Through-the-wall radar target identification method and system based on deep learning

Abstract

The invention discloses a through-the-wall radar target identification method and system based on deep learning, and belongs to the technical field of crossing of radar signal processing and artificial intelligence. The method comprises the steps of collecting echo data of various types of targets behind a wall body through a through-wall radar, and constructing an original echo signal matrix; inhibiting wall clutter and static background interference by adopting a singular value decomposition subspace projection technology, and extracting a purified target signal; short-time Fourier transform is carried out on the purification data, rearrangement optimization of a time-frequency spectrogram is realized by calculating an energy center of gravity, and high-resolution time-frequency features are generated; and training a large convolution kernel redundancy removal network based on the rearranged features, and achieving accurate classification and recognition of various targets such as human bodies and unmanned aerial vehicles behind the wall. Through organic combination of clutter suppression, feature enhancement and deep learning classification, the target recognition precision and real-time performance of the through-the-wall radar in a complex shielding environment are effectively improved, and a reliable detection means is provided for scenes such as urban combat.

Description

Technical Field

[0001] This invention relates to the field of through-wall radar target recognition technology, specifically to a through-wall radar target recognition method and system based on deep learning. Background Technology

[0002] Urban warfare, as one of the main forms of modern warfare, involves numerous personnel and weapons concealed behind buildings, using these structures to disrupt and impede enemy operations. Accurate identification of targets obscured by urban buildings and other environmental factors is crucial for gaining intelligence in urban warfare. Currently, the detection beams of relevant sensors struggle to penetrate these obstructions for accurate detection of the enemy and their equipment, leading to difficulties in target identification and engagement. Accurate identification of targets behind obstructions is a hot research topic both domestically and internationally. Microwave target imaging is currently the primary method for detecting targets through walls. However, this method is limited by the influence of reflected echoes from wall surfaces, requiring close proximity to the wall, which restricts its widespread use. Furthermore, imaging-based methods for identifying and detecting targets behind walls suffer from long imaging times and large radar sizes. Therefore, there is an urgent need for miniaturized, handheld wall-penetrating identification devices for individual soldiers.

[0003] Currently, 2D and 3D through-wall radars are bulky, inconvenient to apply, and have long imaging times, making them unsuitable for scenarios with high real-time requirements. One-dimensional through-wall radar, due to its convenient data acquisition and processing, is more suitable for urban warfare scenarios with high real-time requirements. However, through-wall radar target identification primarily focuses on human posture recognition. For future urban warfare, research on the identification of targets such as small drones, small unmanned vehicles, and people in building defense and attack scenarios is currently lacking. Therefore, starting from the time-frequency characteristics of different targets, this paper proposes a large-core redundancy removal network to classify the time-frequency features of different targets, addressing the problem of ineffective identification of targets obscured by walls. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based through-wall radar target recognition method, which solves the technical problem of how to use deep learning methods to accurately identify targets obscured by walls.

[0005] This invention is achieved through the following technical solution:

[0006] In a first aspect, this application provides a deep learning-based through-wall radar target recognition method, comprising the following steps:

[0007] Based on the real echo data of multiple targets behind the wall collected by through-wall radar, a raw echo signal matrix containing time sequence information is constructed.

[0008] The original echo signal matrix is ​​processed by the singular value decomposition subspace projection method to separate wall clutter from the target signal and filter out static background interference, and output purified echo data.

[0009] Short-time Fourier transform is performed on the purified echo data, and the rearranged time-frequency features are generated by calculating the energy distribution of the energy centroid redistribution spectrum.

[0010] The time-frequency features are divided into training, validation and test sets. A large convolutional kernel redundancy removal network is trained, and the features of the actual collected signals are identified and classified based on the trained model, outputting the results of multiple target categories behind the wall.

[0011] A further optimized solution is that the subspace projection method based on singular value decomposition specifically includes:

[0012] Singular value decomposition is performed on the two-dimensional data matrix obtained by the through-wall radar to obtain multiple singular components, including wall clutter subspace components, target echo subspace components, and noise subspace components.

