An omni-directional human motion feature normalization method based on space-time motion path compensation

Abstract

The application discloses an omnidirectional human body action feature normalization method based on space-time motion path compensation. Firstly, the radar echo signal is processed, and a three-dimensional feature point cloud is constructed in the distance-angle-time space, and a motion path is estimated based on a space-time fitting motion path estimation algorithm; finally, the features of omnidirectional human body actions are compensated to radial human body action features through a feature compensation algorithm based on a motion direction angle, so that the omnidirectional human body action feature normalization is realized. The scheme starts from the fact that the motion direction leads to feature distortion of the human body action radar echo signal, analyzes and solves the angle sensitivity problem from the signal processing angle, restores the feature distortion caused by the motion angle, compensates the omnidirectional human body action features to radial features, eliminates the influence of the motion angle on the radar measurement signal, improves the feature representation efficiency, is superior to the traditional feature extraction algorithm in accuracy, efficiency and robustness, and improves the information utilization efficiency.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of radar human action recognition, and particularly relates to an omnidirectional human action feature normalization method based on space-time motion path compensation. BACKGROUND

[0002] Human action recognition is increasingly applied in public safety monitoring, smart home, medical health monitoring and human-computer interaction, and with the aggravation of social aging, the public's demand for healthy and safe life is continuously improved. Under this background, human action can cause the change of radar echo signal frequency, which is called micro-Doppler effect. Micro-Doppler effect can reveal the unique features of different human actions, thereby laying a solid foundation for radar-based human action recognition technology. Compared with contact sensors, radar sensors do not need to be close to the surface of the human skin to obtain data, and can be used for long-term monitoring; compared with optical sensors, radar-based human action recognition method can overcome the influence of dark and obstacles and other environments, and is less limited by the range of motion; in particular, since only the feature information of the target is carried in the radar echo signal, the visual form of the target will not be captured, so the personal privacy of the measured person can be protected. Due to its low environmental sensitivity (background, light), non-contact, all-weather operation, privacy protection, safety and reliability, and low cost, radar has been widely studied and applied in the field of human action recognition.

[0003] Existing radar sensor-based human action recognition methods mainly focus on radial human action, that is, the human action is close to or away from the radar along the direction of the radar main beam. They achieve human action recognition by extracting the micro-Doppler and distance features of human motion. However, in actual applications, the motion direction of human action is not limited to radial. When the motion direction of human action has an angle with the direction of the radar main beam, the radar detects the component of the real motion micro-Doppler and distance information in the radial direction. This leads to the deviation and distortion of the features measured by the radar from the real features. This deviation and distortion is closely related to the motion direction of the human body, and also makes the effect of human action recognition significantly decrease.

[0004] Therefore, omnidirectional human action recognition is a key problem for human action recognition to practical application, and existing omnidirectional action recognition methods are mainly divided into methods based on multi-base and single-base radar systems. The multi-base radar system can obtain multi-channel information in different directions through multi-system in different places. This deployment can monitor and fuse target information from different angles and ensure effective recording of human activities from a favorable angle, and through information integration of radar echo signals in different directions, the information loss caused by the existence of the motion angle of the single-base radar can be supplemented, which can provide good classification performance for omnidirectional human action recognition. However, the method based on the multi-base radar system still has some problems in practical application, including long time consumption, complex deployment, occupation of the site, and the like, and the calibration between multiple systems needs to be considered in the application, which is not conducive to application promotion. Therefore, the method based on the single-base radar system has attracted great attention. In the existing research methods, although most of the research based on the single-base radar system considers the angle information, they mainly focus on multiple recordings, construction of a large data set covering as many directions as possible, or design of a deep learning model to try to solve the problem. They all need a large amount of radar data as support, and lack of in-depth analysis of the specific influence of the motion direction on the human action features. SUMMARY

[0005] In view of the above problems, the purpose of the present application is to provide an omnidirectional human action feature normalization method based on space-time motion path compensation.

[0006] The specific technical scheme for achieving the purpose of the present application is as follows:

[0007] An omnidirectional human action feature normalization method based on space-time motion path compensation, comprising the following steps:

[0008] Step 1, obtaining the multi-transmit multi-receive radar reflection echo signal of the target, and pre-processing the echo signal to remove the clutter information in the signal, obtaining the Doppler-time data matrix DT and the range-time data matrix RT based on human action;

[0009] Step 2, based on the Doppler-time data matrix DT and the range-time data matrix RT, judging the effective frame through the Doppler envelope, extracting the Doppler feature D and the range feature R of the signal changing with time, and the target azimuth angle α, based on R, D and α, constructing a three-dimensional feature point cloud XYZ in the range-angle-time space;

[0010] Step 3, based on the three-dimensional feature point cloud XYZ, using a motion path estimation algorithm based on space-time fitting to estimate the motion path, obtaining the motion path direction angle β, and correcting the data abnormal points;

[0011] Step 4, based on the geometric relationship between the motion path and the radar main beam direction, the motion direction angle θ corresponding to the human motion feature is obtained, and the Doppler feature D and the distance feature R are restored to the radial feature distribution range D through a feature compensation algorithm corr and R corr , the omnidirectional action feature of the human body is compensated to a radial feature.

[0012] Compared with the prior art, the beneficial effects of the present application are:

[0013] (1) The omnidirectional human action feature normalization method based on space-time motion path compensation proposed in the present application starts from the fact that the motion direction causes the distortion of the human action radar echo signal feature, is based on a monostatic radar system, analyzes and solves the angle sensitivity problem from the perspective of signal processing, restores the feature distortion caused by the existence of the motion angle, compensates the omnidirectional human action feature to a radial feature, eliminates the influence of the motion angle on the radar measurement signal, and improves the representation efficiency of the feature, which is superior to the traditional feature extraction algorithm in accuracy, efficiency and robustness.

[0014] (2) The motion path estimation algorithm based on space-time fitting in the present application, compared with the traditional path estimation algorithm which usually only considers the path in a two-dimensional plane, pays attention to and considers the time relationship between the feature points, tracks the human motion path in the distance-angle-time three-dimensional feature point cloud, and corrects the abnormal data, which can avoid the abnormal change of the feature and ensure that the path estimation is logically smooth, improving the utilization efficiency of information.

