A method and device for identifying moving targets of a through-wall radar

Abstract

The present invention discloses a method and device for identifying moving targets by through-wall radar. The method includes using a plurality of through-wall radar imaging data as training data, performing singular value decomposition on the training data, and enhancing both the training data and the sub-components obtained after decomposing the training data by a wavelet scattering network to obtain an enhanced imaging matrix and enhanced sub-components. Inputting the enhanced imaging matrix and the enhanced sub-components into their respective probability graph models, where the probability graph model is a conditional random field (CRF), training each of the probability graph models to obtain the trained probability graph models. Based on the outputs of the trained probability graph models, performing model fusion. Obtaining the through-wall radar imaging data to be identified and identifying the through-wall radar imaging data to be identified based on the fused model.

Description

Technical Field

[0001] The present invention relates to the field of radar signal processing, and particularly to a method and device for identifying moving targets of a through-wall radar. Background Art

[0002] Due to the advantages of the ultra-wideband (UWB) signal such as certain medium penetration ability and better resolution for close-range targets, the through-wall radar (TWR) technology has developed rapidly in the past few decades. It can serve various fields such as military radars and civilian radars better. In these application fields, the identification of moving targets (especially human bodies) is one of the most challenging topics in through-wall moving target identification. Since the human target exhibits the characteristics of an extended target in through-wall imaging, and problems such as low signal-to-clutter-plus-noise ratio (SCNR) and feature overlap ambiguity, and multipath effects caused by microwave penetration through the wall and focusing need to be solved.

[0003] For the work of human behavior recognition of ultra-wideband through-wall radars, the existing research results at home and abroad can be divided into three categories. The first category is methods based on traditional physical modeling and statistical signal processing theories, including singular value decomposition (SVD), truncated singular value decomposition (TSVD), through-wall direction-of-arrival estimation algorithm (DDOA), and improved Kalman filtering method. These methods are based on physical models, with low computational complexity and the strongest interpretability, but poor generalization ability and are not applicable to complex scenarios. The second category is methods based on compressive sensing, sparse and low-rank modeling theories, and statistical learning, including robust principal component analysis (RPCA) [8] based on reconstructing micro-Doppler features, support vector machine (SVM), and likelihood ratio (LRT) detector, time delay estimation method (TDOE) based on the orthogonal matching pursuit (OMP) algorithm, and cross-validation-based synchronous orthogonal matching pursuit algorithm (CV-CSOMP). These methods have better recognition effects than the first category of traditional methods, but their generalization ability and inference performance are still limited when processing imaging data of complex motion state scenarios. The third category is methods based on the new generation of artificial intelligence theories, including autoencoder network (AEN), convolutional neural network (CNN), generative adversarial network (GAN), fully connected multi-layer perceptron (MLP), and neural network based on simple probability graph (PGM). Artificial intelligence algorithms have strong scene generalization and adaptation abilities, and can work without being affected by many prior knowledge, and are considered to be one of the best choices that may replace the original moving target identification method of through-wall radar in the research in the past five years, but their training process requires a large amount of computational cost.

[0004] The prior art CN202010005940.5 (filing date: January 3, 2020, publication date: May 15, 2020) discloses a method for estimating the micro-Doppler frequency of a human target based on an extended Bessel model, belonging to the technical field of micro-Doppler frequency estimation. By extracting the micro-Doppler frequency of the calf and then performing piecewise fitting, first using the Bessel model to determine the optimal control points, introducing parameters without changing and without adding control points, and completing the separation of the target component of the radar echo signal and the accurate estimation of the calf Doppler frequency characteristics through the Hough transform frequency estimation algorithm based on the extended Bessel model, it has good application prospects in real-time human sensing applications. This prior art solves the task of realizing micro-Doppler feature extraction and behavior recognition by using an extended human body modeling method. Its advantage lies in being able to sort out the micro-Doppler characteristics of human behavior with as few model parameters as possible and as fast an estimation speed as possible, and then conduct comparison, parameter inversion, and perception to achieve the function of accurately estimating the behavior state of the human target. However, due to the non-self-adaptive parameter adjustment characteristic of its model, the intelligence of the algorithm is limited within a relatively small range.

[0005] The prior art CN201810592255.X (filing date: June 11, 2018, publication date: November 9, 2018) discloses an adaptive through-wall radar stationary human target positioning method based on EMD, which relates to through-wall radar technology, especially the detection and positioning technology of stationary targets of through-wall radar. This prior art provides a stationary target positioning algorithm applicable to MIMO through-wall radar. Perform fast-time dimension Fourier transform on the received multi-period data to obtain multi-period range images; then decompose the multi-period range images into multiple column vectors according to range cells, and process the data of each column vector with the EMD algorithm; convert the processed range images to the frequency domain through slow-time dimension Fourier transform, and adaptively select the range images containing target information. Finally, perform fast imaging through the BP imaging algorithm and further suppress noise and clutter in combination with the PCF algorithm. The method proposed by this prior art has the ability of adaptive data processing. However, this prior art decomposes the multi-period range images into multiple column vectors according to range cells and uses the EMD algorithm to process the data of each column vector. However, in the through-wall B-scan data, there are not only column vector relationships but also row vector relationships. Therefore, this prior art loses some feature information and cannot achieve the optimal performance. Summary of the Invention

[0006] In view of this, the present invention provides a method and device for identifying moving targets of a through-wall radar, which can solve the technical problems that the introduction of a wall makes the imaging data of an ultra-wideband (UWB) through-wall radar (TWR) have characteristics such as discontinuous features and low signal-to-clutter-noise ratio (SCNR), resulting in a decrease in the accuracy of the behavior recognition algorithm for human moving targets behind the wall and the robustness of the system.