[0013] Based on the zero-frequency characteristics of the static wall echo, the main frequency of the singular component is analyzed by fast Fourier transform, thereby selecting the singular component with the main frequency of zero frequency as the wall subspace component.

[0014] The processed echo signal matrix is ​​reconstructed by subtracting the wall subspace component from the original two-dimensional data matrix. The processed echo signal matrix contains target subspace and noise subspace signals.

[0015] A further optimized solution is that the method for determining the wall subspace is as follows:

[0016] Frequency analysis is performed on the singular components after singular value decomposition. If the dominant frequency of a singular component is zero, it is determined to belong to the wall subspace, which is represented as follows:

[0017] ;

[0018] in, Zero-frequency threshold, For the first A singular value, and They are the first There are two left singular vectors and one right singular vector. For Fast Fourier Transform operations, This is the Hermitian transpose operation, where M is the number of singular vectors.

[0019] A further optimization scheme is as follows: the spectrum optimization method based on time-frequency rearrangement specifically includes:

[0020] A short-time Fourier transform is performed on the preprocessed echo data to obtain the time-spectrum.

[0021] The centroid of the signal energy is calculated using the smoothed spectrum of the Wigner-Ville distribution.

[0022] The spectral values ​​are moved to the centroid of the signal energy to generate a rearranged time-frequency characteristic map.

[0023] A further optimized solution is that the energy centroid is calculated as follows:

[0024] Calculate the centroid coordinates using the following formula. :

[0025] ;

[0026] ;

[0027] in, and These are the Wigner-Ville distributions of the signal and the window function, respectively. These are the estimated values ​​of the rearranged time coordinates; These are the rearranged frequency coordinate estimates; The original signal to be analyzed; For integration time; For the integral frequency variable; Original time coordinates; These are the original frequency coordinates.

[0028] A further optimized solution is that the large convolutional kernel and redundancy-removing target recognition network includes the following structure:

[0029] The initial layer, consisting of multiple regular convolutional layers and depthwise separable convolutional layers, is used for preliminary feature extraction.

[0030] Multiple intermediate modules, each including a stage module and a transition layer, wherein the stage module consists of p repeated RepLK blocks, ConvFFN blocks and De-R Conv blocks cascaded together;

[0031] The final stage module, including the RepLK block and the ConvFFN block, is used for final feature compression;

[0032] The fully connected layer maps the feature vectors to the target category output.

[0033] A further optimization scheme is that the specific structure of the RepLK block includes a BN layer with a size of 1×1× The convolutional layer Conv_Rij and its size are 7×7× The depthwise separable convolutional layer DW_Rij has the following output:

[0034] ;

[0035] In the formula, For the first Phase, First Output feature maps of heavily parameterized large convolutional kernel blocks; For the first Phase, First A reparameterized large convolutional kernel block is a reparameterized convolutional layer; BN stands for batch normalization. For the first Phase, First Depth-separable convolutional layers with heavily parameterized large convolutional kernel blocks.

[0036] A further optimized solution involves the following steps in the processing of the large convolutional kernel and the redundancy-removing target recognition network:

[0037] Input a two-dimensional time-frequency rearranged feature map starting layer, which is composed of multiple ordinary convolutional layers and depthwise separable convolutional layers alternately, perform preliminary feature extraction, and obtain the output of the starting layer;

[0038] The output of the start layer is fed into multiple cascaded intermediate modules for deep feature extraction. Each intermediate module includes a stage module and a transition layer. The stage module consists of p cascaded RepLK blocks, ConvFFN blocks and De-R Conv blocks of the same number, and the output of the intermediate module is obtained.

[0039] The intermediate module outputs to the final stage module, which includes p equal numbers of RepLK blocks and ConvFFN blocks, to perform final feature extraction and obtain the output of the final stage module.

[0040] The output of the stretching end stage module is a one-dimensional vector, which is fed into the fully connected layer to output the target category prediction result. The loss function used for network training is the mean square error between the prediction result and the true label.

[0041] A further optimization is that the loss function used for network training is the mean squared error loss. As shown in the following formula:

[0042] ;

[0043] in, For the network's predicted output, This is a real label.

