Human motion multi-dimensional joint representation method based on tensor sparse point cloud

Abstract

The application discloses a human action space-time-frequency multi-dimensional joint representation method based on tensor sparse point cloud, which comprises the following steps: obtaining a radar echo signal, and preprocessing the radar echo signal; constructing a fourth-order tensor model R tensor of the human action echo signal; constructing a tensor sparse dictionary Ψ and a tensor measurement matrix Φ, and combining the fourth-order tensor model R tensor of the human action echo signal to obtain a tensor sparse projection R stensor ; determining a core tensor X based on a core tensor deformation orthogonal matching algorithm; rearranging and rearranging based on the position index to physical information mapping of the core tensor X to generate a tensor sparse human action point cloud representing the space-time-frequency multi-dimensional features of the human action. The scheme optimizes the traditional tensor orthogonal matching algorithm based on core tensor deformation, removes redundant and useless points, improves the representation efficiency of information, effectively reduces the computational complexity, simultaneously takes the sparsity as an index, reduces the dependence of the feature extraction algorithm on the empirical parameters, retains the coupling relationship of each dimension of the signal, and completely extracts the tensor features of the signal from the multi-dimensional and overall angle, and is superior to the low-dimensional feature extraction algorithm in accuracy, efficiency and robustness.

Description

Technical Field

[0001] This invention belongs to the field of radar human motion recognition, specifically involving a multi-dimensional joint representation method of human motion based on tensor sparse point cloud in space, time and frequency. Background Technology

[0002] Human motion recognition has garnered widespread attention for its applications in public surveillance, disaster relief, intelligent interaction, and daily living assistance. Radar's ability to penetrate obstacles unaffected by light and weather conditions has made radar-based human motion recognition technology a popular choice. Compared to optical sensors, radar-based human motion recognition methods perform exceptionally well in diverse environments, overcoming the effects of darkness and obstacles. As the target moves, radar can easily obtain information such as the subject's distance, speed, and angle. Most importantly, radar echo signals only carry the target's feature information and do not capture its visual form, thus protecting the subject's privacy.

[0003] Existing human motion recognition methods based on radar sensors primarily employ image-based feature extraction algorithms. These methods convert radar echo signals into images with feature domains such as distance, Doppler, and angle, and then extract information like depth and contour to achieve feature extraction and representation. However, these image-based feature extraction algorithms suffer from drawbacks such as susceptibility to noise interference and low feature representation efficiency. Therefore, new feature extraction algorithms with strong anti-interference capabilities and high information representation efficiency have become a hot topic in human motion recognition research.

[0004] Sparse representation theory can preserve most of the information in the original signal using a very small number of non-zero sparse solutions, effectively improving the efficiency of information representation while reducing dependence on empirical parameters. However, existing algorithms for representing human motion features based on sparse representation theory are limited to the time-frequency domain or the range-Doppler domain, making it difficult to comprehensively and effectively represent the multidimensional features of human motion in the spatiotemporal domain. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to provide a spatiotemporal-frequency multidimensional joint representation method for human movement based on tensor sparse point clouds. Utilizing the concept of tensor algebra, the sparse representation theory is extended from two-dimensional space to high-dimensional space, proposing a spatiotemporal-frequency multidimensional joint representation method for human movement based on tensor sparse point clouds. First, a high-order tensor model with dimensions such as fast time dimension, slow time dimension, frame time dimension, and channel dimension is established for the echo signal. Then, a corresponding tensor sparse dictionary and measurement matrix are constructed to obtain the tensor sparse projection of the human movement echo signal in the distance-Doppler-time-angle domain. Finally, an orthogonal matching pursuit algorithm based on core tensor deformation is designed to solve for the core tensor, and tensor sparse human activity point clouds (TSHAP) are obtained through mapping relationships, realizing the spatiotemporal-frequency multidimensional joint representation of human movement.

[0006] The specific technical solution for achieving the objective of this invention is as follows:

[0007] A method for spatiotemporal-frequency joint representation of human motion based on tensor sparse point clouds includes the following steps:

[0008] Step 1: Acquire radar echo signals and preprocess them to construct a fourth-order tensor model R of human motion echo signals. tensor ;

[0009] Step 2: Construct a tensor sparse dictionary Ψ and a tensor measurement matrix Φ, and combine them with a fourth-order tensor model R of human motion echo signals. tensor Obtain the tensor sparse projection R stensor ;

[0010] Step 3: Determine the core tensor using an orthogonal matching algorithm based on core tensor deformation. X ;

[0011] Step 4: Based on the core tensor X The mapping from location index to physical information is rearranged to generate a tensor sparse human motion point cloud representing the spatiotemporal-frequency multidimensional features of human motion.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0013] (1) The spatiotemporal-frequency multidimensional joint representation method of human motion based on tensor sparse point cloud proposed in this invention extracts the tensor features of the signal from a multidimensional and holistic perspective by preserving the coupling relationship of each dimension of the signal. It is superior to low-dimensional feature extraction algorithms in terms of accuracy, efficiency and robustness.

[0014] (2) This invention optimizes the traditional tensor orthogonal matching algorithm based on core tensor deformation, removes redundant and useless points, improves the information representation efficiency, effectively reduces computational complexity, and uses sparsity as an indicator to reduce the dependence of feature extraction algorithms on empirical parameters.