[0007] To solve the above technical problems, the present invention is implemented as follows.

[0008] A method for identifying moving targets of a through-wall radar, comprising:

[0009] Step S1: Using a plurality of through-wall radar imaging data as training data, performing singular value decomposition on the training data, and enhancing both the training data and the sub-components obtained after decomposing the training data by a wavelet scattering network to obtain an enhanced imaging matrix and enhanced sub-components;

[0010] Step S2: Inputting the enhanced imaging matrix and the enhanced sub-components into their respective corresponding probabilistic graphical models, where the probabilistic graphical model is a conditional random field (CRF), training each of the probabilistic graphical models to obtain the trained probabilistic graphical models; based on the outputs of the trained probabilistic graphical models, performing model fusion;

[0011] Step S3: Obtaining the through-wall radar imaging data to be identified, and identifying the through-wall radar imaging data to be identified based on the fused model.

[0012] Preferably, in step S1: the singular value decomposition of the training data is performed in the following manner:

[0013]

[0014] where, φ r is the echo matrix corresponding to the through-wall radar imaging data, with a dimension of N×M, σ i represents the singular value after the SVD decomposition of the echo matrix, u i and v i respectively represent the left and right singular vectors after the SVD decomposition of the echo matrix, H represents the Hermitian transform of the echo matrix, i is the singular value ordinal number arranged in descending order, W is the wall clutter subspace, T is the target subspace, N is the noise subspace; the output matrix represents three subspaces containing only human target imaging information, wall clutter information, and noise information.

[0015] Preferably, the wavelet function of the wavelet scattering network is:

[0016]

[0017] where, WT(a,τ) is the representation of the input data in the wavelet domain, and the input data is the echo signal; f(t) is a measure of the time-domain representation of the input echo signal, ψ represents the wavelet transform basis function, t is the integration time variable, a is the dilation scale of the wavelet transform, and τ is the time delay in the wavelet domain.

[0018] Preferably, the convolutional basis function of the wavelet transform of the wavelet scattering network is expressed as

[0019] ψ j (x) = 2 -2j ψ(2 -j x)(3)

[0020] where ψ j (x) is the output of the j-th layer of the wavelet scattering network, x represents the input of the nodes in the wavelet scattering network, and θ is defined as the rotation angle of the wavelet transform in the wavelet scattering network. Then

[0021]

[0022] The convolutional basis function corresponding to each layer in the wavelet scattering network is expressed as

[0023]

[0024] where ψ l,θ,1 (x) is the first convolutional basis function of the l-th layer of the wavelet scattering network, l is the number of layers of the wavelet scattering network, ψ i,θ,2 (x) is the second convolutional basis function of the l-th layer of the wavelet scattering network, r θ is the rotation angle between each layer of the wavelet scattering network, is the transfer function of the wavelet scattering network, and k = log2a represents the logarithm of the scale transformation coefficient a in the wavelet transform.

[0025] Preferably, the probability of the conditional random field model is expressed as

[0026]

[0027] where p(y|x) represents the conditional probability of output y under the condition of input x, w s represents the weight of the s-th node in the order of the conditional random field, S is the total number of nodes, exp represents the exponential function, and f s (y, x) is the conditional random field feature function corresponding to the conditional variable x and the state variable y, and z(x) is the normalization constant.

[0028] The probability graph model is the conditional random field (CRF). All kinds of subclasses of the probability graph model are applicable to constructing the probability graph model.

[0029] Preferably, based on the outputs of the trained conditional random field models, model fusion is performed, and the fused model is BIC(MLF):

[0030]

[0031] where p represents the input data under the hyperparameters The maximum likelihood probability under represents that the joint likelihood estimation of the MLF makes the estimation result consistent with the feature distribution of the data Train after radar echo preprocessing in terms of probability. is a hyperparameter that can be dynamically adjusted. MLF is the marginal likelihood function. is the regularization term, M is the size of the training samples, and Dim() represents the number of independent parameters.

[0032] A device for identifying moving targets of a through-wall radar provided by the present invention, the device includes:

[0033] Decomposition module: configured to take a plurality of through-wall radar imaging data as training data, perform singular value decomposition on the training data, and perform enhancement processing on the training data and the sub-components after decomposition of the training data by a wavelet scattering network to obtain an enhanced imaging matrix and enhanced sub-components;

[0034] Fusion module: configured to input the enhanced imaging matrix and enhanced sub-components into their respective corresponding probabilistic graphical models, where the probabilistic graphical model is a conditional random field (CRF), train each of the probabilistic graphical models to obtain each trained probabilistic graphical model; based on the outputs of each trained probabilistic graphical model, perform model fusion;

[0035] Recognition module: configured to obtain the through-wall radar imaging data to be recognized and recognize the through-wall radar imaging data to be recognized based on the fused model.

[0036] A system, the system includes:

[0037] A processor for executing multiple instructions;

[0038] A memory for storing multiple instructions;

[0039] Wherein, the multiple instructions are used to be stored by the memory and loaded and executed by the processor to perform the method as described above.

[0040] A computer-readable storage medium, in which multiple instructions are stored; the multiple instructions are used to be loaded and executed by a processor to perform the method as described above.