[0015] (3) The feature point sequence extracted by the present application for human action recognition satisfies the discrimination degree of human action, effectively removes redundant information, and has smaller data volume and higher representation efficiency compared with the traditional spectrum feature, which greatly helps the efficiency and accuracy of subsequent network processing; at the same time, the present application compensates the omnidirectional activity feature to a radial feature through a feature compensation algorithm based on the moving direction to suppress feature distortion, can only use the recorded data of the radial angle, and can be extended to all angle cases, greatly reducing the cost of radar data collection, and can further improve the economic benefit and operation efficiency of the system.

[0016] The present application will be further described below in conjunction with the specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The present application is an omnidirectional human action feature normalization method based on space-time motion path compensation.

[0018] Figure 2 The present application is an echo data preprocessing schematic diagram.

[0019] Figure 3 Fig. 1 is a schematic diagram of feature extraction and three-dimensional feature point cloud construction according to the present application.

[0020] Figure 4 Fig. 2 is a schematic diagram of motion path estimation effect in an embodiment of the present application.

[0021] Figure 5 Fig. 3 is a geometric relationship between the motion path and the radar main beam direction according to the present application.

[0022] Figure 6 Fig. 4 is a schematic diagram of effect comparison of feature compensation algorithm in an embodiment of the present application. DETAILED DESCRIPTION

[0023] An omnidirectional human action feature normalization method based on space-time motion path compensation, comprising the following steps:

[0024] Step 1, obtaining the multi-input multi-output (MIMO) radar reflection echo signal of the target, and pre-processing the echo signal to remove clutter information in the signal, obtaining the Doppler-time data matrix DT and the range-time data matrix RT based on human action:

[0025] Step 1-1, obtaining the radar echo signal R rec , and performing data rearrangement: rec

[0026] According to the number of single chirp sampling points M, the number of antenna channels C, and the number of chirps N, the radar echo signal R rec is rearranged, and the entire echo signal is rearranged into an M×N×C three-dimensional radar echo data matrix R0, where m=[1,2,...,M] represents a fast time sampling point, n=[1,2,...,N] represents a slow time sampling point, and c=[1,2,...,C] represents a virtual antenna channel sequence;

[0027] Step 1-2, based on the radar echo data matrix R0, performing a fast Fourier operation on N fft points along the fast time dimension, obtaining a data matrix RPC temp , and since RPC temp has symmetry in the fast time dimension, taking half of it to obtain a range-time-channel three-dimensional data matrix RPC:

[0028]

[0029] RPC(r,n,c)=RPC temp (N fft / 2:N fft ,n,c) ​

[0030] Where, r temp =[1,2,...,N fft ] indicates RPC temp The distance gate index, r = [1, 2, ..., N] fft / 2] represents the distance gate index of RPC; the symbol "A:B" indicates indexing elements within the range from index A to index B in a specific dimension, N fft The number of Fourier operation points is set;

[0031] Steps 1-3: Perform Moving Target Indicator (MTI) clutter cancellation on the range-time-channel 3D data matrix RPC. Differential subtraction is performed along the slow time dimension to eliminate the echoes of stationary targets with the same amplitude and phase, thus obtaining the range-gated signal data matrix RPC after removing stationary targets. filtOut Select RPC filtOut From a single antenna channel, the RT data matrix is ​​obtained;

[0032]

[0033] RT(r,n) = RPC filiOut (r,n,1)

[0034] Steps 1-4: Based on the radar echo data matrix R0, select a single virtual antenna channel data matrix. Calculate the average value of each fast-time sampling point in the slow-time dimension as a reference value for background noise to obtain the noise-removed single-channel echo data matrix TPC. filtOut :

[0035]

[0036] TPC filtOut (m,n)=R0(m,n)-noise(m,n)

[0037] Steps 1-5: TPC based on single-channel echo data filtOut Select a single fast-time sampling point sequence and perform N operations in the slow-time dimension. stft The short-time fourier transform (STFT) of the points yields the DT data matrix:

[0038] Data dt (n)=TPC filtOut (pIdx,n)

[0039]

[0040] Where m1 = [1,2,…,N]stft ], n1 = [1, 2, …, N / WIN stft ], WIN stft is the window size of STFT transform.

[0041] Step 2, based on the Doppler-time (DT) data matrix DT and the Range-time (RT) data matrix RT of human action, the effective frame is determined by Doppler envelope, the Doppler feature D and the distance feature R of signal change over time, and the target azimuth angle a are extracted, and based on R, D and a, a three-dimensional feature point cloud XYZ in the range-angle-time space is constructed:

[0042] Step 2-1, based on the Doppler-time data matrix DT of human action, find all data greater than the set energy threshold P th in each time column, and take the maximum index and the minimum index as the upper and lower envelopes Eu and Ed respectively:

[0043] I(m1, n1) = {1 | DT(m1, n1) > P th}

[0044] Eu(n1) = maxidx(I(:, n1))

[0045] Ed(n1) = minidx(I(:, n1))

[0046] Wherein, minidx(I(:, n1)) and maxidx(I(:, n1)) are the minimum and maximum index values of the index I(:, n1);

[0047] Step 2-2, divide the upper and lower envelopes Eu and Ed into N f subsections, according to and and the minimum offset of the upper envelope offset1 and the minimum offset of the lower envelope offst2, the corresponding frame meeting the condition is taken as the effective frame for subsequent calculation:

[0048]

[0049] Ed i = Ed((i-1)*V+1:V*i)

[0050] Eu i = Eu((i-1)*V+1:V*i)

[0051] Wherein n f = [1, 2, …, N f ] represents the sequence of divided data frames, and N fV = N / WIN stft / N f denotes the length of each sub-segment, and effFrame(n f ) is considered as an effective frame when it is equal to 1;

[0052] Step 2-3, select the distance-time data matrix RT corresponding to the time of the effective frame, obtain the corresponding distance by calculating the maximum row and the number of rows, find the corresponding maximum envelope, and obtain the corresponding Doppler feature D and distance feature R:

[0053]

[0054] r norm = 20log 10 (|r frame |)

[0055]

[0056] R(index) = (maxrow + 256) * ΔR

[0057] D(index) = max(Eu index )

[0058] wherein index is the effective frame number obtained by effFrame in step 2-2; N chirp is the number of chirps corresponding to each frame in the data matrix; wherein the symbol argmax represents the index of the maximum value; ΔR is the distance gate width;

[0059] Step 2-4, according to the radar antenna arrangement, select the corresponding virtual channel in TPC to form the azimuth angle data matrix, and perform angle estimation by multiple signal classification (MUSIC) algorithm to obtain the target azimuth angle a of the corresponding frame;