[0015] (3) The present invention extracts tensor sparse human motion point cloud based on the sparsity of human motion in the spatiotemporal frequency domain. Compared with traditional spectral features, it has higher representation efficiency, lower data volume, and improves the signal-to-noise ratio of the features, which is beneficial to subsequent network processing.

[0016] The present invention will be further described below with reference to specific embodiments. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the process of the spatiotemporal-frequency multidimensional joint representation method of human motion based on tensor sparse point cloud according to the present invention.

[0018] Figure 2 This is a schematic diagram of echo data preprocessing according to the present invention.

[0019] Figure 3 This is a schematic diagram of a fourth-order tensor model of the human motion echo signal of the present invention.

[0020] Figure 4 This is a comparison of the fourth-order tensor model of human motion signals and the traditional spectrum of the fourth-order reconstructed tensor in this embodiment of the invention.

[0021] Figure 5 This is a schematic diagram of the projection point cloud of the Ming tensor sparse human motion point cloud in the distance, Doppler and time domains in an embodiment of the present invention.

[0022] Figure 6 This is a schematic diagram comparing the extracted tensor sparse human motion point cloud features with traditional spectrograms in an embodiment of the present invention. Detailed Implementation

[0023] A method for spatiotemporal-frequency joint representation of human motion based on tensor sparse point clouds includes the following steps:

[0024] Step 1: Acquire radar echo signals and preprocess them, including data rearrangement, removal of DC components, moving target detection (removal of static clutter interference), and slow time axis segmentation, to construct a fourth-order tensor model R of the human motion echo signal. tensor :

[0025] Step 1-1: Acquire radar echo signal R rec For radar echo signal R rec Perform data rearrangement:

[0026] Each signal transmitted by the radar platform is called a chirp, and the captured radar echo signal R... rec It consists of two LVDS channels in a non-interleaved format and is stored in a binary file; the radar echo signal R is analyzed according to the number of individual chirped sampling points M, the number of chirps N, and the number of antenna channels C. rec The entire echo signal is divided and rearranged into an M×N×C radar echo data block R0, where m=[1,2,...,M] represents fast time sampling points, n=[1,2,...,N] represents slow time series, and c=[1,2,...,C] represents the number of antenna channels;

[0027] Steps 1-2: Based on radar echo data block R0, determine the mean of fast time sampling points in the slow time series for each antenna channel, and obtain radar echo data block R1 after removing DC:

[0028] R1(:,n,c)=R0(:,n,c)-avgDC(n,c)

[0029]

[0030] The symbol “:” indicates that all elements in a specific dimension are indexed.

[0031] Steps 1-3: Subtract radar echo data blocks R1 from each other along the slow time axis to eliminate the echoes of stationary targets with the same amplitude and phase. However, the phase of moving human targets changes continuously and is preserved in the layer-by-layer subtraction structure, thus obtaining the human target echo data block R after removing stationary targets. cube :

[0032]

[0033] Steps 1-4: Echo data block R of the human target cube By segmenting along the slow time axis, the slow time domain is divided into a slow time domain and a frame time domain, forming a fourth-order tensor model R of the human motion echo signal in the fast time-slow time-frame time-channel domain. tensor :

[0034] R tensor =[Frame1,Frame2,…,Frame i ,…Frame F ]

[0035] in Represents the fourth-order human target echo tensor. f = [1,2,...,F] represents a time-domain sequence of frames, and "a:b" indicates indexing from element a to element b in a specific dimension.

[0036] Step 2: Combining tensor algebra and sparse representation theory, construct a tensor sparse dictionary Ψ and a tensor measurement matrix Φ, and combine this with a fourth-order tensor model R of human motion echo signals. tensor Mapped to the measurement space, the tensor sparse projection R is obtained. stensor :

[0037] Step 2-1: Construct a tensor sparse dictionary Ψ and a tensor measurement matrix Φ based on different domains:

[0038] Based on the construction of a tensor sparse dictionary Ψ in different domains, Fourier transform plays a crucial role in radar signal processing. The Doppler frequency shift of the target causes a change in the phase of the echo signal. By obtaining the spectral information of the echo signal through Fourier transform, human target parameters can be estimated. Therefore, based on the four-order human target echo tensor R in the fast-time-slow-time-frame-time-channel domain... tensor Construct the tensor Fourier sparse dictionary Ψ sequentially. The tensor sparse dictionary Ψ is as follows:

[0039] Ψ = {D1, D2, D3, D4}

[0040]

[0041] D3 = I F

[0042]

[0043] in, 0≤i≤M-1,α i Represents the i-th column of the fast-time Fourier dictionary D1; 0≤i≤G-1,β i This represents the i-th column of the slow-time Fourier dictionary D2; 0≤i≤C-1,γ i This represents the atom in the i-th column of the channel domain dimension dictionary D4;

[0044] The frame time dimension is obtained by dividing the frame into slow-time frames, without the need for Fourier transform. D3 represents the frame time dimension dictionary, I F Let F be the identity matrix;

[0045] Tensor measurement matrices Φ are constructed based on different domains. The performance of the measurement matrix plays a crucial role in the accurate reconstruction of the signal. The smaller the correlation coefficient of the sensing matrix, the higher the sparse recovery accuracy and the stronger the system's noise resistance. The tensor measurement matrix Φ is:

[0046] Φ = {SM1, SM2, SM3, SM4}

[0047] SM1=φ M

[0048]

