum image with a similarity higher than a set threshold into the same event; training data using a clustering algorithm; performing gait recognition according to the clustering result. 2.The FMCW millimeter wave radar-based gait recognition method of claim 1, wherein, The method for constructing the millimeter wave radar echo data matrix comprises the following steps: collecting radar echo raw data of different testers in different gaits using a millimeter wave radar; Sampling the raw data of radar echo, the FMCW radar echo data is expressed as a two-dimensional matrix wherein n = 0, 1, …, N-1, m = 0, 1, …, M-1, N is the number of sampling points of the radar in one frequency modulation period, and M is the total number of frequency modulation periods. 3.The FMCW millimeter wave radar-based gait recognition method of claim 2, wherein, The method for extracting micro-Doppler time-frequency spectrum image features of the data matrix comprises the following steps: performing a discrete Fourier transform on each column of the matrix results in a matrix performing a moving target display filter pre-processing on the matrix results in a range-time image; Stacking the range cell signals in the range-time image results in a matrix ; On the matrix The short-time Fourier transform is performed to obtain a matrix , and a micro-Doppler time-frequency spectrogram is obtained. Discrete Fourier transform is performed on each row of the matrix to obtain a matrix Discrete Fourier transform is performed on each row of the matrix Modulo operation can obtain a range-Doppler image. 4.The FMCW millimeter wave radar-based gait recognition method of claim 1, wherein, The method for training data using a clustering algorithm comprises the following steps: dividing all data into a training set and a test set on average; The event is represented as where e denotes the number of equal divisions of the time-frequency spectrum. taking all events of one category in the training set as a training data set; For each event , respectively calculate the Euclidean distance from the event to each cluster center, and divide the event to the cluster with the shortest Euclidean distance; ​ recalculating the center of each cluster according to the number of events in the cluster; taking the adjusted center point as the cluster center and repeating the above steps until the cluster no longer changes. 5.The FMCW millimeter wave radar-based gait recognition method of claim 4, wherein, The method for performing gait recognition according to the clustering result comprises the following steps: setting the threshold of the nearest neighbor algorithm as the number of clusters after clustering; For each data in the test set, the Euclidean distance between the data and the n cluster centers of the m categories is calculated, and then for each event, n values are obtained. ​ for the same test data, calculating the average distance of each event of the data to each category prototype, and then obtaining m average values; for each category, calculating the sum of the distances of all events of the test data to the category prototype, and then the category represented by the minimum value is the classification result of the test data; repeating the above steps to obtain the classification result of each test data, which is the result of gait recognition.

6. A computer readable storage medium storing a computer program for electronic data interchange, wherein, The computer program enables a computer to perform the method of any one of claims 1-5.

7. A gait recognition system based on FMCW millimeter wave radar, characterized by The method comprises the following steps: a data acquisition unit; a processor; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the processor, and the programs enable a computer to perform the method of any one of claims 1-5.

Images