[0041] The present invention proposes a method and device for identifying moving targets of a through-wall radar, which is used for identifying human behaviors of a through-wall radar. The method first collects imaging data of seven different states and establishes a data set, and then applies the subspace separation method based on SVD to the data. Except for the global information space corresponding to the original data, it is decomposed into three different feature levels, namely an approximate noise subspace, an approximate target subspace, and an approximate wall subspace, according to the rank from low to high. The wavelet scattering network (WSN) method is used to enhance the contrast and feature information for the subsequent input and understanding of the classifier. Then, the multi-layer conditional random field (CRF) probability graph model is used to train, verify, and optimize the global and three local information matrices respectively, and finally form the weights corresponding to each classification state of the four levels. Finally, the weight layers are fused to output the final recognition result. The present invention discloses a classifier with higher accuracy, faster convergence speed, and greater reliability, and combines existing clutter, noise suppression, and data enhancement algorithms to establish a complete solution for the through-wall radar recognition system. Under the same background conditions, the convergence speed is faster and the reliability is higher, while the accuracy can still be maintained at a high level.

[0042] Beneficial effects:

[0043] (1) The present invention makes full use of the information of the low-rank, medium-rank, and high-rank features of through-wall radar imaging, so as to achieve high-accuracy recognition without the need for a large number of parameters like traditional convolutional neural networks. It effectively improves the recognition performance in the scenario of extracting scattered and blurred features of human bodies passing through walls under strong clutter and noise interference.

[0044] (2) The present invention uses the probability graph model for modeling, which can combine the relationship of mutual influence and restriction at the pixel level of target imaging, and the method has strong interpretability.

[0045] (3) The method of the present invention realizes the integrated operation of recognition, data enhancement, and clutter suppression, avoiding additional computational expenses.

[0046] (4) Different from traditional through-wall radar human behavior data enhancement technologies, the method of the present invention combines the specific complex multipath effects of through-wall radar imaging and the physical background in the scenario of low signal-to-clutter ratio, so that in terms of the recognition accuracy of the method, better results can be achieved with less cost.

[0047] (5) The present invention uses the SVD subspace separation method to innovatively improve the clutter suppression method into three subspace separation algorithms, and uses the method of setting thresholds to obtain the information and imaging matrices of the clutter subspace, the target feature subspace, and the noise subspace respectively. Therefore, all information is retained to the greatest extent, and only the interval with partial low-rank and high-rank target information superimposed is cut off.

[0048] (6) The present invention utilizes a data augmentation scheme based on the Wavelet Scattering Network (WSN) and an edge detection algorithm based on the Canny operator, aiming to enhance image contrast, increase the detection threshold of feature contours, and make the performance of the detection results better.

[0049] (7) The present invention utilizes a probabilistic graphical model. Through the verification of multiple models, it is proved that the Conditional Random Field (CRF) is the best choice, thereby optimizing the algorithm to achieve the best. By building the comparison of the corresponding recognition accuracies using networks with different depths, the optimal characteristics of the method are proved.

[0050] (8) On the basis of ensuring a recognition accuracy of 97%, the present invention improves the existing accuracy by 2%, and the number of parameters is much less. By comparing with other types of probabilistic graphical models, the optimality of the method of the present invention is proved. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a schematic diagram of the geometric principle of the electromagnetic wave propagation characteristics of through-wall radar imaging;

[0052] Figure 2 It is a schematic diagram of the flow of the method for identifying moving targets by through-wall radar provided by the present invention;

[0053] Figure 3 It is a schematic diagram of the architecture of the method provided by the present invention;

[0054] Figure 4(A) is a schematic diagram of the function image of the wavelet basis function provided by the present invention;

[0055] Figure 4(B) is a schematic diagram of the signal flow of the wavelet scattering enhancement network provided by the present invention;

[0056] Figure 5 It is a schematic diagram of the implementation effect of the SVD subspace separation performed by the present invention;

[0057] Figure 6 It is a schematic diagram of the CRF (Conditional Random Field) of the present invention;

[0058] Figure 7(A)-Figure 7(D) It is a schematic diagram for comparing the solution of the present invention with other existing technical solutions;

[0059] Figure 8 It is a schematic diagram of the structure of the device for identifying moving targets by through-wall radar provided by the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0060] First, in combination with Figure 1 , the working principle of the through-wall radar is introduced and explained.

[0061] The stepped-frequency waveform is used as the transmission signal model of the through-wall radar system. The stepped-frequency signal transmits K frequency points, and the frequency of the transmission signal increases with the increase of the frequency points. The mathematical model of the transmission signal is:

[0062]

[0063] Among them, f0 represents the starting frequency of the stepped-frequency signal, and Δf represents the frequency step interval. The function rect represents the gate function, which is defined as follows:

[0064]

[0065] After the electromagnetic scattering by the wall and the human body, the echo signal received by the radar receiver can be expressed as

[0066]

[0067] Among them, S w (t) represents the scattered echo from the wall. S n (t) represents the noise component superimposed by the wall and the background effect received by the through-wall radar signal. a p represents the echo gain ratio of the radar to the target at the p-th point. τ p is the time delay of the point target. Figure 1 Shows the through-wall radar human behavior detection scenario, signal propagation model and geometric relationship unfolded along the cross-section perpendicular to the wall of the through-wall radar. Using the delay characteristic of the echo and summing up the point targets, the delay focusing imaging result of the echo is established.