[0049] SM3=φ F

[0050] SM4=φ C

[0051] Where SM1, SM3, and SM4 are random Gaussian measurement matrices in the fast time domain, frame time domain, and channel domain, respectively; SM2 is a circular convolution matrix in the slow time domain, generated by cyclic shifting randomly generated row vectors; φ M φ is an M×M matrix M , φ M Each element in the matrix independently follows a Gaussian distribution with a mean of 0 and a variance of 1 / M, φ F With φ C Similarly;

[0052] Step 2-2: Convert the fourth-order tensor model R of the human motion echo signal tensor Mapping to the measurement space, we obtain the fourth-order tensor observation model R of the human motion echo signal. mtensor :

[0053] R mtensor =R tensor ×1SM1×2SM2×3SM3×4SM4

[0054] in, A fourth-order tensor observation model representing human motion echo signals.

[0055] R, a fourth-order tensor model representing human motion echo signals tensor , × n n=

[0056] 1, 2, 3, 4 represent tensors multiplied by modulo n;

[0057] Steps 2-3: Based on the fourth-order tensor observation model R of human motion echo signals mtensor Tensor sparse projection R of the tensor signal of human motion echo in the distance-Doppler-time-angle domain is obtained through tensor sparseness. stensor :

[0058] Y = X × 1D1 × 2D2 × 3D3 × 4D4

[0059] in, This represents a tensor sparse projection. Represents the core tensor, × n n = 1, 2, 3, 4 represents the tensor modulo n product.

[0060] Rstensor The joint sparse representation can be equivalent to a vector form:

[0061]

[0062] Where y = vec(R) stensor x = vec(X), where the vec operation represents vectorizing the tensor. Represents the Kronecker product;

[0063] According to sparse signal theory, when S << M * G * F * C, the sparse representation vector... The solution can be obtained from the following formula:

[0064]

[0065] in, ||·||0 represents the l0 norm, indicating the number of non-zero elements in vector x. S represents sparsity.

[0066] Step 3: Apply the tensor sparse projection R calculated in Step 2. stensor An orthogonal matching algorithm based on core tensor deformation is proposed to search in the distance-Doppler-time-angle domain, that is, the orthogonal matching algorithm based on core tensor deformation determines the core tensor X:

[0067] Step 3-1: Initialize parameters, input the multidimensional sparse dictionary Ψ, and the fourth-order human target sparse tensor R. stensor Set sparsity S, initialize residual R = R stensor The number of iterations k = 1, and the core tensor X = 0;

[0068] Step 3-2: Determine the coordinate indices [i1, i2, ..., i] of the maximum value of the multidimensional inner product of the residual R and the multidimensional sparse dictionary Ψ. N ]:

[0069]

[0070] Step 3-3: The non-zero elements in a sparse tensor have a certain structure, and their characteristic of being concentrated in a certain region is called block sparsity. The distribution region is called the sparse block. According to the coordinate index [i1,i2,…,i…] N ], merge the index into the index set According to the index set Update sub-dictionary B n :

[0071]

[0072] in The set of indices representing the coordinates of the nth dimension of a tensor. B represents the coordinates of the nth dimension of the maximum value of the multidimensional inner product in the k-th iteration; n The sub-dictionary is composed of the atoms (columns of the sparse dictionary) of the corresponding indexes selected under the nth dimension;

[0073] Steps 3-4: Based on sub-dictionary B n The least squares method is applied to solve the linear equations to determine the redundant core tensor in the sparse block. A The vectorized form a:

[0074]

[0075] y = vec(R) stensor )

[0076] in Let y denote the Kronecker product of two matrices, where y is in R. stensor The vectorized form;

[0077] Steps 3-5: Quantize a sheets into A For redundant core tensors in sparse blocks A Redundant and useless points are filtered out based on energy scale to obtain a set of redundant and useless points:

[0078]

[0079] Where Λ is a set of redundant and useless point coordinates, Coordinates representing redundant and useless points;

[0080] Based on the coordinates of redundant and useless points, the redundant core tensor A By performing deformation to remove the influence of redundant and useless points in the block on the iteration, computational efficiency is improved while sparsity is significantly reduced, making... A (Λ) = 0, for redundant core tensors A Perform deduplication to obtain the core tensor. X ;

[0081] Steps 3-6: Based on the core tensor X Perform iterations to update the residuals. The residual update formula is as follows:

[0082] R =R stensor - X ×1B1×2B2×3B3…× n B n

[0083] Steps 3-7, based on the core tensor X Determine the sparsity s = || X||0, determine if the sparsity s is greater than or equal to the sparsity S. If the sparsity is less than the sparsity S, execute k = k + 1 and return to step 3-2; if the sparsity s is greater than or equal to the sparsity S, then output the core tensor at this time. X With index set

[0084] Then the core tensor X It can be represented as:

[0085]

[0086] in, Represents the non-zero elements in the core tensor;

[0087] Steps 3-8: Based on the core tensor X calculated in Step 3-7 and the sensing matrix H, the core tensor is reconstructed to recover the fourth-order reconstructed tensor of the human motion echo signal.

[0088] The sensing matrix H is the product of the tensor measurement matrix Φ and the corresponding dimensions of the tensor sparse dictionary Ψ.