[0060] Data A1 (n,c array ) = R0(pIdx,n,C array1 )

[0061] Data A2 (n,c array ) = R0(pIdx,n,C array2 )

[0062]

[0063] Rxx = U·S·V

[0064]

[0065] wherein c array =[1,2,…,z] is the virtual antenna channel sequence; C array1 , C array2 represent two groups of virtual channel numbers for azimuth angle detection, respectively; N s is the number of detection targets, and Rxx is the array manifold vector of the channel domain data matrix; is the array manifold vector; α takes the average of the angles calculated from the two channel domain data matrices, Data Ai is the Data A1 or Data A2 ;

[0066] Step 2-5, determine the planar motion sequence point coordinates XY according to the distance feature R and the azimuth angle α, generate the relative time axis coordinates according to the number of effective frames, and construct the three-dimensional feature point cloud XYZ in the distance-angle-time space:

[0067] X(index) = R(index) * sin(α(index))

[0068] Y(index) = R(index) * cos(α(index))

[0069] Z(index) = t start + 0.2 * index

[0070] XYZ = [X; Y; Z]

[0071] wherein t start is the set starting time point.

[0072] Step 3, based on the three-dimensional feature point cloud XYZ, use the motion path estimation algorithm based on space-time fitting to estimate the motion path, obtain the motion path direction angle β, and correct the data abnormal points:

[0073] Step 3-1, initialize the neighborhood width gate of the straight line, set the minimum core point ratio as rate, and set the maximum offset degree of the point set as bestd;

[0074] Step 3-2, randomly select any two points A and B in the three-dimensional feature point cloud XYZ, connect them into a straight line L k , calculate the distance from the remaining points in the three-dimensional feature point cloud XYZ to the straight line:

[0075]

[0076] wherein A and B represent two points in the kth combination; k = [1, 2, …, K] represents the index of all point combinations; L kThis represents the straight line formed by connecting the k-th points; Let x be the i-th valid point other than A and B. i y i z i These are the coordinates of the point; the symbol ‖·‖ represents the magnitude of the vector;

[0077] Step 3-3: The points obtained in Step 3-2 are connected to line L. k The distance is used to determine whether each point is on the straight line L. k Neighborhood points:

[0078] N in ={XYZ i |d i ≤gate}

[0079] N out ={XYZ i |d i gate}

[0080] Where, N in Let d represent a set containing all sets that satisfy d. i Points ≤ gate are considered normal points; correspondingly, N out This indicates an outlier group;

[0081] Step 3-4: From each point obtained in step 3-2, transfer it to line L. k The distance is calculated by excluding the data points with the largest and second largest distances to reduce the influence of outliers on the fitted line, thus obtaining the degree of deviation between the point set and the line:

[0082] distance=Σd i -d max1 -d max2

[0083] Where, d max1 For all d i The maximum value in; d max2 For all d i The second largest value in;

[0084] Step 3-5: Based on the point set N obtained in steps 3-3 and 3-4 in And distance, determine the line L k If the line is the optimal line, update the parameters; otherwise, skip to step 3-2 for iteration until all point combinations are traversed to obtain the optimal combination point S. b1 and S b2 :

[0085] if|Nin | ≥ | XYZ | * rate and distance < bestd then

[0086] bestd = distance

[0087] else continue

[0088] wherein the symbol | · | represents the number of point sets;

[0089] Step 3-6, the optimal combination point S determined based on the result of step 3-5 b1 and S b2 , the abnormal point correction is realized by using the space-time based abnormal point correction method, a plane perpendicular to the time axis and corresponding to the abnormal value timestamp is selected as the correction plane to realize the correction point, that is, the abnormal point group N out is corrected by its time characteristics in the space-time dimension to obtain the corrected correct point set N c :

[0090] N out = N (x out ,y out ,z out )

[0091] z out = z corr

[0092]

[0093] N corr = N (x corr ,y corr ,z corr )

[0094] N c = N in +N corr

[0095] wherein N corr is the point group obtained after correction of the abnormal point group N out , y corr , x corr are the horizontal and vertical coordinates of the correction point obtained by solving the collinear point relationship in three-dimensional space;

[0096] Step 3-7, the corrected point set N c = N (x c ,y c ,z c ) obtained by step 3-6, according to the XY coordinates of N c , the azimuth angle α corresponding to each valid point is obtained, and the optimal combination point S obtained according to step 3-5 is fittedb1 and S b2 , obtain the motion path angle β:

[0097]

[0098]

[0099] α=A c ·μ

[0100]

[0101] wherein μ is a sign correction factor, used to determine the correct sign of the azimuth angle, A c represents the azimuth angle of each point in the point group N c .

[0102] Step 4, based on the geometric relationship between the motion path and the direction of the radar main beam, obtain the motion direction angle θ corresponding to the human motion feature, and restore the Doppler feature D and the distance feature R to the radial feature distribution range D corr and R corr , compensate the omnidirectional action feature of the human body to the radial feature:

[0103] Step 4-1, based on the point set N c obtained in step 3-7, calculate the distance between each point in N c and the first point of the motion sequence, obtain the real distance span of the target motion, and thus compensate the original distance feature R to the radial distance feature R corr :

[0104]

[0105] wherein x c1 , y c1 are the horizontal coordinate and the vertical coordinate of the first point in the point set N c .

[0106] Step 4-2, according to the determined azimuth angle α and the motion path direction angle β corresponding to each effective point, according to the geometric relationship between the motion path and the direction of the radar main beam, obtain the corresponding motion direction angle θ, compensate the Doppler feature through θ, and obtain the radial Doppler feature D corr :

[0107] θ=α+β

[0108] D corr =D / cos(θ)

[0109] The application also provides an omnidirectional human action feature normalization system based on space-time motion path compensation, comprising the following modules:

[0110] a radar echo signal processing module for obtaining multi-transmit multi-receive radar reflection echo signals of a target, and pre-processing the echo signals to remove clutter information in the signals, to obtain a Doppler-time data matrix DT and a range-time data matrix RT based on human action;

[0111] a feature extraction module for judging valid frames based on the Doppler-time data matrix DT and the range-time data matrix RT, extracting Doppler features D and range features R of signals changing with time, and a target azimuth angle a, and constructing a three-dimensional feature point cloud XYZ in a range-angle-time space based on R, D and a;

[0112] a path estimation module for estimating a motion path based on the three-dimensional feature point cloud XYZ, using a motion path estimation algorithm based on space-time fitting, obtaining a motion path direction angle β, and correcting data abnormal points;

[0113] a radial feature restoration module for obtaining a motion direction angle θ corresponding to a human motion feature based on a geometric relationship between the motion path and a radar main beam direction, and restoring Doppler features D and range features R to a radial feature distribution range D corr and R corr by a feature compensation algorithm.