[0044] Secondly, this application provides a deep learning-based through-wall radar target recognition system, comprising:

[0045] The data acquisition module is used to collect real echo data of various targets behind the wall based on the through-wall radar and construct a raw echo signal matrix containing time sequence information.

[0046] The preprocessing module, which is communicatively connected to the data acquisition module, is used to process the original echo signal matrix using the singular value decomposition subspace projection method, separate wall clutter from the target signal and filter out static background interference, and output purified echo data.

[0047] The feature extraction module, which is communicatively connected to the preprocessing module, is used to perform short-time Fourier transform on the purified echo data and generate rearranged time-frequency features by calculating the energy centroid and redistributing the spectral energy distribution.

[0048] The identification and classification module is communicatively connected to the feature extraction module. It is used to divide the rearranged time-frequency features into training set, validation set and test set, train a large convolutional kernel redundancy removal network, and identify and classify the features of the actual sampled signal based on the trained model, and output the results of multiple target categories behind the wall.

[0049] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0050] By utilizing singular value decomposition subspace projection, wall clutter and static background interference can be effectively filtered out, improving the signal-to-noise ratio of the target echo and overcoming the problem of low recognition accuracy caused by wall reflection and noise in traditional technologies.

[0051] By combining short-time Fourier transform with time-frequency rearrangement techniques, the energy centroid is calculated to optimize the spectral distribution, which enhances the discriminative power of the target Doppler features and thus improves the ability to characterize weak or overlapping signals.

[0052] Based on the design of a large convolutional kernel redundancy removal network, efficient classification and recognition of multiple types of targets is achieved, which solves the shortcomings of existing technologies that are limited to human posture recognition. It also optimizes real-time processing performance, making it more suitable for scenarios with high timeliness requirements, such as urban warfare, and provides a reliable solution for through-wall radar target recognition. Attached Figure Description

[0053] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0054] In the attached image:

[0055] Figure 1 A flowchart illustrating the deep learning-based through-wall radar target recognition method provided in this application embodiment;

[0056] Figure 2 Another flowchart of the deep learning-based through-wall radar target recognition method provided in the embodiments of this application;

[0057] Figure 3 This is a schematic diagram of a RepLK block provided in an embodiment of this application;

[0058] Figure 4 A schematic diagram of a ConvFFN block provided in an embodiment of this application;

[0059] Figure 5 This is a schematic diagram of the De-R Conv block provided in an embodiment of this application;

[0060] Figure 6 This is a schematic diagram of the convergence curve of network training loss provided in an embodiment of this application. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0062] First, some of the technical terms used in this application will be explained to help those skilled in the art understand this application.

[0063] FFT: Fast Fourier Transform;

[0064] SVD: Singular Value Decomposition;

[0065] STFT: Short-Time Fourier Transform;

[0066] WAD: Wigner-Ville Distribution;

[0067] RepLK: Reparameterized Large Kernel Convolution Block;

[0068] ConvFFN: Convolutional Feedforward Network Block;

[0069] De-R Conv: De-Redundancy Convolution Block;

[0070] BN: Batch Normalization;

[0071] GeLU: Gaussian Error Linear Unit;

[0072] Conv: Convolution;

[0073] DW: Depthwise Separable Convolution.

[0074] Firstly, such as Figure 1 As shown, this application provides a deep learning-based through-wall radar target recognition method, including the following steps:

[0075] Step S1: Based on the real echo data of multiple targets behind the wall collected by the through-wall radar, construct the original echo signal matrix containing time sequence information;

[0076] Step S2: The original echo signal matrix is ​​processed using the singular value decomposition subspace projection method to separate wall clutter from the target signal and filter out static background interference, and output the purified echo data.

[0077] Step S3: Perform a short-time Fourier transform on the purified echo data, and generate rearranged time-frequency features by calculating the energy distribution of the energy centroid redistribution spectrum.

[0078] Step S4: Divide the rearranged time-frequency features into training set, validation set and test set, train a large convolutional kernel redundancy removal network, and identify and classify the features of the actual collected signal based on the trained model, and output the results of multiple target categories behind the wall.