[0089] H = {H1, H2, H3, H4}

[0090] Where H1 = SM1 × D1, H2 = SM2 × D2, H3 = SM3 × D3, H4 = SM4 × D4, represent the product of the corresponding measurement matrix and the sparse dictionary;

[0091] Core Tensor X By performing a modular n-product with the sensing matrices of each dimension, the fourth-order reconstruction tensor of the human motion echo signal is calculated.

[0092]

[0093] Step 4: Based on the core tensor X The mapping from location index to physical information is rearranged to generate a tensor sparse human motion point cloud (TSHAP) representing the spatiotemporal-frequency multidimensional features of human motion.

[0094] Step 4-1, Non-zero elements In the core tensor, the position index represents distance, Doppler, time, and angle information; non-zero elements... The absolute value represents the power intensity, relative to the core tensor. X Central African zero element Sort in descending order and record the corresponding distance index ind_r k Doppler index ind_d k Time index ind_t k and angle index ind_a kBased on the physical information represented by each atom in the multidimensional sparse dictionary Ψ, non-zero elements... Map the index to the distance index ind_r k Doppler index ind_d k Time index ind_t k and angle index ind_a k Mapped to distance information RNG k Doppler information (dop) k Time information t k and angle information ang k :

[0095] rng k =R Max *ind_r k / M

[0096]

[0097] t k =ind_t k *t f

[0098] ang k =A Max *ind_a k / C

[0099] Among them, R Max M represents the maximum ranging range of the radar, and M is the length of the fast time dimension Fourier dictionary D1. Max G represents the maximum frequency measurement range of the radar, G is the length of the slow-time dimension Fourier dictionary D2, and t f Indicates the duration of each frame, A Max This represents the maximum angle measurement range of the radar, and C is the length of the channel dimension Fourier dictionary D2;

[0100] Step 4-2: Based on the calculation results of Step 4-1, convert the non-zero elements... Rearranged into a tensor sparse human motion point cloud (TSHAP) representing the distance-Doppler-time-angle domain:

[0101]

[0102] in, Represents the core tensor X Location at (rng) k ,dop k ,t k ,ang k Non-zero elements at position ) The power intensity, s = 1, 2, ..., S represents the index of the non-zero element item.

[0103] A spatiotemporal-frequency multidimensional joint representation system for human motion based on tensor sparse point clouds includes the following modules:

[0104] The fourth-order tensor model construction module is used to acquire radar echo signals, preprocess the radar echo signals, and construct a fourth-order tensor model R of the human motion echo signals. tensor ;

[0105] Tensor Sparse Projection Module: Used to construct the tensor sparse dictionary Ψ and the tensor measurement matrix Φ, and combine it with the fourth-order tensor model R of human motion echo signals. tensor Obtain the tensor sparse projection R stensor ;

[0106] Core Tensor Module: Used to determine the core tensor using an orthogonal matching algorithm based on core tensor deformation. X ;

[0107] Tensor-Sparse Human Motion Point Cloud Module: Used for core tensor-based... X The mapping from location index to physical information is rearranged to generate a tensor sparse human motion point cloud representing the spatiotemporal-frequency multidimensional features of human motion.

[0108] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, performs the following steps:

[0109] Step 1: Acquire radar echo signals and preprocess them to construct a fourth-order tensor model R of human motion echo signals. tensor ;

[0110] Step 2: Construct a tensor sparse dictionary Ψ and a tensor measurement matrix Φ, and combine them with a fourth-order tensor model R of human motion echo signals. tensor Obtain the tensor sparse projection R stensor ;

[0111] Step 3: Determine the core tensor using an orthogonal matching algorithm based on core tensor deformation. X ;

[0112] Step 4: Based on the core tensor X The mapping from location index to physical information is rearranged to generate a tensor sparse human motion point cloud representing the spatiotemporal-frequency multidimensional features of human motion.

[0113] A computer-storable medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0114] Step 1: Acquire radar echo signals and preprocess them to construct a fourth-order tensor model R of human motion echo signals. tensor ;

[0115] Step 2: Construct a tensor sparse dictionary Ψ and a tensor measurement matrix Φ, and combine them with a fourth-order tensor model R of human motion echo signals. tensor Obtain the tensor sparse projection R stensor ;

[0116] Step 3: Determine the core tensor X using an orthogonal matching algorithm based on core tensor deformation;

[0117] Step 4: Based on the core tensor X The mapping from location index to physical information is rearranged to generate a tensor sparse human motion point cloud representing the spatiotemporal-frequency multidimensional features of human motion.

[0118] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0119] Example

[0120] Combination Figure 1 A spatiotemporal-frequency joint representation method for human motion based on tensor sparse point clouds includes the following steps:

[0121] Step 1: Elevate the radar system to 0.85 meters above the ground. Have a volunteer fall towards the radar to obtain radar echo signals from the platform. Preprocess the radar echo signals, including data rearrangement, removal of DC components, moving target detection (removal of static clutter interference), and slow time axis segmentation, to obtain a fourth-order tensor model R of the human motion echo signal with sparse characteristics and after removing stationary targets. tensor ,like Figure 2 As shown:

[0122] Step 1-1: Acquire radar echo signal R rec For radar echo signal R rec Perform data rearrangement:

[0123] Each signal transmitted by the radar platform is called a chirp, and the captured radar echo signal R rec It consists of two LVDS channels in a non-interleaved format and is stored in a binary file; according to the number of sampling points per chirp (64), the number of chirps (32000), and the number of antenna channels (4), the radar echo signal R is analyzed. recThe entire echo signal is divided and rearranged into a 64×32000×4 radar echo data block R0, where m=[1,2,...,64] represents fast time sampling points, n=[1,2,...,32000] represents slow time sequence, and c=[1,2,...,4] represents the number of antenna channels;