[0114] A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements the following steps when executing the computer program:

[0115] Step 1, obtaining multi-transmit multi-receive radar reflection echo signals of a target, and pre-processing the echo signals to remove clutter information in the signals, to obtain a Doppler-time data matrix DT and a range-time data matrix RT based on human action;

[0116] Step 2, judging valid frames based on the Doppler-time data matrix DT and the range-time data matrix RT, extracting Doppler features D and range features r of signals changing with time, and a target azimuth angle a, and constructing a three-dimensional feature point cloud XYZ in a range-angle-time space based on R, D and a;

[0117] Step 3, estimating a motion path based on the three-dimensional feature point cloud XYZ, using a motion path estimation algorithm based on space-time fitting, obtaining a motion path direction angle β, and correcting data abnormal points;

[0118] Step 4, based on the geometric relationship between the motion path and the radar main beam direction, the motion direction angle θ corresponding to the human motion feature is obtained, and the Doppler feature D and the distance feature R are restored to the radial feature distribution range D through a feature compensation algorithm corr and R corr compensate the omnidirectional action feature of the human body to the radial feature.

[0119] A computer storage medium, which stores a computer program, the computer program is executed by a processor to implement the following steps:

[0120] Step 1, obtain the multi-transmit multi-receive radar reflection echo signal of the target, and pre-process the echo signal to remove the clutter information in the signal, obtain the Doppler-time data matrix DT and the range-time data matrix RT based on human action;

[0121] Step 2, based on the Doppler-time data matrix DT and the range-time data matrix RT of the human action, the effective frame is judged through the Doppler envelope, the Doppler feature D and the range feature R of the signal changing with time, and the target azimuth angle α are extracted, and based on R, D and α, a three-dimensional feature point cloud XYZ in the range-angle-time space is constructed;

[0122] Step 3, based on the three-dimensional feature point cloud XYZ, the motion path is estimated by using the motion path estimation algorithm based on space-time fitting, the motion path direction angle β is obtained, and the data abnormal points are corrected;

[0123] Step 4, based on the geometric relationship between the motion path and the radar main beam direction, the motion direction angle θ corresponding to the human motion feature is obtained, and the Doppler feature D and the distance feature R are restored to the radial feature distribution range D through a feature compensation algorithm corr and R corr compensate the omnidirectional action feature of the human body to the radial feature.

[0124] Embodiments

[0125] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. The described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0126] As used in this application and the appended claims, the singular forms "a," "an," and "the" include plural referents unless the context clearly dictates otherwise. Thus, for example, reference to "a" or "the" component can include from one or more of such components, and reference to "the" method can include practice of one or more of such methods, and so forth. Similarly, words of approximation such as "approximately," "substantially," or "about" which will be understood by those of ordinary skill in the art, can be used herein in connection with at least one of the terms they modify throughout the entire application unless specifically indicated otherwise. As used herein, the term "include" and its variants are intended to be equivalent to the term "comprise," and are intended to be used in their broadest sense. Similarly, the term "comprising" is used herein to mean, and is used in its broadest sense, that "comprising" can be replaced with "consisting essentially of" or "consisting of."

[0127] The relative arrangement of parts and steps, numerical expressions, and numerical values set forth in the examples herein are not limiting to the scope of the present application unless otherwise specifically indicated. It should also be understood that the drawings are not necessarily drawn to scale of the actual proportions of the parts being depicted. Techniques, methods, and apparatus known to those of ordinary skill in the relevant art can not be discussed in detail herein. However, the techniques, methods, and apparatus are fully intended to be part of the description of the application. In all examples shown and discussed herein, any specific values are to be interpreted as merely illustrative and not limiting. Other examples of the exemplary embodiments can therefore have different values. It is to be noted that like reference numerals and letters refer to like items in the drawings, and once an item is defined in one drawing, it need not be discussed further in subsequent drawings.

[0128] In conjunction with Figure 1 A method for omnidirectional human action feature normalization based on space-time motion path compensation, comprising the following steps:

[0129] Step 1, the radar system is elevated to 0.85 meters from the ground, and the volunteer faces the radar to perform falling in different directions. The multi-input multi-output (MIMO) radar reflection signal of the target is obtained, and the echo signal is preprocessed, including data rearrangement, removal of direct current component, MTI processing (removal of static clutter interference), to obtain human action echo signal with characteristic distortion due to non-radar radial direction of motion, as shown in Figure 2

[0130] Remove clutter information in the signal to obtain Doppler-time data matrix DT and range-time data matrix RT based on human action:

[0131] Step 1-1, obtain radar echo signal R rec , data rearrangement is performed on the radar echo signal R rec

[0132] The radar platform sends a signal each time, which is called a chirp signal, and the received radar echo signal R rec ​​The two LVDS channels are constituted in a non-interleaved format and stored in a binary file, each chirp sampling point is a 16-bit unsigned integer, and two bytes are required for storage; according to a single chirp sampling point number M = 64, an antenna channel number C = 2 * 4 = 8, and a chirp number N = 250 * 128 = 32000, the radar echo signal R rec is rearranged to form an M * N * C three-dimensional radar echo data matrix R0, where m = [1, 2,..., 64] represents a fast time sampling point, n = [1, 2,..., 32000] represents a slow time sampling point, and c = [1, 2,..., 8] represents a virtual antenna channel sequence;

[0133] Step 1-2, based on the radar echo data matrix R0, N fft point fast Fourier FFT operation is performed along the fast time dimension to obtain a data matrix RPC temp , since RPC temp has symmetry in the fast time dimension, half of it is taken to obtain a distance-time-channel three-dimensional data matrix RPC:

[0134]

[0135] RPC(r,n,c) = RPC temp (N fft / 2:N fft ,n,c)

[0136] wherein r temp = [1, 2,..., N fft ] represents a distance gate index of RPC temp , r = [1, 2,..., N fft / 2] represents a distance gate index of RPC; the symbol “A: B” represents indexing elements in a specific dimension from sequence number A to sequence number B, and N fft is a set Fourier operation point number;