[0079] This embodiment uses through-wall radar to collect real echo data of various targets behind walls, constructs an original echo signal matrix containing time-series information, and employs subspace projection technology based on singular value decomposition to effectively separate wall clutter from target signals and filter out static background interference. Based on this, a short-time Fourier transform is performed on the purified data, and the energy distribution of the spectrum is redistributed by calculating the energy centroid, generating rearranged time-frequency features with higher time-frequency resolution. Finally, a large convolutional kernel redundancy removal network is used to train and recognize the features, achieving accurate classification of various targets behind walls, such as humans and drones. This scheme, through the organic combination of clutter suppression, feature enhancement, and deep learning classification, improves the target recognition capability of through-wall radar in complex obstructed environments, overcoming the limitations of traditional imaging radars such as large size and slow processing speed.

[0080] In one embodiment, step S1: Based on the real echo data of multiple targets behind the wall collected by through-wall radar, a raw echo signal matrix containing time-series information is constructed, which specifically includes the following steps:

[0081] Step S11: Deploy the through-wall radar system and collect echo data for different types of targets. During data collection, the radar operates in the microwave band (e.g., 1-4 GHz) to penetrate common walls (e.g., brick walls, concrete walls) and record the original radar signal of the target behind the wall.

[0082] Step S12: Organize the acquired raw signals in a structured manner; specifically, set the radar to have P consecutive measurement cycles, sample Q points in each cycle, and construct the data into a raw echo signal matrix of size P×Q. ; where each column vector of the matrix This represents the time-series data sequence obtained across P measurement periods at the q-th sampling point; the original echo signal matrix is... The mathematical representation of is:

[0083] Equation (1)

[0084] In the formula, column vector This represents the time-series data sequence spanning P measurement cycles at the q-th sampling point. This matrix completely preserves the target's motion timing information, avoiding signal loss.

[0085] In one embodiment, step S2: The original echo signal matrix is ​​processed using the singular value decomposition subspace projection method to separate wall clutter from the target signal and filter out static background interference, outputting purified echo data. This specifically includes the following sub-steps:

[0086] Step S21: Process the original echo signal matrix Singular Value Decomposition (SVD) is the sum of M singular components, and its decomposition formula is:

[0087] Equation (2)

[0088] Physically, these singular components belong to three subspaces: the wall clutter subspace (W), the target echo subspace (T), and other noise subspaces (N). Based on the zero-frequency characteristics of stationary wall echoes, the wall subspace components are determined by frequency analysis and screening of the components obtained after singular value decomposition. Specifically, Fast Fourier Transform (FFT) is used to analyze each singular component to obtain its frequency composition. The dominant frequency component of the component is determined by identifying the spectral peaks. Components with dominant frequencies close to zero are identified as singular values ​​related to the wall echo, thereby determining and separating the wall clutter subspace.

[0089] Step S22: Based on the characteristic that the static wall echo has zero frequency (or near-zero frequency), perform Fast Fourier Transform (FFT) analysis on each singular component obtained from SVD decomposition. By identifying its dominant frequency component, the component with the dominant frequency of zero frequency (Z) is determined to belong to the wall clutter subspace, and the wall subspace signal is constructed accordingly. The judgment condition is:

[0090] Equation (3)

[0091] in, Zero-frequency threshold, For the first A singular value, and They are the first There are two left singular vectors and one right singular vector. For Fast Fourier Transform operations, This is the Hermitian transpose operation, where M is the number of singular vectors;

[0092] Step S23: Given that the amplitude of the noise subspace signal is relatively weak and random, to simplify the processing, the main focus is on suppressing the strong interference signal from the wall subspace, based on the original echo signal matrix. Subtract wall subspace signal The resulting echo signal matrix, which mainly contains target information after purification, is obtained. , is represented as:

[0093] Equation (4)

[0094] In the formula, This is the original echo signal matrix; For the wall subspace signal matrix; the first One component; This is the set of indices for the target signal components; This is the set of indices for the noise components.

[0095] This embodiment accurately suppresses wall clutter, improves the signal-to-noise ratio in typical scenarios, and overcomes the problem of target signal loss caused by traditional filtering.