[0068] Relative to the stationary radar wave source, the moving human body will produce the Doppler effect, which is expressed as

[0069]

[0070] Among them, f d (t) and represent the Doppler frequency shift and the Doppler phase shift respectively. v(t) and r(t) represent the relative motion speed and the relative rotation angle of the wave source and the moving target respectively. Similarly, the micro-Doppler effect is a physical phenomenon generated by the micro-motion of the object itself. From the theory of short-time Fourier transform (STFT), we know that the micro-Doppler characteristics of the human body can be characterized as the sum of the characteristics formed by a series of Doppler frequency shifts under the micro-motion of the target. Therefore

[0071] S r = S T + S W + S N (12)

[0072] For the frequency point f c , the corresponding denoised micro-Doppler expression is

[0073]

[0074] After I-Q demodulation, the following can be obtained

[0075]

[0076] Therefore, the wall enhances the noise and DC components of the micro-Doppler characteristics of people. The magnitude of the DC component value depends on the magnitude of the wall component and the ratio of the distance to the wavelength. Generally speaking, the globally effective information that can be extracted during the subsequent recognition process includes coordinates and macroscopic motion characteristics, while the locally effective information is mainly the micro-Doppler characteristics.

[0077] The present invention will be described in detail below in conjunction with the accompanying drawings and embodiments.

[0078] As Figure 2-Figure 3 shown, the present invention proposes a method for identifying moving targets by through-wall radar, including the following steps:

[0079] Step S1: Use a number of through-wall radar imaging data as training data, perform singular value decomposition on the training data, and enhance both the training data and the sub-components obtained after decomposing the training data by a wavelet scattering network to obtain an enhanced imaging matrix and enhanced sub-components;

[0080] Step S2: Input the enhanced imaging matrix and enhanced sub-components into their respective corresponding probabilistic graphical models. The probabilistic graphical model is a conditional random field (CRF), train each of the probabilistic graphical models to obtain the trained probabilistic graphical models; based on the outputs of the trained probabilistic graphical models, perform model fusion;

[0081] Step S3: Obtain the through-wall radar imaging data to be recognized, and recognize the through-wall radar imaging data to be recognized based on the fused model.

[0082] The present invention first performs cascaded singular value decomposition (SVD) and wavelet scattering network (WSN) data feature separation and enhancement processing on the radar echo imaging information. Secondly, the imaging matrix and the sub-components corresponding to its singular values arranged in descending order are respectively input into the corresponding weighted conditional random field (CRF) for training and inference to generate their respective weight outputs. Finally, model fusion is performed in the test stage, and the recognition result is given.

[0083] In the said step S1: Use a number of through-wall radar imaging data as training data, perform singular value decomposition on the training data, and enhance both the training data and the sub-components obtained after decomposing the training data by a wavelet scattering network to obtain an enhanced imaging matrix and enhanced sub-components, where:

[0084] The method of performing singular value decomposition on the training data is as follows:

[0085]

[0086] where φ r is the echo matrix corresponding to the through-wall radar imaging data, with dimensions N×M, and σ i represents the singular value after the SVD decomposition of the echo matrix, u i and v i respectively represent the left singular vector and the right singular vector after the SVD decomposition of the echo matrix, H represents the Hermitian transform of the echo matrix, i is the ordinal number of the singular value arranged in descending order, W is the wall clutter subspace, T is the target subspace, and N is the noise subspace; the output matrix represents three subspaces containing only human target imaging information, wall clutter information, and noise information.

[0087] In this embodiment, through SVD decomposition, subspaces can be obtained, the wall clutter subspace and the noise subspace can be suppressed, so as to retain the target subspace, and the echo matrix corresponding to the through-wall radar imaging data is decomposed into the sum of submatrices corresponding to ranks of different sizes.

[0088] The three output sub-components can all be used for recognition and detection work. However, due to the influence of the wall, the signal-to-clutter ratio (SCR) and signal-to-noise ratio (SNR) of the echo signal remain at a very low level. Such useful target signals are not conducive to the training and understanding of subsequent classifiers. To improve the recognition success rate, the present invention cascades the above singular value decomposition algorithm with a wavelet scattering network, that is, the sub-components after decomposing the training data are enhanced by the wavelet scattering network. In this embodiment, both the training data and the sub-components after decomposing the training data are enhanced by the wavelet scattering network to obtain an enhanced imaging matrix and enhanced sub-components.

[0089] The present invention uses the SVD method to improve the clutter suppression algorithm into a three-class subspace separation method, so that when identifying an object, the target motion characteristics at various rank levels can be retained to a great extent. By fusing, judging, tuning parameters, verifying, and optimizing the characteristics corresponding to each subspace and the characteristics corresponding to the global subspace, it is possible to achieve a sufficiently high recognition accuracy while ensuring a relatively small number of parameters of the subsequent conditional random field model.

[0090] The structure of the wavelet scattering network in this embodiment is a conventional wavelet scattering network structure in the art. Compared with the traditional Fourier transform, the wavelet scattering network uses a wavelet function to replace the activation function of the hidden nodes of the neural network. As Figure 4(A)-Figure 4(B) shown, in this embodiment, the corresponding weights of the input layer to the hidden layer and the threshold of the hidden layer are respectively characterized by the scale factor and time translation factor of the wavelet function.

[0091] The wavelet function is:

[0092]

[0093] Among them, WT(a, τ) is the representation of the input data in the wavelet domain, and the input data is the echo signal; f(t) is the time-domain representation of the input echo signal, ψ represents the wavelet transform basis function, t is the integration time variable, a is the dilation scale of the wavelet transform, and τ is the time delay in the wavelet domain. In this embodiment, the Morlet wavelet basis function is selected. In this embodiment, the data input to the wavelet scattering network is the enhanced imaging matrix and the enhanced sub-components, and the input data is actually the echo signal.