[0137] Step 1-3, moving target indicator (MTI) clutter elimination is performed on the distance-time-channel three-dimensional data matrix RPC, and the amplitude and phase of the stationary target echo are eliminated by difference along the slow time; since the human body is always moving (such as breathing), the amplitude and phase of the reflected signal are different, and the stationary object can be removed by time-by-time subtraction, and the signal that changes is retained, to obtain a distance gate signal data matrix RPC filtOut after removing the stationary target, and a single antenna channel in RPC filtOut is selected to obtain an RT data matrix;

[0138]

[0139] RT(r,n)=RPC filtOut (r,n,1)

[0140] Step 1-4, based on the radar echo data matrix R0, a single virtual antenna channel data matrix is selected, and by calculating the average value of each fast time sampling point in the slow time dimension, a reference value of background noise is obtained, and a single-channel echo data matrix TPC after removing noise is obtained filtOut :

[0141]

[0142] TPC filtOut (m,n)=R0(m,n)-noise(m,n)

[0143] Step 1-5, based on the single-channel echo data TPC filtOut , a single fast time sampling point sequence is selected, and a short-time Fourier transform (STFT) of N stft point is performed in the slow time dimension to obtain a DT data matrix:

[0144] Data dt (n)=TPC filtout (pIdx,n)

[0145]

[0146] Wherein, m1=[1,2,…,N stft ], n1=[1,2,…,N / WIN stft ], WIN stft is the window size of STFT, pIdx=50 is the serial number of the selected fast time sampling point in the single chirp signal, WIN stft =128 is the sliding window size of STFT; N stft =640 is the FFT point number of each sliding window of STFT;

[0147] Step 2, based on the Doppler-time (DT) data matrix DT and the range-time (RT) data matrix RT of human body action, the effective frame is judged by Doppler envelope, the Doppler feature D and the distance feature R of signal change with time, and the target azimuth angle a are extracted, and based on R, D and a, a three-dimensional feature point cloud XYZ in the range-angle-time space is constructed, as shown in Figure 3 ;

[0148] Step 2-1, find all values greater than a set energy threshold P in each time column of the Doppler-time data matrix DT based on human motion th From these qualified data, take the maximum index and the minimum index as the upper and lower envelopes Eu and Ed respectively:

[0149] I(m1,n1) = {1 | DT(m t ,n1) > P th}

[0150] Eu(n1) = maxidx(I(:,n1))

[0151] Ed(n1) = minidx(I(:,n1))

[0152] Wherein, minidx(I(:,n1)) and maxidx(I(:,n1)) are the minimum and maximum index values in the index I(:,n1);

[0153] Step 2-2, divide the upper and lower envelopes Eu and Ed into N f = 23 sub-sections, according to and and the minimum offset of the upper envelope offset1 and the minimum offset of the lower envelope offset2, the corresponding frames that meet the conditions are taken as valid frames for subsequent calculation:

[0154]

[0155]

[0156] Ed i = Ed((i-1)*V+1:V*i)

[0157] Eu i = Eu((i-1)*V+1:V*i)

[0158] Wherein n f = [1, 2,..., N f ] represents the sequence of divided data frames, N f is the number of divided data frames, V = N / WIN stft / N f represents the length of each subsection, when effFrame(n f ) is 1, it is considered as a valid frame;

[0159] In this embodiment, offset1 = 100 and offset2 = -80 are the effective frame judgment thresholds;

[0160] Step 2-3, select the distance-time data matrix RT corresponding to the time of the effective frame, obtain the corresponding distance by calculating the maximum row and the corresponding row number, find the corresponding maximum envelope, and obtain the corresponding Doppler feature D and distance feature R:

[0161]

[0162] r norm =20log 10 (|r frame |)

[0163]

[0164] R(index)=(maxrow+256)*ΔR

[0165] D(index)=max(Eu index )

[0166] Wherein, index is the effective frame number obtained by effFrame in step 2-2; N chirp =2560 is the number of chirps corresponding to each frame in the data matrix; wherein the symbol argmax represents the index of the maximum value; is the distance gate width;

[0167] Step 2-4, according to the radar antenna arrangement, select the corresponding virtual channel in TPC to form the azimuth angle data matrix, and estimate the angle by multiple signal classification (MUSIC) algorithm to obtain the target azimuth angle α of the corresponding frame;

[0168] Data A1 (n,c array )=R0(pIdx,n,C arra )

[0169] Data A2 (n,c array )=R0(pIdx,n,C array2 )

[0170]

[0171] Rxx=U·S·V

[0172]

[0173] Wherein, c array =[1,2,3,4] is the virtual antenna channel sequence; C arra =[1,4,5,8], C array2= [2, 3, 6, 7] represent two groups of virtual channel numbers for azimuth angle detection; N s Rxx is the array manifold vector of channel domain data matrix for detecting target number; Rxx is the array manifold vector; a is the average of the angles calculated from two channel domain data matrices, Data Ai Data A1 or Data A2 ;

[0174] Step 2-5, determine the plane motion sequence point coordinates XY according to the distance feature R and the azimuth angle a, generate the relative time axis coordinates according to the number of effective frames, and construct the three-dimensional feature point cloud XYZ in the distance-angle-time space:

[0175] X(index) = R(index) * sin(a(index))

[0176] Y(index) = R(index) * cos(a(index))

[0177] Z(index) = t start + 0.2 * index

[0178] XYZ = [X; Y; Z]

[0179] where t start = 0.4 is the set starting time point.

[0180] Step 3, based on the three-dimensional feature point cloud XYZ, use the motion path estimation algorithm based on space-time fitting to estimate the motion path, obtain the motion path direction angle β, and correct the data abnormal points, as shown in Figure 4

[0181] Step 3-1, initialize the neighborhood width of the straight line gate = 0.15, set the minimum core point ratio rate = 0.5, and the maximum offset degree of the point set bestd = 100;

[0182] Step 3-2, randomly select any two points A and B in the three-dimensional feature point cloud XYZ, connect them into a straight line L k , calculate the distance from the remaining points in the three-dimensional feature point cloud XYZ to the straight line:

[0183]

[0184]

[0185] where A and B represent two points in the kth combination; k = [1, 2, …, K] represents the index of all point combinations; L k ​represents the straight line formed by the kth combination of points; x i , y i , z i are the coordinates of the ith valid point other than A and B, respectively; the symbol ||·|| represents the modulus of a vector;

[0186] Step 3-3, the distance of each point obtained by step 3-2 to the straight line L k , determines whether each point is a neighborhood point of the straight line L k :