[0096] In one embodiment, step S3: performing a short-time Fourier transform on the purified echo data, and generating rearranged time-frequency features by calculating the energy distribution of the energy centroid redistribution spectrum, specifically including the following sub-steps:

[0097] Step S31: Process the purified echo data matrix Short-Time Fourier Transform (STFT) is performed, windowing is applied in the time domain and sliding along the time axis to calculate the local spectrum of the signal in different time frames, thus obtaining an initial time-spectrum diagram to reveal the relationship between the signal's frequency components and time. The mathematical expression for STFT is:

[0098] Equation (5)

[0099] in, It is a window function (such as the Hamming window). For time and frequency Two-dimensional function, For summation index variables, The length of the window.

[0100] Step S32: Smooth the initial time-frequency spectrum using the Wigner-Ville distribution to obtain the spectrum. Then, for each point on the time-frequency surface... Calculate the centroid coordinates of the signal energy in its vicinity. The centroid coordinates are obtained using the following weighted average formula:

[0101] Equation (6)

[0102] Equation (7)

[0103] In the formula, and These are the Wigner-Ville distributions of the signal and the window function, respectively.

[0104] This step enhances the time-frequency resolution of the Doppler features of targets obscured by walls through time-frequency rearrangement technology, thereby improving the accuracy of target identification.

[0105] Step S33: Obtain the original time spectrum. Every point The value at that location is redistributed (rearranged) to the calculated energy barycenter location. Ultimately, this yields a rearranged time-frequency feature map with more concentrated energy and clearer characteristics. , is represented as:

[0106] Equation (8)

[0107] This step makes the rearranged feature energy more concentrated, improving the distinction between weak and overlapping targets;

[0108] In one embodiment, step S4: dividing the rearranged time-frequency features into a training set, a validation set, and a test set, training a large convolutional kernel redundancy removal network, and identifying and classifying the actual sampled signal features based on the trained model to output the results of multiple target categories behind the wall, specifically including the following sub-steps:

[0109] Step S41: Divide the generated large number of target rearranged time-frequency feature maps into training set, validation set and test set;

[0110] Step S42: Based on the training set, validation set, and test set, construct and train a target recognition network with large convolutional kernels to remove redundancy, such as... Figure 2 As shown, the processing flow is as follows:

[0111] Step S421: Input the rearranged time-frequency feature map SPEC to the start layer. This layer consists of alternating 1×1 ordinary convolutions (Conv) and 3×3 depthwise separable convolutions (DW) for preliminary feature extraction, resulting in... The formula for its calculation is:

[0112] Equation (9)

[0113] Step S422: Input Deep feature extraction is performed through q cascaded identical intermediate modules. Each intermediate module contains a stage module and a transition layer; the stage module is used for feature transformation and enhancement; the transition layer is used for transitioning feature map size or number of channels.

[0114] For the q-th intermediate module, its output It can be recursively defined as all front-end modules' inputs to it. The mathematical expression for the continuous processing result is:

[0115] Equation (10)

[0116] in, Representing the A complete intermediate module operation.

[0117] Each stage module consists of p repeated, structurally identical cascaded blocks to achieve deep feature learning. Specifically, the output of the i-th stage module... It is generated by the combined action of its internally cascaded RepLK blocks, ConvFFN blocks, and De-RConv blocks, and its calculation process is as follows:

[0118] Equation (11)

[0119] In the formula, This is the output of the previous intermediate module;

[0120] The core components and their specific operations within the stage module are as follows:

[0121] Reparameterized Large Kernel Convolutional Block (RepLK): This component employs large-size convolutional kernels (e.g., 7×7) to significantly expand the receptive field, thereby effectively capturing global contextual information in the feature map. It also utilizes structural reparameterization techniques to achieve a balance between depth and efficiency during training and inference. Its structure is as follows: Figure 3 As shown, for the j-th RepLK block in the i-th stage, let its input be... The output is Its core calculation formula is:

[0122] Equation (12)

[0123] In the formula, This represents point convolution (1×1 convolution). represents depthwise separable convolution, BN represents batch normalization, and residual connections ensure the effective propagation of gradients in the network.