[0094] The convolutional basis function of the wavelet transform of the wavelet scattering network is expressed as

[0095] ψ j (x) = 2 -2j ψ(2 -j x) (17)

[0096] Among them, ψ j (x) is the output of the j-th layer of the wavelet scattering network, x represents the input of the node in the wavelet scattering network, and θ is defined as the rotation angle of the wavelet transform in the wavelet scattering network. Then:

[0097]

[0098] Therefore, the convolutional basis function corresponding to each layer in the wavelet scattering network is expressed as

[0099]

[0100] Among them, ψ l,θ,1 (x) is the first convolutional basis function of the l-th layer of the wavelet scattering network, l is the number of layers of the wavelet scattering network, ψ i,θ,2 (x) is the second convolutional basis function of the l-th layer of the wavelet scattering network, r θ is the rotation angle between each layer of the wavelet scattering network, is the transfer function of the wavelet scattering network, and k = log2a represents the logarithm of the scale transformation coefficient a in the wavelet transform.

[0101] Therefore, for an input echo information φ, the output obtained is:

[0102]

[0103] Among them, j represents the total number of scalings, L represents the horizontal scale of one layer of the wavelet scattering network. φ is the input echo information, is the total amount of the convolutional function between the hidden layers of the wavelet scattering network, is the echo information after weighting in the input layer, is the echo information after weighting in the first hidden layer, Provide 2 j Average energy distribution information of intra-field imaging diagrams at different scales and directions. Is an enhanced Scale-Invariant Feature Transform (SIFT) feature vector representation that provides amplitude and angle interaction information in a multi-scale domain. After passing through the wavelet scattering network, the useful information and contrast of the input corresponding image are enhanced, which makes the subsequent detection and recognition of weak signals more reliable. In the wavelet scattering network, in order to enable the imaging data to obtain the best enhanced contrast information, in this embodiment, θ = 1.5708° is selected as the condition for subsequent processing. After the imaging information matrix itself and its three SVD sub-components are respectively processed by the above-mentioned feature contrast enhancement, they enter the probability graph model training and inference process of the next stage. Both the training data and the sub-components obtained after decomposing the training data are enhanced by the wavelet scattering network to obtain an enhanced imaging matrix and enhanced sub-components. In this embodiment, the effects after SVD decomposition and enhancement of the entire image and each component are as Figure 5 shown.

[0104] Using the data enhancement scheme based on the wavelet scattering network (WSN) can effectively improve the contrast of imaging information and make the contour of imaging information clearer. In this way, when performing target recognition, the feature extraction pressure of the classifier network can be reduced, which is the theoretical basis for the algorithm to have a lower computational cost.

[0105] As Figure 6 shown, in step S2: the enhanced imaging matrix and the enhanced sub-components are respectively input into their corresponding probability graph models, and the probability graph models are probability graph models constructed by a Conditional Random Field (CRF). Each of the probability graph models is trained to obtain each trained probability graph model; based on the outputs of each trained probability graph model, model fusion is performed, where:

[0106] The enhanced imaging matrix and the enhanced sub-components are respectively input into their corresponding probability graph models, and the probability graph models are probability graph models constructed by a Conditional Random Field (CRF). From the signal model given by formula (3), it is known that the sampling of the through-wall radar is an equidistant process, and the human body movement is also a continuous overall process tending to linear change. Therefore, using the conditional random field of the undirected probability graph model to build a sub-model can obtain a higher physical model fit and convergence ability. From the Hidden Markov Model (HMM) assumption in the undirected probability graph, it is known that given the observation variable x, the state variable y forms a Markov random field. In the CRF, when the observation variable is given, the probability model of the state variable can be expressed as:

[0107] P(y r ∣x,y1,…,y r-1 ,y r+1,…,y R ) = P(y r ∣x,y r-1 ,y r+1 ), r = 1, 2, …, R (21)

[0108] The CRF models the conditional probability P(y∣x), and its parametric form is:

[0109]

[0110] where t q (y r+1 ,y r ,x,r) is a transition feature function defined on two adjacent label positions r and r + 1, used to characterize the correlation between adjacent label variables y r and y r+1 and the influence of the observation sequence x on them, and v s (y r ,x,r) is a state feature function defined on the label position r, used to characterize the influence of the observation sequence x on the label variable y r .

[0111] Simplify the above conditional random field form. For a hidden Markov model (HMM), assume that s1 is the number of its state transition features and s2 is the number of its state features, and the state probability density function is expressed as follows

[0112]

[0113] where t s and v s′ represent the forward and backward state transition functions.

[0114] Sum over the position s, and assume S = s1 + s2, to get

[0115]

[0116] Assume that w s is the weight of the feature f s (y,x), then

[0117]

[0118] Therefore, the probability representation of the conditional random field model is

[0119]

[0120] where p(y∣x) represents the conditional probability that the input x outputs y under certain conditions, and w srepresents the weight of the s-th node in the order of the conditional random field, S is the total number of nodes, exp represents the exponential function, and f s (y, x) is the conditional random field feature function corresponding to the conditional variable x and the state variable y, and z(x) is the normalization constant;

[0121] The probability graph model is the conditional random field (CRF). Various subclasses of the probability graph model are applicable to constructing the probability graph model.

[0122] The training process of the probability graph model is equivalent to the following optimization problem:

[0123]

[0124] where L(w) represents the loss function, w is the weight vector, and x r represents the r-th component of the input sequence, y r represents the r-th component of the output sequence, R is the total number of state nodes, and f s (y r , x r ) is the transfer feature function under the r-th node y r and the r-th condition x r , and z(x r ) is the normalization coefficient corresponding to the r-th node.