[0124] Steps 1-2: Based on radar echo data block R0, determine the mean of fast time sampling points in the slow time series for each antenna channel, and obtain radar echo data block R1 after removing DC:

[0125] R1(:,n,c)=R0(:,n,c)-avgDC(n,c)

[0126]

[0127] Where n = [1,2,...,32000] corresponds to different slow times, c = [1,2,...,4] corresponds to different numbers of antenna channels, and the symbol ":" indicates indexing all elements in a specific dimension;

[0128] Steps 1-3: Subtract radar echo data blocks R1 from each other along the slow time axis to eliminate the echoes of stationary targets with the same amplitude and phase. However, the phase of moving human targets changes continuously and is preserved in the layer-by-layer subtraction structure, thus obtaining the human target echo data block R after removing stationary targets. cube :

[0129]

[0130] Where m = [1,2,...,64] represents fast time sampling points, n = [1,...,32000] represents slow time series, and c = [1,2,...,4] represents the number of antenna channels;

[0131] Steps 1-4: Echo data block R of the human target cube Segmenting along the slow time axis, the slow time domain is divided into a slow time domain and a frame time domain, forming a fourth-order tensor model R of the human motion echo signal in the fast time-slow time-frame time-channel domain. tensor :

[0132] R tensor =[Frame1,Frame2,…,Frame i ,…Frame F ]

[0133] in Represents the fourth-order human target echo tensor. f = [1,2,...,250] represents a time-domain sequence of frames, and "a:b" indicates indexing from element a to element b in a specific dimension.

[0134] Step 2: Combining tensor algebra and sparse representation theory, construct a tensor sparse dictionary Ψ and a tensor measurement matrix Φ, and combine this with a fourth-order tensor model R of human motion echo signals. tensor Mapped to the measurement space, the tensor signal of the human motion echo is obtained as a tensor sparse projection R in the distance-Doppler-time-angle domain. stensor ,like Figure 3 As shown:

[0135] Step 2-1: Construct a tensor sparse dictionary Ψ and a tensor measurement matrix Φ based on different domains:

[0136] Based on the construction of a tensor sparse dictionary Ψ in different domains, Fourier transform plays a crucial role in radar signal processing. The Doppler frequency shift of the target causes a change in the phase of the echo signal. By obtaining the spectral information of the echo signal through Fourier transform, human target parameters can be estimated. Therefore, based on the four-order human target echo tensor R in the fast-time-slow-time-frame-time-channel domain... tensor Construct a tensor Fourier sparse dictionary Ψ sequentially;

[0137] D1 is the fast-time Fourier dictionary matrix, with a size of 64×64, which can be represented as:

[0138]

[0139] in, 0≤i≤63,α i This represents the atom in the i-th column of dictionary D1;

[0140] Similarly, D2 and D4 are the Fourier dictionary matrices in the slow time and channel domain dimensions, respectively, with sizes of 128×128 and 4×4, which can be represented as:

[0141]

[0142] in, 0≤i≤127, β i Represents the atom in the i-th column of dictionary D2;

[0143] 0≤i≤3, γ i This represents the atom in the i-th column of dictionary D4.

[0144] The frame time dimension is obtained by dividing the frame into slow-time frames, without the need for Fourier transform. D3 is an identity matrix of length 250 in the frame time dimension: D3 = I F .

[0145] The final tensor sparse dictionary can be represented as:

[0146] Ψ = {D1, D2, D3, D4}

[0147] Tensor measurement matrices Φ are constructed based on different domains. The performance of the measurement matrix plays a key role in the accurate reconstruction of the signal. The smaller the correlation coefficient of the sensing matrix, the higher the sparse recovery accuracy and the stronger the system's noise resistance.

[0148] SM1, SM3, and SM4 are the random Gaussian measurement matrices in the fast time domain, frame time domain, and channel domain, respectively, as shown below:

[0149] SM1=φ 64

[0150] SM3=φ 250

[0151] SM4 = φ4

[0152] Where, φ 64 A 64×64 matrix, φ 64 Each element in the matrix independently follows a Gaussian distribution with a mean of 0 and a variance of 1 / 64, φ 250 The same applies to φ4.

[0153] SM2 selects a circular convolution matrix in the slow time domain and generates the matrix by cyclic shifting randomly generated row vectors, which can be represented as:

[0154]

[0155] The final tensor measurement matrix Φ can be expressed as:

[0156] Φ = {SM1, SM2, SM3, SM4}

[0157] Step 2-2: Convert the fourth-order tensor model R of the human motion echo signal tensor Mapping to the measurement space, we obtain the fourth-order tensor observation model R of the human motion echo signal. mtensor :

[0158] R mtensor =R tensor ×1SM1×2SM2×3SM3×4SM4

[0159] in, A fourth-order tensor observation model representing human motion echo signals. R, a fourth-order tensor model representing human motion echo signals tensor ,

[0160] × nn = 1, 2, 3, 4 represents the tensor modulo n product;

[0161] Steps 2-3: Based on the fourth-order tensor observation model R of human motion echo signals mtensor Tensor sparse projection R of the tensor signal of human motion echo in the distance-Doppler-time-angle domain is obtained through tensor sparseness. stensor :

[0162] Y = X ×1D1×2D2×3D3×4D4

[0163] in, This represents a tensor sparse projection. Represents the core tensor, × n n = 1, 2, 3, 4 represents the tensor modulo n product.