[0124] Convolutional Feedforward Network Block (ConvFFN): This component introduces complex nonlinear transformations and information interactions between feature channels through a feedforward network consisting of two 1×1 convolutional layers and a GeLU activation function, thereby significantly enhancing the model's representational capabilities. Its structure is as follows: Figure 4 As shown, for the j-th ConvFFN block in the i-th stage, let its input be... The output is The formula for its calculation is:

[0125] Equation (13)

[0126] The residual structure here also stabilizes the training process of deep networks.

[0127] De-RConv (Deredundant Convolutional Block): This component is specifically designed to reduce redundant information between feature channels, improving feature compactness and discriminativeness through a unique dual-branch attention mechanism. Its structure is as follows: Figure 5 As shown, the specific operation process is as follows:

[0128] For the upper branch, first, the input Perform global average pooling, then pass it through two 1×1×C1 convolutional layers to obtain upper branch weights of size 1×1×C1. Utilize the weights of the upper branches Weighting is performed according to the channel direction to obtain the weighted features. , Divide the passageway into four equal parts along its direction based on the number of passageways. , , , .Will , , , Element-wise addition, followed by a 3×3×C1 convolutional layer, yields the upper branch features. ;

[0129] For the lower branch, adjust the weight of the upper branch. Flip along the channel direction to obtain the lower branch weight. Utilize the weight of the lower branch Weighting is performed according to the channel direction to obtain the weighted features. , Divide the passageway into four equal parts along its direction based on the number of passageways. , , , .Will , , , Element-wise addition, followed by a 3×3×C1 / 4 convolutional layer and a 1×1×C1 / 4 convolutional layer, yields the upper branch features. .

[0130] Then and By piecing the pieces together along the channel direction, the final result is obtained. ;

[0131] The mathematical expression for this process can be summarized as follows:

[0132] Equation (14)

[0133] This design uses attention weighting and grouping strategies to force the network to focus on more discriminative features, thereby effectively suppressing channel redundancy.

[0134] Through the sequential stacking and cascading collaboration of the three core components, the stage module achieves progressive learning from global context capture and non-linear feature enhancement to efficient and compact feature extraction. Large-kernel convolution ensures a broad perceptual field, the feedforward network deepens feature transformation capabilities, and redundancy removal significantly improves feature information density. These three components work closely together to form the powerful hierarchical feature extraction capability of the intermediate module, laying a solid feature foundation for the network to ultimately achieve accurate target classification.

[0135] Step S423: The features extracted by the intermediate module are further processed by a final stage module (mainly composed of RepLK and ConvFFN blocks). The output of the final stage module is:

[0136] Equation (15)

[0137] In the formula, This refers to the p-th convolutional feedforward network module in the i-th stage.

[0138] The stretched feature map is converted into a vector and then mapped to the dimension of the number of target classes through a fully connected layer, outputting a prediction vector. ;

[0139] Step S424: Using the mean squared error loss function as follows, optimize all network weight parameters through backpropagation until the model converges. For example... Figure 6 As shown, the training process is monitored using a validation set to prevent overfitting:

[0140] Equation (16)

[0141] Repeat the above steps until the parameters of each network or module in the entire model converge.

[0142] Step S43: The through-wall radar echo data to be identified is preprocessed and feature extracted in steps S1 to S3 in sequence. The resulting rearranged time-frequency feature map is input into the trained large convolutional kernel redundancy removal network model. Through the forward propagation of the network, the final output is the identification result of the target category behind the wall.

[0143] Secondly, this application provides a deep learning-based through-wall radar target recognition system, comprising:

[0144] The data acquisition module is used to collect real echo data of various targets behind the wall based on the through-wall radar and construct a raw echo signal matrix containing time sequence information.

[0145] The preprocessing module, which is communicatively connected to the data acquisition module, is used to process the original echo signal matrix using the singular value decomposition subspace projection method, separate wall clutter from the target signal and filter out static background interference, and output purified echo data.

[0146] The feature extraction module, which is communicatively connected to the preprocessing module, is used to perform short-time Fourier transform on the purified echo data and generate rearranged time-frequency features by calculating the energy centroid and redistributing the spectral energy distribution.