[0125] Similarly, the training process of the CRF model is equivalent to the following optimization problem:

[0126]

[0127] where

[0128]

[0129] and w represents the weight vector, and F(y, x) represents the feature vector corresponding to the current probability graph link:

[0130]

[0131] The probability graph models corresponding to the enhanced imaging matrix and the enhancer components, namely the four probability graph models corresponding to the global matrix, the wall subspace, the target subspace, and the noise subspace, output the confidence levels characterized by the marginal likelihood probabilities corresponding to each subspace respectively. By weighted summing the confidence levels of these four spaces, the total output confidence level can be obtained. This total output confidence level reflects the probability that the input belongs to a certain type of human behavior. In this embodiment, it is the probability of a certain type among seven types of human behaviors. Specifically, assume that the independently distributed training samples are Train = [Train(x, y)1, Train(x, y)2, ……, Train(x, y) M, the marginal likelihood function MLF is established as follows

[0132] MLF * = argmax p(Train∣MLF) (31)

[0133] Thus, the marginal likelihood result can be expressed as:

[0134]

[0135] where p represents the maximum likelihood probability of the input data under the hyperparameters , represents the joint likelihood estimation of MLF such that the estimated result is probabilistically consistent with the feature distribution of the data Train after radar echo preprocessing, is a hyperparameter that can be dynamically adjusted, MLF is the marginal likelihood function, is the regularization term, M is the training sample size, and Dim() represents the number of independent parameters.

[0136] Based on the outputs of the trained conditional random field models, model fusion is performed, and the fused model is BIC(MLF).

[0137] In step S3: The through-wall radar imaging data to be recognized is obtained, and based on the fused model, the through-wall radar imaging data to be recognized is recognized, where:

[0138] The through-wall radar imaging data φ to be recognized r The total output confidence score is MLF, and take fus = [w fus,1 , w fus,2 , w fus,3 , w fus,G , Solve:

[0139]

[0140] R BIC = fus · BIC(MLF) (34)

[0141] That is, the final recognition result is output, where F model represents the CRF model transfer function of the ordinal model, w is the weight vector to be learned, y model-1 is the state model of the previous training step, y model is the state model of the current training step, x is the conditional vector, BIC(MLF) is the prediction preference score based on the likelihood function, R BIC represents the total output confidence of the fused model, fus represents the fusion step of dot product summation, y *Represents the value of the variable y when the weighted sum of the outputs of each model is maximum.

[0142] Furthermore, before step S1, it includes: dividing the states of the human body behind the wall into seven categories, including scenes of the human body moving back and forth parallel to the wall, scenes of the human body moving back and forth obliquely behind the wall, scenes of the human body rotating behind the wall, scenes of the human body squatting behind the wall, scenes of the human body stepping on the spot behind the wall, scenes of the human body standing still behind the wall and empty scenes, and the acquired training data covers the above seven categories of data.

[0143] The division method of the present invention includes most of the movement states of people in a room under normal circumstances and is not limited by the problem of visual angle asymmetry.

[0144] In order to better illustrate the purpose and advantages of the present invention, specific implementation methods are described, and further description is given below in conjunction with the accompanying drawings, experiments and data analysis.

[0145] The effectiveness of the algorithm is proved by systematically verifying the performance of this method in the actual scenario of through-wall radar, and the optimality of the algorithm is proved by comparing the existing probability graph algorithm and the verification effect of replacing the conditional random field in the model with other sub-models. The scene of experimental data collection and the processing flow remain consistent with those in theoretical analysis. The equipment used in the experiment is an ultra-wideband radar system based on Keysight N9923A handheld vector network analyzer (VNA) and Vivaldi antenna. The antenna operation and control are completed by the guide rail and the host computer. By initializing the coordinates of the two-dimensional guide rail platform, the end of the antenna can be fixed close to the wall for detection. The scene used in the experiment is a two-layer concrete brick wall.

[0146] Through the method of decomposing the singular value of global imaging information into three sub-components mentioned in the above algorithm theory, the wall, target and noise subspaces are effectively focused. Then they are contrast enhanced by the wavelet scattering method, where a short walk of the human body parallel to the wall is taken as an example, and the back of the wall is used as the initial value of the distance coordinate of the human activity space. In order to more intuitively describe the characteristics of the matrix components corresponding to the singular value size distribution, the Canny edge detection operator is used to process the imaging results. The experiment proves the reliability of the data preprocessing module. The recognition state behind the wall is divided into 7 categories, and the same steps are used for preprocessing to establish a data set containing 2160×4 frame data (divided into 80% training set and 20% verification set), and all the 2160 sets of global and sub-component data are input into the proposed multi-layer hybrid dynamic probability graph model for training and verification.

[0147] In the prior art article "Triple-Link Fusion Decision Method for Through-the-Wall Radar Human Motion Recognition", team members proposed a method of residual neural network improved by permutation attention (SA-Inception-ResNet) for through-the-wall radar human recognition. Figure 7(A)-Figure 7(D) The accuracy curves of the training process and verification process of the present invention are shown and compared with the methods of the prior art. The final convergence accuracy of the network training process of the present invention is 100%, and the final convergence accuracy of the verification process is 97.69%. This result exceeds the training convergence accuracy of 99.07% and the verification convergence accuracy of 95.60% of the previously proposed SA-Inception-ResNet method. And during the iteration process, the proposed multi-layer hybrid dynamic probability map method can reach convergence faster, and the fluctuation range of the accuracy in the verification process is smaller, which means that the efficiency and reliability of the algorithm are significantly improved. When training with the hardware configuration, due to its relatively large number of parameters, the SA-Inception-ResNet network takes about 4 hours to complete training. Compared with the training time required for the training process of the 12-layer 4-link model of the present invention, it only takes about 2.5 hours, effectively saving about 38% of the training time cost. The confusion matrices in Figure 7(C) and Figure 7(D) give the specific performance of the two algorithms in each category. It can be seen that the present invention can exceed the original method because of its more superior conditional random field topological modeling relationship, making the algorithm more efficient in identifying three more complex states of human motion information: standing still, walking back and forth in parallel, and walking diagonally.