[0164] R stensor The joint sparse representation can be equivalent to a vector form:

[0165]

[0166] Where y = vec(R) stensor x = vec(X), where the vec operation represents vectorizing the tensor. Represents the Kronecker product;

[0167] According to sparse signal theory, when S << M * G * F * C, the sparse representation vector... The solution can be obtained from the following formula:

[0168]

[0169] in, ||·‖0 represents the l0 norm, indicating the number of non-zero elements in vector x. S represents sparsity.

[0170] Step 3: Apply the tensor sparse projection R calculated in Step 2. stensor This paper proposes an orthogonal matching algorithm based on core tensor deformation to search in the distance-Doppler-time-angle domain, that is, to determine the core tensor based on the orthogonal matching algorithm based on core tensor deformation. X And recover the fourth-order reconstructed tensor of the human motion echo signal; The fourth-order tensor model R of human motion echo signals tensor Comparative verification was performed using traditional distance-time plots, time-frequency plots, distance-Doppler plots, and distance-angle plots, such as... Figure 4 As shown;

[0171] Step 3-1: Initialize parameters. For the fall samples, the sparsity S = 1000. Input the tensor sparse dictionary Ψ and the fourth-order human target sparse tensor R. stensor Sparsity S = 1000. Initialize residuals. R =R stensor The number of iterations k=1, and the core tensor X =0;

[0172] Step 3-2: Determine the residuals R The coordinate indices [i1, i2, ..., i] of the maximum value of the multidimensional inner product with the multidimensional sparse dictionary Ψ. N ]:

[0173]

[0174] Step 3-3: The non-zero elements in a sparse tensor have a certain structure, and their characteristic of being concentrated in a certain region is called block sparsity. The distribution region is called the sparse block. According to the coordinate index [i1,i2,…,i…] N ], merge the index into the index set According to the index set Update sub-dictionary B n :

[0175]

[0176] in The set of indices representing the coordinates of the nth dimension of a tensor. B represents the coordinates of the nth dimension of the maximum value of the multidimensional inner product in the k-th iteration; n Size is a sub-dictionary composed of the atoms (columns of the sparse dictionary) of the corresponding indexes selected under the nth dimension. b Indicates the size of blocks in a sparse block;

[0177] Steps 3-4: Based on sub-dictionary B n The least squares method is applied to solve the linear equations to determine the redundant core tensor in the sparse block. A The vectorized form a:

[0178]

[0179] y = vec(R) stensor )

[0180] in Let y denote the Kronecker product of two matrices, where y is in R. stensor The vectorized form;

[0181] Steps 3-5: Quantize a sheets into A For redundant core tensors in sparse blocks ARedundant and useless points are filtered out based on energy scale to obtain a set of redundant and useless points:

[0182]

[0183] Where Λ is a set of redundant and useless point coordinates, Coordinates representing redundant and useless points;

[0184] Based on the coordinates of redundant and useless points, the redundant core tensor A By performing deformation to remove the influence of redundant and useless points in the block on the iteration, computational efficiency is improved while sparsity is significantly reduced, making... A (Λ) = 0, for redundant core tensors A Perform deduplication to obtain the core tensor. X ;

[0185] Steps 3-6: Based on the core tensor X Perform iterations to update the residuals. The residual update formula is as follows:

[0186] R =R stensor -X×1B1×2B2×3B3…× n B n

[0187] Steps 3-7, based on the core tensor X Determine the sparsity s = || X ||0, determine if the sparsity s is greater than or equal to the sparsity S. If the sparsity is less than the sparsity S, execute k = k + 1 and return to step 3-2; if the sparsity s is greater than or equal to the sparsity S, then output the core tensor at this time. X With index set

[0188] Then the core tensor X It can be represented as:

[0189]

[0190] in, Represents the non-zero elements in the core tensor;

[0191] Steps 3-8: Based on the core tensor X calculated in Step 3-7 and the sensing matrix H, the core tensor is reconstructed to recover the fourth-order reconstructed tensor of the human motion echo signal.

[0192] The sensing matrix H is the product of the tensor measurement matrix Φ and the corresponding dimensions of the tensor sparse dictionary Ψ.

[0193] H = {H1, H2, H3, H4}

[0194] Where H1 = SM1 × D1, H2 = SM2 × D2, H3 = SM3 × D3, H4 = SM4 × D4, represent the product of the corresponding measurement matrix and the sparse dictionary;

[0195] The core tensor X is multiplied by the sensing matrices of each dimension modulo n to calculate the fourth-order reconstruction tensor of the human motion echo signal.