[0187] N in ={XYZ i |d i ≤gate}

[0188] N out ={XYZ i |d i >gate}

[0189] wherein, N in represents a set, and this set contains all points satisfying d i ≤gate, which is a normal point group; correspondingly, N out represents an abnormal point group;

[0190] Step 3-4, the distance of each point obtained by step 3-2 to the straight line L k , by excluding the data points with the largest and second largest distance, the influence of abnormal points on the fitting straight line is reduced, and the degree of deviation distance of the point set from the straight line is obtained:

[0191] distance=∑d i -d max1 -d ma

[0192] wherein, d max1 is the maximum value of all d i ; d max2 is the second largest value of all d i ;

[0193] Step 3-5, according to the point set N in and distance obtained by steps 3-3 and 3-4, determines whether the straight line L k is the current optimal straight line, if it is the optimal straight line, the parameters are updated, if it is not the optimal straight line, jumps to step 3-2 for iteration, until all point set combinations are traversed, to obtain the optimal combination of points S b1 and S b2 :

[0194] if|Nin |≥|XYZ|*rateanddistance <bestd then

[0195] bestd = distance

[0196] else continue

[0197] where |·| denotes the number (i.e. length) of a point set;

[0198] Step 3-6, the optimal combination point S determined based on the result of step 3-5 b1 and S b2 , the abnormal point correction method based on space-time is used to realize the abnormal point correction, the plane perpendicular to the time axis and corresponding to the abnormal value timestamp is selected as the correction plane to realize the correction point, that is, the abnormal point group N out is corrected by its time characteristics in the space-time dimension to obtain the corrected correct point set N c :

[0199] N out = N(x out ,y out ,z out )

[0200] z out = z corr

[0201]

[0202] N corr = N(x corr ,y corr ,z corr )

[0203] N c = N in +N corr

[0204] where N corr is the point group obtained after correction of the abnormal point group N out , y corr and x corr are the horizontal and vertical coordinates of the correction point obtained by solving the collinear point relationship in three-dimensional space;

[0205] Step 3-7, the corrected point set N c = N(x c ,y c ,z c ) obtained by step 3-6, according to N cThe azimuth angle a corresponding to each valid point is obtained according to the XY coordinates, and the combined point S obtained in step 3-5 is fitted to be optimal b1 and S b2 , the motion path angle β is obtained

[0206]

[0207] a = A c · μ

[0208]

[0209] Wherein, μ is a sign correction factor, used to determine the correct sign of the azimuth angle, A c represents the azimuth angle of each point in the N c point group.

[0210] Step 4, based on the geometric relationship between the motion path and the direction of the radar main beam, the motion direction angle θ corresponding to the human motion feature is obtained, as shown in Figure 5 The Doppler feature D and the distance feature R are restored to the radial feature distribution range D corr and R corr , and the omnidirectional action feature of the human body is compensated to the radial feature, as shown in Figure 6 :

[0211] Step 4-1, set the action starting point distance R0 = 3, so that the initial distance is a fixed value, highlight the distance feature according to the characteristics of different actions, based on the point set N c corrected in step 3-7, the distance between each point in N c and the first point of the motion sequence is calculated, the real distance span of the target motion is obtained, and the original distance feature R is compensated to the radial distance feature R corr :

[0212]

[0213]

[0214] Wherein, x c1 , y c1 are the horizontal coordinate and vertical coordinate of the first point in the point set N c :

[0215] Step 4-2, according to the azimuth angle a corresponding to each valid point and the motion path direction angle β, according to the geometric relationship between the motion path and the direction of the radar main beam, the corresponding motion direction angle θ is obtained, and the radial Doppler feature D corr is obtained by compensating the Doppler feature through θ:

[0216] θ = a + β

[0217] D corr = D / cos (theta)

[0218] By Figure 4 and Figure 6 It can be seen that after the omnidirectional human motion feature normalization method based on space-time motion path compensation is proposed, the human motion features (Doppler and distance features) recovered by the feature compensation algorithm are more consistent with the radial motion feature distribution, Figure 6 In the figure, blue is the feature before compensation, orange is the feature after compensation, and yellow is the feature measured in the radial case. It can be seen that when the motion angle is not in the main beam of the radar, the Doppler and distance features of the human motion before compensation are distorted, and the Doppler is attenuated to half of the true value, while the Doppler and distance features of the human motion after compensation return to the normal range. Similarly, Figure 4 In the figure, blue points are normal points, purple points are abnormal points, red points are corrected points of abnormal points, and green straight lines are fitted straight lines. It can be proved from the figure that the motion path estimation based on space-time fitting can correctly extract the human motion path angle, and the time sequence feature of human motion can be used to correct abnormal points in space and logic.

[0219] The above-described embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it cannot be understood as limiting the scope of the patent. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are within the scope of protection of the present application. Therefore, the scope of protection of the patent of the present application should be subject to the appended claims.