[0148] As shown in the present invention Figure 3 During the probability map feature annotation stage, while keeping the number of layers of the probability map sub-model in each link unchanged, by replacing the conditional random field with other probability map models. Through comparison, it is found that the conditional random field is the sub-model of the probability map model commonly used in engineering that is most suitable for the current algorithm. Its algorithm convergence speed is the fastest among the four methods, and it can reach the highest verification accuracy. The comparison of the current experiment also proves that for the construction of the moving target recognition algorithm of through-the-wall radar, the undirected graph network algorithm can obtain higher convergence accuracy and faster convergence speed than the directed graph algorithm. Similarly, during the probability map feature annotation stage, while keeping the type of the probability map sub-model in each link constantly as the conditional random field, by changing the number of layers of the stacked conditional random fields, the performance of the comparison algorithm on the existing through-the-wall human behavior recognition dataset is shown as Figure 7(A)-Figure 7(D)As shown in the figure. Since the present invention is an improvement solution based on the idea of widening and deepening, when the number of layers of the probability graph model of each link exceeds 12 layers, the total number of layers of its algorithm network exceeds 50 layers of the previous algorithm SA-Inception-ResNet. When the number of layers of the probability graph model does not exceed 3 layers, it is verified that the network does not converge under the same parameter conditions. By comparing the network depth setting schemes of 4 to 12 layers, it is found that as the number of layers of the probability graph model increases, both the convergence ability and the prediction accuracy do not decrease monotonically. When the network depth reaches 11 conditional random field superposition layers, the training accuracy of the model of the present invention converges to 100%, and the verification accuracy also exceeds the existing algorithms. When the network depth reaches 12 conditional random field superposition layers, the algorithm achieves the best verification accuracy of 97.69%. This proves that the design scheme of the present invention is superior in terms of accuracy, convergence ability, parameter scale, and reliability.

[0149] During the experiment, the present invention was compared with the prior art solutions, and it was proved that the present invention can achieve an accuracy of 97.69%, effectively improving the accuracy of the best convolutional neural network algorithm by 2.09%. In addition, through the idea of controlling variables, while keeping the hyperparameters of model training unchanged, the types of probability graph submodels and the overlapping layers of probability graph submodels were adjusted respectively, and the technical effects of the solution of the present invention were verified.

[0150] In this embodiment, the three subspaces decomposed by the SVD method respectively pass through a 12-layer conditional random field sub-module, and then the global information graph of radar imaging is input into another 12-layer conditional random field sub-module, and the iterative results of the global information are respectively applied to the verification, adjustment, and optimization of itself and the three sub-space modules. After training is completed, the output model can be directly used for the application of through-wall radar human behavior recognition. The inference process still maintains the same model structure. The finally output weights are for the three subspaces and the global space. Therefore, the four weights need to be combined, and finally the total decision weight is output. The weights can intuitively show the understanding and judgment results of the network algorithm for the recognition task. Through a large number of experiments, it is proved that the hyperparameters involved in the construction of this method are the optimal choices.

[0151] The scenario and processing flow of the experimental data collection in the verification process of the present invention are consistent with those in the theoretical analysis. The equipment used in the experiment is an ultra-wideband radar system based on a Keysight N9923A handheld vector network analyzer (VNA) and Vivaldi antennas. The antenna motion control is completed by the guide rail and the upper computer. By initializing the coordinates of the two-dimensional guide rail placement platform, the end of the antenna can be fixed close to the wall for detection. The experimental scenario used is a two-layer concrete brick wall. Using the present invention, better imaging quality, better algorithm deployment and verification results can be obtained, and its settings can be flexibly adjusted.

[0152] The present invention also provides a through-wall radar moving target recognition device, as Figure 8 shown. The device includes:

[0153] Decomposition module: configured to use a plurality of through-wall radar imaging data as training data, perform singular value decomposition on the training data, and enhance both the training data and the sub-components after decomposition of the training data by a wavelet scattering network to obtain an enhanced imaging matrix and enhanced sub-components;

[0154] Fusion module: configured to input the enhanced imaging matrix and enhanced sub-components into their respective corresponding probability graph models, where the probability graph model is a conditional random field (CRF), train each of the probability graph models to obtain the trained probability graph models; based on the outputs of the trained probability graph models, perform model fusion;

[0155] Recognition module: configured to obtain the through-wall radar imaging data to be recognized and recognize the through-wall radar imaging data to be recognized based on the fused model.

[0156] The above specific embodiments only describe the design principle of the present invention. The shapes and names of the components in this description can be different and are not limited. Therefore, those skilled in the art of the present invention can modify or equivalently replace the technical solutions recorded in the foregoing embodiments; and these modifications and replacements do not depart from the spirit and technical solutions of the present invention and should all fall within the protection scope of the present invention.