[0196]

[0197] Step 4: Based on the core tensor X, the mapping from position index to physical information is rearranged to generate a tensor sparse human motion point cloud TSHAP representing the spatiotemporal-frequency multidimensional features of human motion. The TSHAP is then dimensionality-reduced and projected in the distance-time-Doppler domain as (TSHAP-RangeTimeDoppler)TSHAP-RTD, as shown below. Figure 5 As shown, the projections of TSHAP in the time-range domain (TSHAP-RangeTime) TSHAP-RT, the projections in the time-Doppler domain (TSHAP-DopplerTime) TSHAP-DT, and the projections in the range-Doppler domain (TSHAP-RangeDoppler) TSHAP-RD are compared with the fourth-order human target echo tensor R. tensor Comparative verification was performed using traditional distance-time plots, time-frequency plots, and distance-Doppler plots, such as... Figure 6 As shown:

[0198] Step 4-1, Non-zero elements In the core tensor, the position index represents distance, Doppler, time, and angle information; non-zero elements... The absolute value represents the power intensity, relative to the core tensor. X Central African zero element Sort in descending order and record the corresponding distance index ind_r k Doppler index ind_d k Time index ind_t k and angle index ind_a k Based on the physical information represented by each atom in the multidimensional sparse dictionary Ψ, non-zero elements... Map the index to the distance index ind_r k Doppler index ind_d k Time index ind_t k and angle index ind_a k Mapped to distance information RNG k Doppler information (dop) k Time information t k and angle information ang k :

[0199] rng k =R Max *ind_r k / M

[0200]

[0201] t k =ind_t k *t f

[0202] ang k =A Max *ind_a k / C

[0203] Among them, R Max M represents the maximum ranging range of the radar, and M is the length of the fast time dimension Fourier dictionary D1. Max G represents the maximum frequency measurement range of the radar, G is the length of the slow-time dimension Fourier dictionary D2, and t f Indicates the duration of each frame, A Ma This represents the maximum angle measurement range of the radar, and C is the length of the channel dimension Fourier dictionary D2;

[0204] Step 4-2: Based on the calculation results of Step 4-1, convert the non-zero elements... Rearranged into a tensor sparse human motion point cloud (TSHAP) representing the distance-Doppler-time-angle domain:

[0205]

[0206] in, Represents the core tensor X Location at (rng) k ,dop k ,t k ,ang k Non-zero elements at position ) The power intensity, s = 1, 2, ..., S represents the index of the non-zero element item.

[0207] Combination Figure 4 and Figure 6 It can be seen that after the orthogonal matching algorithm based on core tensor deformation proposed in this invention, the core tensor... X The fourth-order reconstruction tensor of the recovered human motion echo signal The fourth-order tensor model r of human motion echo signals tensor It shows high agreement with traditional spectral features: distance-time plot, time-frequency plot, distance-Doppler plot, and distance-angle plot, and the fourth-order reconstruction tensor... Noise is suppressed, therefore the fourth-order reconstruction tensor To achieve accurate reconstruction, the core tensor contains a fourth-order tensor model r. tensor Feature information; at the same time by Figure 6 It was observed that the dimensionality reduction projection of the tensor sparse human motion point cloud TSHAP and the fourth-order tensor model R... tensor The consistent performance of traditional spectrogram features demonstrates that TSHAP can extract human motion features with lower data volume and higher representation efficiency through sparse point cloud. Furthermore, unlike single spectrograms, TSHAP's tensor properties preserve the coupling relationship of features in various dimensions, allowing for the complete extraction of tensor features of the signal from a multi-dimensional and holistic perspective.

[0208] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A multi-dimensional joint representation method for human motion based on tensor sparse point clouds, characterized in that, Includes the following steps: Step 1: Acquire radar echo signals and preprocess them to construct a fourth-order tensor model of human motion echo signals. : Step 1-1: Acquire radar echo signal For radar echo signals Perform data rearrangement: Based on the number of individual chirp sampling points Number of chirps Number of antenna channels radar echo signal Divide the signal and rearrange the entire echo signal into Radar echo data block ,in Indicates fast time sampling points, Indicates a slow time series. Indicates the number of antenna channels; Steps 1-2: Based on radar echo data blocks Determine the mean of fast time sampling points in the slow time series for each antenna channel, and obtain the radar echo data block after removing DC. : ； ； The symbol ":" indicates that all elements in a specific dimension are indexed. Steps 1-3: Transfer radar echo data blocks Subtracting each other along the slow time axis eliminates stationary target echoes with the same amplitude and phase, thus obtaining human target echo data blocks after removing stationary targets. : ； Steps 1-4: Echo data blocks of human target By segmenting along the slow time axis, the slow time domain is divided into a slow time domain and a frame time domain, forming a fourth-order tensor model of the human motion echo signal in the fast time-slow time-frame time-channel domain. : ； in Represents the fourth-order human target echo tensor. , , Represents a frame time-domain sequence, " indicates indexing from element a to element b in a specific dimension; Step 2: Construct a tensor sparse dictionary and tensor measurement matrix And combined with the fourth-order tensor model of human motion echo signals. Obtain tensor sparse projection ; Step 3: Determine the core tensor using an orthogonal matching algorithm based on core tensor deformation. : Step 3-1: Initialize parameters and input a multidimensional sparse dictionary. Fourth-order human target sparse tensor Set sparsity Initialize residuals Number of iterations Core tensor ; Step 3-2: Determine the residuals With multidimensional sparse dictionaries Coordinate index of the maximum value of the multidimensional inner product [ ]: ； Step 3-3, based on coordinate index [ ], merge the index into the index set According to the index set Update sub-dictionary : ； ； in The set of indices representing the coordinates of the nth dimension of a tensor. The coordinate of the nth dimension represents the maximum value of the multidimensional inner product in the k-th iteration; This is a sub-dictionary composed of atoms corresponding to the indexes selected under the nth dimension; Steps 3-4: Based on sub-dictionaries Determine the redundant core tensor in the sparse block. vectorized form : ； ； in Denotes the Kronecker product of two matrices. for The vectorized form; Steps 3-5, Zhang Liangliang For redundant core tensors in sparse blocks Redundant and useless points are filtered out based on energy scale to obtain a set of redundant and useless points: ； in It is a set of redundant and useless point coordinates. Coordinates representing redundant and useless points; Based on the coordinates of redundant and useless points, the redundant core tensor To deform, so that =0, for redundant core tensors Perform deduplication to obtain the core tensor. ; Steps 3-6: Based on the core tensor Perform iterations to update the residuals. The residual update formula is as follows: ； Steps 3-7, based on the core tensor Determine sparsity Determine sparsity Is it greater than or equal to sparsity? If the sparsity is less than the sparsity ,implement Return to step 3-2; if sparsity greater than or equal to sparsity Then, output the core tensor at this point. With index set ; Then the core tensor It can be represented as: ； in, Represents the non-zero elements in the core tensor; Step 3-8: Based on the core tensor calculated in step 3-7 and sensor matrix The core tensor is reconstructed to recover a fourth-order reconstructed tensor of the human motion echo signal. ; Sensing Matrix Tensor measurement matrix tensor sparse dictionary Product of corresponding dimensions: ； in , , , , representing the product of the corresponding measurement matrix and the sparse dictionary; Core Tensor By performing a modular n-product with the sensing matrices of each dimension, the fourth-order reconstruction tensor of the human motion echo signal is calculated. : ； Step 4: Based on the core tensor The mapping from location index to physical information is rearranged to generate a tensor sparse human motion point cloud representing the spatiotemporal-frequency multidimensional features of human motion.