Claims

1. A method for omnidirectional human action feature normalization based on space-time motion path compensation, characterized in that, The method comprises the following steps: Step 1, obtaining a multi-transmitting and multi-receiving radar reflection echo signal of a target, and pre-processing the echo signal to remove clutter information in the signal, and obtaining a Doppler-time data matrix DT and a range-time data matrix RT based on human action; Step 2, judging an effective frame through a Doppler envelope based on the Doppler-time data matrix DT and the range-time data matrix RT, extracting Doppler features D and range features R of the signal changing with time, and a target azimuth angle α, and constructing a three-dimensional feature point cloud XYZ in a range-angle-time space based on R, D and α; Step 3, estimating a motion path based on the three-dimensional feature point cloud XYZ by using a motion path estimation algorithm based on space-time fitting, obtaining a motion path direction angle β, and correcting data abnormal points: Step 3-1, initializing a neighborhood width gate of a straight line, setting a minimum core point ratio as rate, and setting a maximum offset degree of a point set as bestd; Step 3-2, randomly select any two points A and B in the three-dimensional feature point cloud XYZ, and connect them into a straight line L k , calculate the distance from the remaining points in the three-dimensional feature point cloud XYZ to the straight line: where A and B represent two points in the kth combination; k = [1, 2, …, K] represents the index of all point combinations; L k represents a straight line formed by the kth point combination; xi, yi, zi i , y i , z i are the coordinates of the point, respectively; the symbol ||·|| represents the modulus of the vector; Step 3-3, judging whether each point is a neighborhood point of the straight line L by the distance from each point to the straight line L obtained by step 3-2 k Step 3-4, if the point is a neighborhood point of the straight line L, then the point is a point on the straight line L k Step 3-5, if the point is not a neighborhood point of the straight line L, then the point is not a point N in = { XYZ i | d i ≤ gate} N out = { XYZ i | d i > gate} where N in represents a set containing all points satisfying d i ≤ gate, which is the normal point set; correspondingly, N out represents the abnormal point set; Step 3-4: From each point obtained in step 3-2, transfer it to line L. k The distance is calculated by excluding the data points with the largest and second largest distances to reduce the influence of outliers on the fitted line, thus obtaining the degree of deviation between the point set and the line: distance =∑d i - d max1 - d max2 where d max1 is the maximum value among all d i ; d max2 is the second maximum value among all d i ; Step 3-5, judging whether the straight line L in is the current optimal straight line according to the distance between the point set N k and S b1 and updating the parameters if it is the optimal straight line, or jumping to step 3-2 for iteration if it is not the optimal straight line, until all point set combinations are traversed to obtain the combination point S b2 ​ if |N in |≥|XYZ|*rate and distance <bestd then bestd = distance else continue Wherein, the symbol |·| represents the number of point sets; Step 3-6: The optimal combination point S determined based on the result of step 3-5 b1 and S b2 , the abnormal point correction is realized by using the space-time based abnormal point correction method, a plane perpendicular to the time axis and corresponding to the abnormal value timestamp is selected as the correction plane to realize the correction point, that is, the abnormal point group N out The space-time dimension is corrected by its time characteristics, and the corrected correct point set N c : N out = N(x out ,y out ,z out ) z out = z corr N corr = N(x corr ,y corr ,z corr ) N c = N in + N corr wherein N corr is the group of abnormal points N out is the group of points y corr , x corr are the horizontal and vertical coordinates of the corrected points obtained by solving the collinear point relationship in three-dimensional space; Step 3-7, the revised point set N obtained by step 3-6 c = N(x c ,y c ,z c ), according to the XY coordinates of N c , the azimuth angle α corresponding to each valid point is obtained, and the motion path angle β is obtained according to the fitting optimal combination point S b1 and S b2 obtained by step 3-5 a = A c • μ where μ is a sign correction factor used to determine the correct sign of the azimuth angle, A c denotes N c the azimuth angle of each point in the point group; Step 4, based on the geometric relationship between the motion path and the radar main beam direction, the motion direction angle θ corresponding to the human motion feature is obtained, and the Doppler feature D and the distance feature R are restored to the radial feature distribution range D through the feature compensation algorithm corr and R corr , the omnidirectional action feature of the human body is compensated to the radial feature. 2.The omnidirectional human action feature normalization method based on space-time motion path compensation according to claim 1, characterized in that, The pre-processing of the radar echo signal in step 1 is specifically: Step 1-1, obtaining radar echo signal R rec , the radar echo signal R rec Data rearrangement: According to the number of single chirp sampling points M, the number of antenna channels C, and the number of chirps N, the radar echo signal R rec The entire echo signal is rearranged into an MxNxC three-dimensional radar echo data matrix R0, where m=[1, 2,..., M] represents a fast time sampling point, n=[1, 2,..., N] represents a slow time sampling point, and c=[1, 2,..., C] represents a virtual antenna channel sequence. Step 1-2, based on the radar echo data matrix R0, along the fast time dimension, N fft point Fourier operation, get the data matrix RPC temp , since RPC temp has symmetry in the fast time dimension, take half of it, get the distance-time-channel three-dimensional data matrix RPC: RPC(r, n, c) = RPC(r, n, c) temp (N fft / 2: N fft , n, c) wherein r temp = [1, 2,..., N fft ] represents the distance gate index of the RPC temp , and r = [1, 2,..., N fft / 2] represents the distance gate index of the RPC; the symbol "A:B" represents indexing elements in a certain dimension from the sequence number A to the sequence number B, N fft is the set number of Fourier operation points; Step 1-3, moving target display clutter elimination is performed on the distance-time-channel three-dimensional data matrix RPC, difference is subtracted along the slow time dimension, amplitude and phase of the stationary target echo are eliminated, and a distance gate signal data matrix RPC after removing the stationary target is obtained filtOut , a single antenna channel in the RPC filtOut is selected to obtain an RT data matrix; RT(r,n) = RPC filtOut (r,n,1) Step 1-4, based on the radar echo data matrix R0, select a single virtual antenna channel data matrix, by calculating the average value of each fast time sampling point in the slow time dimension, as the reference value of the background noise, obtain the single channel echo data matrix TPC after removing the noise filtOut : TPC filtOut (m, n) = R0(m, n) - noise(m, n) Step 1-5, TPC based on single channel echo data filtOut , select a single fast time sampling point sequence, and perform a short time Fourier transform on the N stft point dimension to obtain a DT data matrix: Data dt (n) = TPC filtOut (pIdx,n) where m1 = [1, 2, …, N stft ], n1 = [1, 2, …, N / WIN stft ], WIN stft is the window size of the STFT transform. 3.The omnidirectional human action feature normalization method based on space-time motion path compensation according to claim 2, characterized in that, The construction of the three-dimensional feature point cloud in the range-angle-time space in step 2 is specifically: Step 2-1, based on the human motion Doppler-time data matrix DT, find all values greater than a set energy threshold P in each time column th The data, respectively, take the maximum index and the minimum index as the upper and lower envelopes Eu and Ed: I(m1,n1) = { 1 | DT(m1,n1) > P th} Eu(n1) = maxidx(I(:,n1)) Ed(n1) = minidx(I(:,n1)) Wherein, minidx(I(:,n1)) and maxidx(I(:,n1)) are the minimum and maximum index values of the index I(:,n1); Step 2-2, divide the upper and lower envelopes Eu and Ed into N sub-segments, according to f and and and the upper envelope minimum offset offset1 and the lower envelope minimum offset offset2, the corresponding frame meeting the condition is taken as a valid frame for subsequent calculation: Ed i = Ed((i-1)*V+1:V*i) Eu i = Eu((i-1)*V+1:V*i) where n f = [1, 2,..., N f ] represents the sequence of divided data frames, N f is the number of divided data frames, V = N / WIN stft / N f represents the length of each sub-segment, and effFrame(n f ) is considered as an effective frame when it is 1. Step 2-3, selecting a range-time data matrix RT corresponding to the time of the effective frame, obtaining the corresponding distance by calculating the maximum row and the corresponding row number, finding the corresponding maximum envelope, and obtaining the corresponding Doppler features D and range features R: r norm = 20 log 10 (|r frame |) R(index) = (maxrow+256)*ΔR D(index) = max(Eu index ) wherein index is the valid frame number obtained from effFrame in step 2-2; N chirp is the number of chirps corresponding to each frame in the data matrix; wherein the symbol argmax represents the index corresponding to the maximum value; AR is the range gate width; Step 2-4, selecting a corresponding virtual channel in the TPC according to the radar antenna arrangement, forming an azimuth angle data matrix, and estimating the angle by a multiple signal classification algorithm to obtain the target azimuth angle α of the corresponding frame; Data A1 (n,c array )=R0(pIdx,n,C array1 ) Data A2 (n,c array )=R0(pIdx,n,C array2 ) Rxx = U·S·V wherein c array = [1, 2, …, z] is a virtual antenna channel sequence; C array1 , C array2 respectively represent two groups of virtual channel numbers for azimuth angle detection; N s is the number of detection targets, and Rxxis an array manifold vector of the channel domain data matrix; is an array manifold vector; a takes the average of the angles calculated from the two channel domain data matrices, Data Ai is Data A1 or Data A2 ; Step 2-5, determining the plane motion sequence point coordinates XY according to the range features R and the azimuth angle α, generating a relative time axis coordinate according to the number of effective frames, and constructing a three-dimensional feature point cloud XYZ in a range-angle-time space: X(index) = R(index)*sin(α(index)) Y(index) = R(index)*cos(α(index)) Z(index) = t start + 0.2 * index XYZ = [X;Y;Z] wherein t start is a set start time point. 4.The space-time motion path compensated omnidirectional human action feature normalization method according to claim 1, characterized in that, The step 4 is to restore the Doppler feature D and the distance feature R to the radial feature distribution range D by a feature compensation algorithm corr and R corr compensate the omnidirectional action features of the human body to radial features, specifically: Step 4-1, based on the revised point set N obtained in step 3-7 c , calculate the distance of each point in N c to the first point of the motion sequence, obtain the real distance span of the target motion, and compensate the original distance feature R to the radial distance feature R corr : where x c1 , y c1 are the horizontal and vertical coordinates of the first point in the point set N c , respectively. Step 4-2, according to the determined azimuth angle a and the motion path direction angle β corresponding to each valid point, the corresponding motion direction angle θ is obtained according to the geometric relationship between the motion path and the radar main beam direction, and the radial Doppler feature D is obtained by compensating the Doppler feature through θ corr : θ = α+β D corr = D / cos(θ).