[0147] The identification and classification module is communicatively connected to the feature extraction module. It is used to divide the rearranged time-frequency features into training set, validation set and test set, train a large convolutional kernel redundancy removal network, and identify and classify the features of the actual sampled signal based on the trained model, and output the results of multiple target categories behind the wall.

[0148] The functions of each module in the deep learning-based through-wall radar target recognition system correspond to the steps in the deep learning-based through-wall radar target recognition method embodiment. Their functions and implementation processes will not be described in detail here.

[0149] Thirdly, embodiments of this application provide a deep learning-based through-wall radar target recognition device, which can be a personal computer (PC), laptop, server, or other device with data processing capabilities.

[0150] In this embodiment of the application, the deep learning-based through-wall radar target recognition device may include a processor, a memory, a communication interface, and a communication bus.

[0151] The communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.

[0152] The communication interface includes input / output (I / O) interfaces, physical interfaces, and logical interfaces used for interconnecting internal components of the deep learning-based through-wall radar target recognition device, as well as interfaces used for interconnecting the deep learning-based through-wall radar target recognition device with other devices (such as other computing devices or user equipment). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user equipment can be displays, keyboards, etc.

[0153] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0154] The processor can be a general-purpose processor, which can call a deep learning-based through-wall radar target recognition program stored in memory and execute the deep learning-based through-wall radar target recognition method provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the deep learning-based through-wall radar target recognition program is called can be referred to in the various embodiments of the deep learning-based through-wall radar target recognition method of this application, and will not be repeated here.

[0155] Fourthly, embodiments of this application also provide a readable storage medium.

[0156] This application has a readable storage medium storing a deep learning-based through-wall radar target recognition program, wherein when the deep learning-based through-wall radar target recognition program is executed by a processor, it implements the steps of the deep learning-based through-wall radar target recognition method described above.

[0157] The method implemented when the deep learning-based through-wall radar target recognition program is executed can be referred to in the various embodiments of the deep learning-based through-wall radar target recognition method of this application, and will not be repeated here.

[0158] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A deep learning-based through-wall radar target recognition method, characterized in that, Includes the following steps: Based on the real echo data of multiple targets behind the wall collected by through-wall radar, a raw echo signal matrix containing time sequence information is constructed. The original echo signal matrix is ​​processed by the singular value decomposition subspace projection method to separate wall clutter from the target signal and filter out static background interference, and output purified echo data. Short-time Fourier transform is performed on the purified echo data, and the rearranged time-frequency features are generated by calculating the energy distribution of the energy centroid redistribution spectrum. The time-frequency features are divided into training, validation and test sets. A large convolutional kernel redundancy removal network is trained, and the features of the actual collected signals are identified and classified based on the trained model, outputting the results of multiple target categories behind the wall.

2. The deep learning-based through-wall radar target recognition method according to claim 1, characterized in that, The subspace projection method based on singular value decomposition specifically includes: Singular value decomposition is performed on the two-dimensional data matrix obtained by the through-wall radar to obtain multiple singular components, including wall clutter subspace components, target echo subspace components, and noise subspace components. Based on the zero-frequency characteristics of the static wall echo, the main frequency of the singular component is analyzed by fast Fourier transform, thereby selecting the singular component with the main frequency of zero frequency as the wall subspace component. The processed echo signal matrix is ​​reconstructed by subtracting the wall subspace component from the original two-dimensional data matrix. The processed echo signal matrix contains target subspace and noise subspace signals.

3. The deep learning-based through-wall radar target recognition method according to claim 2, characterized in that, The method for determining the wall subspace is as follows: Frequency analysis is performed on the singular components after singular value decomposition. If the dominant frequency of a singular component is zero, it is determined to belong to the wall subspace. Represented as: ； in, Zero-frequency threshold, For the first A singular value, and They are the first There are two left singular vectors and one right singular vector. For Fast Fourier Transform operations, This is the Hermitian transpose operation, where M is the number of singular vectors.