2. The method for multidimensional joint representation of human motion based on tensor sparse point clouds according to claim 1, characterized in that, Step 2: Obtaining the tensor sparse projection Specifically: Step 2-1: Construct tensor sparse dictionaries based on different domains and tensor measurement matrix : Tensor Sparse Dictionary for: ； ； ； ； ； in, , , Fourier Dictionary for Expressing Fast Time The i-th column of atoms; , , Fourier Dictionary representing slow time The i-th column of atoms; , Dictionary representing channel domain dimensions The i-th column of atoms, A dictionary representing the time dimension of a frame. Let F be the identity matrix; Tensor measurement matrix for: ； ； ； ； ； in, , and These are random Gaussian measurement matrices in the fast time domain, frame time domain, and channel domain, respectively. For slow-time domain circular convolution matrices, for Matrix of size , Each element in the dataset independently follows a Gaussian distribution with a mean of 0 and a variance of 1 / M. and Similarly; Step 2-2: Develop a fourth-order tensor model of the human motion echo signal. Mapping to the measurement space yields a fourth-order tensor observation model of the human motion echo signal. : ； in, A fourth-order tensor observation model representing human motion echo signals. A fourth-order tensor model representing human motion echo signals , Represented as the product of tensors modulo n; Steps 2-3: Based on the fourth-order tensor observation model of human motion echo signals Tensor sparse projection of the tensor signal of human motion echo in the distance-Doppler-time-angle domain is obtained through tensor sparse projection. : ； in, This represents a tensor sparse projection. Represents the core tensor, It is represented as the product of tensors modulo n.

3. The method for multidimensional joint representation of human motion based on tensor sparse point clouds according to claim 1, characterized in that, The step 4 of generating a tensor sparse human motion point cloud representing the spatiotemporal-frequency multidimensional features of human motion specifically involves: Step 4-1, for the core tensor Central African zero element Sort in descending order and record the corresponding distance index. Doppler Index Time Index and angle index According to the multidimensional sparse dictionary The physical information represented by each atom in the non-zero element Map the index to the distance index. Doppler Index Time Index and angle index Mapped to distance information Doppler information Time information and angle information : ； ； ； ； in, Indicates the maximum ranging range of the radar. Fourier Dictionary for Fast Time Dimension Length, Indicates the maximum frequency measurement range of the radar. Fourier dictionary for slow time dimension Length, Indicates the duration of each frame. Indicates the maximum angle measurement range of the radar. Fourier dictionary for channel dimensions Length; Step 4-2: Based on the calculation results of Step 4-1, convert the non-zero elements... Rearranged into a tensor sparse human motion point cloud representing the distance-Doppler-time-angle domain. : ； in, Represents the core tensor Location at ( Non-zero elements at the location power intensity, Represents the index of a non-zero element.

4. A multidimensional joint representation system for human motion based on tensor sparse point clouds, used to execute the method described in claim 1, characterized in that, Includes the following modules: The fourth-order tensor model construction module is used to acquire radar echo signals, preprocess the radar echo signals, and construct a fourth-order tensor model of human motion echo signals. ; Tensor Sparse Projection Module: Used to construct tensor sparse dictionaries and tensor measurement matrix And combined with the fourth-order tensor model of human motion echo signals. Obtain tensor sparse projection ; Core Tensor Module: Used to determine the core tensor using an orthogonal matching algorithm based on core tensor deformation. ; Tensor-Sparse Human Motion Point Cloud Module: Used for core tensor-based... The mapping from location index to physical information is rearranged to generate a tensor sparse human motion point cloud representing the spatiotemporal-frequency multidimensional features of human motion.

5. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-3.

6. A computer-storable medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-3.

Images