5. An omni-directional human motion feature normalization system based on space-time motion path compensation, characterized by, The method comprises the following modules: A radar echo signal processing module; A radar echo signal processing module; A radar echo signal processing module; Feature extraction module: for based on human motion Doppler-time data matrix DT, distance-time data matrix RT, through the Doppler envelope to determine the effective frame, extract the Doppler feature D, distance feature R and target azimuth angle a of signal change with time, based on R, D and a, construct three-dimensional feature point cloud XYZ in distance-angle-time space; Path estimation module: for based on three-dimensional feature point cloud XYZ, using motion path estimation algorithm based on space-time fitting to estimate the motion path, obtain the motion path direction angle β, correct the data abnormal points: Initialize the neighborhood width gate of the straight line, set the minimum core point ratio as rate, and the maximum offset degree of the point set as bestd; Randomly select any two points A and B in the three-dimensional feature point cloud XYZ, connect them into a straight line L k , calculate the distance from the remaining points in the three-dimensional feature point cloud XYZ to the straight line: where A and B represent two points in the kth combination; k = [1, 2, …, K] represents the index of all point combinations; L k represents a straight line formed by the kth point combination; xi, yi, zi i , y i , z i are the coordinates of the point, respectively; the symbol ||·|| represents the modulus of the vector; By obtaining the distance of each point to the straight line L k , it is determined whether each point is a neighborhood point of the straight line L k . N in = { XYZ i | d i ≤ gate} N out = { XYZ i | d i > gate} where N in represents a set containing all points satisfying d i ≤ gate, which is the normal point set; correspondingly, N out represents the abnormal point set; By getting the distance of each point to the line L k , by excluding the distance of the largest and second largest data points, the influence of abnormal points on the fitting line is reduced, and the degree of deviation of the point set from the line distance is obtained: distance =∑d i - d max1 - d max2 where d max1 is the maximum value among all d i ; d max2 is the second maximum value among all d i ; According to the obtained point set N in and distance, determine whether the straight line L k is the current optimal straight line, if it is the optimal straight line, then update the parameters, if it is not the optimal straight line, then jump to step 3-2 for iteration, until all point set combinations are traversed, to obtain the combination points S b1 and S b2 : if |N in |≥|XYZ|*rate and distance <bestd then bestd = distance else continue Wherein, the symbol |·| represents the number of point set; The determined optimal combination point S b1 and S b2 , the abnormal point correction is implemented by using a space-time-based abnormal point correction method, a plane perpendicular to the time axis and corresponding to the abnormal value timestamp is selected as a correction plane to implement a correction point, that is, the abnormal point group N out is corrected in the space-time dimension through its time characteristics, and a corrected correct point set N c : N out = N(x out ,y out ,z out ) z out = z corr N corr = N(x corr ,y corr ,z corr ) N c = N in + N corr wherein N corr is the group of abnormal points N out is the group of points y corr , x corr are the horizontal and vertical coordinates of the corrected points obtained by solving the collinear point relationship in three-dimensional space; By the revised point set N c = N(x c ,y c ,z c ), according to the XY coordinates of N c , the azimuth angle α corresponding to each valid point is obtained, and the motion path angle β is obtained according to the combination points S b1 and S b2 obtained by fitting optimization in step 3-5. a = A c • μ where μ is a sign correction factor used to determine the correct sign of the azimuth angle, A c denotes N c the azimuth angle of each point in the point group; The radial feature reduction module is configured to obtain a motion direction angle θ corresponding to a human motion feature based on a geometric relationship between a motion path and a radar main beam direction, and reduce the Doppler feature D and the distance feature R to a radial feature distribution range D through a feature compensation algorithm. corr and R corr compensate the omnidirectional action feature of the human body to a radial feature.

6. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the steps of the method of any one of claims 1-4.

7. A computer storable medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to realize the steps of the method of any one of claims 1-4.

Images