4. The deep learning-based through-wall radar target recognition method according to claim 1, characterized in that, The spectrum optimization method based on time-frequency rearrangement specifically includes: A short-time Fourier transform is performed on the preprocessed echo data to obtain the time-spectrum. The centroid of the signal energy is calculated using the smoothed spectrum of the Wigner-Ville distribution. The spectral values ​​are moved to the centroid of the signal energy to generate a rearranged time-frequency characteristic map.

5. The deep learning-based through-wall radar target recognition method according to claim 4, characterized in that, The calculation method for the energy centroid is as follows: Calculate the centroid coordinates using the following formula. : ； ； in, and These are the Wigner-Ville distributions of the signal and the window function, respectively. These are the estimated values ​​of the rearranged time coordinates; These are the rearranged frequency coordinate estimates; The original signal to be analyzed; For integration time; For the integral frequency variable; Original time coordinates; These are the original frequency coordinates.

6. The deep learning-based through-wall radar target recognition method according to claim 1, characterized in that, The large convolutional kernel and redundancy-removing target recognition network includes the following structure: The initial layer, consisting of multiple regular convolutional layers and depthwise separable convolutional layers, is used for preliminary feature extraction. Multiple intermediate modules, each including a stage module and a transition layer, wherein the stage module consists of p repeated RepLK blocks, ConvFFN blocks and De-R Conv blocks cascaded together; The final stage module, including the RepLK block and the ConvFFN block, is used for final feature compression; The fully connected layer maps the feature vectors to the target category output.

7. The deep learning-based through-wall radar target recognition method according to claim 6, characterized in that, The specific structure of the RepLK block includes a BN layer with a size of 1×1× convolutional layers and the size is 7×7× Depth-separable convolutional layers Its output is calculated as follows: ； In the formula, For the first Phase, First Output feature maps of heavily parameterized large convolutional kernel blocks; For the first Phase, First A reparameterized large convolutional kernel block is a reparameterized convolutional layer; BN stands for batch normalization. For the first Phase, First Depth-separable convolutional layers with heavily parameterized large convolutional kernel blocks.

8. The deep learning-based through-wall radar target recognition method according to claim 6, characterized in that, The processing of the large convolutional kernel and the redundancy-removing target recognition network includes the following steps: Input a two-dimensional time-frequency rearranged feature map starting layer, which is composed of multiple ordinary convolutional layers and depthwise separable convolutional layers alternately, perform preliminary feature extraction, and obtain the output of the starting layer; The output of the start layer is fed into multiple cascaded intermediate modules for deep feature extraction. Each intermediate module includes a stage module and a transition layer. The stage module consists of p cascaded RepLK blocks, ConvFFN blocks and De-RConv blocks of equal number, and the output of the intermediate module is obtained. The intermediate module outputs to the final stage module, which includes p equal numbers of RepLK blocks and ConvFFN blocks, to perform final feature extraction and obtain the output of the final stage module. The output of the stretching end stage module is a one-dimensional vector, which is fed into the fully connected layer to output the target category prediction result. The loss function used for network training is the mean square error between the prediction result and the true label.

9. The deep learning-based through-wall radar target recognition method according to claim 1, characterized in that, The loss function used in the network training is the mean squared error loss. As shown in the following formula: ； in, For the network's predicted output, This is a real label.

10. A deep learning-based through-wall radar target recognition system, characterized in that, include: The data acquisition module is used to collect real echo data of various targets behind the wall based on the through-wall radar and construct a raw echo signal matrix containing time sequence information. The preprocessing module, which is communicatively connected to the data acquisition module, is used to process the original echo signal matrix using the singular value decomposition subspace projection method, separate wall clutter from the target signal and filter out static background interference, and output purified echo data. The feature extraction module, which is communicatively connected to the preprocessing module, is used to perform short-time Fourier transform on the purified echo data and generate rearranged time-frequency features by calculating the energy centroid and redistributing the spectral energy distribution. The identification and classification module is communicatively connected to the feature extraction module. It is used to divide the rearranged time-frequency features into training set, validation set and test set, train a large convolutional kernel redundancy removal network, and identify and classify the features of the actual sampled signal based on the trained model, and output the results of multiple target categories behind the wall.

Images