Claims

1. A method for identifying moving targets by through-wall radar, characterized in that, Including: Step S1: Use a number of through-wall radar imaging data as training data, perform singular value decomposition on the training data, and enhance both the training data and the sub-components obtained after decomposing the training data using a wavelet scattering network to obtain an enhanced imaging matrix and enhanced sub-components; Step S2: Input the enhanced imaging matrix and enhanced sub-components into their respective corresponding probability graph models. The probability graph model is a conditional random field. Train each of the probability graph models to obtain the trained probability graph models; Based on the outputs of the trained probability graph models, perform model fusion; Step S3: Obtain the to-be-recognized through-wall radar imaging data, and recognize the to-be-recognized through-wall radar imaging data based on the fused model; The convolutional basis function of the wavelet transform of the wavelet scattering network is represented as ψ j (x) = 2 -2j ψ(2 -j x) where ψ j (x) is the output of the j-th layer of the wavelet scattering network, x represents the input of the nodes in the wavelet scattering network, and θ is defined as the rotation angle of the wavelet transform in the wavelet scattering network, then The convolutional basis function corresponding to each layer in the wavelet scattering network is expressed as Among them, ψ l,θ,1 (x) is the first convolutional basis function of the l-th layer of the wavelet scattering network, where l is the number of layers of the wavelet scattering network, and ψ l,θ,2 (x) is the second convolutional basis function of the l-th layer of the wavelet scattering network, and r θ is the rotation angle between each layer of the wavelet scattering network, is the transfer function of the wavelet scattering network, k = log2a represents the logarithm of the scale transformation coefficient a in the wavelet transform, and ψ() represents the Morlet basis function of the wavelet scattering network.

2. The method according to claim 1, characterized in that, In step S1: The singular value decomposition of the training data is performed in the following way: where φ r is the echo matrix corresponding to the through-wall radar imaging data, with dimensions N×M, and σ i represents the singular value after the SVD decomposition of the echo matrix, u i and v i respectively represent the left and right singular vectors after the SVD decomposition of the echo matrix, H represents the Hermitian transform of the echo matrix, i is the ordinal number of the singular value arranged in descending order, W is the wall clutter subspace, T is the target subspace, and N is the noise subspace; the output matrix represents three subspaces containing only human target imaging information, wall clutter information, and noise information.

3. The method according to claim 1, characterized in that, The wavelet function of the wavelet scattering network is: Where WT(a,τ) is the representation of the input data in the wavelet domain, and the input data is the echo signal; f(t) is a measure of the time-domain representation of the input echo signal, t is the integration time variable, a is the dilation scale of the wavelet transform, and τ is the time delay in the wavelet domain.

4. The method according to claim 1, characterized in that, The probability representation of the conditional random field model is Among them, p(y∣x) represents the conditional probability of output y under the condition of input x, and w s represents the weight of the s-th node in the conditional random field order, S is the total number of nodes, exp represents the exponential function, and f s (y, x) is the conditional random field feature function corresponding to the conditional variable x and the state variable y, and z(x) is the normalization constant; The probability graph model is the conditional random field.

5. The method according to any one of claims 1-4, characterized in that, Based on the outputs of the trained conditional random field models, perform model fusion. The fused model is BIC(MLF): where p represents the maximum likelihood probability of the input data under the hyperparameter , represents the joint likelihood estimation of the MLF, making the estimation result probabilistically consistent with the feature distribution of the data Train after radar echo preprocessing, is a hyperparameter that can be dynamically adjusted, MLF is the marginal likelihood function, is the regularization term, M is the training sample size, and Dim() represents the number of independent parameters.

6. A device for identifying moving targets of a through-wall radar, characterized in that, Including: Decomposition module: Configured to use a number of through-wall radar imaging data as training data, perform singular value decomposition on the training data, and enhance both the training data and the sub-components obtained after decomposing the training data using a wavelet scattering network to obtain an enhanced imaging matrix and enhanced sub-components; The convolutional basis function of the wavelet transform of the wavelet scattering network is expressed as ψ j (x) = 2 -2j ψ( 2-j x) where ψ j (x) is the output of the j-th layer of the wavelet scattering network, x represents the input of the nodes in the wavelet scattering network, and θ is defined as the rotation angle of the wavelet transform in the wavelet scattering network. Then The convolutional basis function corresponding to each layer in the wavelet scattering network is expressed as where, ψ l,θ,1 (x) is the first convolutional basis function of the l-th layer of the wavelet scattering network, l is the number of layers of the wavelet scattering network, ψ l,θ,2 (x) is the second convolutional basis function of the l-th layer of the wavelet scattering network, r θ is the rotation angle between each layer of the wavelet scattering network, is the transfer function of the wavelet scattering network, k = log2a represents the logarithm of the scale transformation coefficient a in the wavelet transform; ψ() represents the Morlet basis function of the wavelet scattering network; Fusion module: Configured to input the enhanced imaging matrix and enhanced sub-components into their respective corresponding probability graph models. The probability graph model is a conditional random field. Train each of the probability graph models to obtain the trained probability graph models; Based on the outputs of the trained probability graph models, perform model fusion; Recognition module: Configured to obtain the to-be-recognized through-wall radar imaging data, and recognize the to-be-recognized through-wall radar imaging data based on the fused model.

7. A system, characterized in that, The system includes: A processor for executing multiple instructions; A memory for storing multiple instructions; Among them, the multiple instructions are used to be stored by the memory and loaded and executed by the processor to perform the method according to any one of claims 1-5.

8. A computer-readable storage medium, in which multiple instructions are stored; the multiple instructions are used to be loaded and executed by a processor to perform the method according to any one of claims 1-5.

Images