Target recognition method and device for single-platform radar multimodal data fusion

Abstract

This application relates to a target recognition method and device for single-platform radar multimodal data fusion. The method preprocesses radar scattering cross-section data and high-resolution one-dimensional range images obtained by detecting a target using a single platform. The preprocessed radar echo data is divided using Monte Carlo sampling, and the resulting subsequences are encoded to obtain a coded sequence. A trained target recognition network is then used to extract fusion features of the two modal echo data based on the coded sequence, and the target is then identified using a classification head based on the fusion features. This method is used to achieve cross-modal fusion of RCS and HRRP data, and provide well-characterized data features for subsequent target recognition tasks, thereby improving recognition performance and robustness.

Description

Technical Field

[0001] The present application relates to the field of radar signal processing and computer intelligence technology, and in particular to a target recognition method and device for multi-modal data fusion of a single-platform radar. Background Art

[0002] Radar Auto Target Recognition (RATR) plays an important role in national security and aviation safety. How to quickly, accurately and robustly identify aerospace targets is the most important task of RATR.

[0003] In RATR missions, the most commonly used radar signals are Radar Cross Section (RCS) and High Range Resolution Profile (HRRP). RCS data is easier to obtain in combat but contains less target information, while HRRP data contains more information about the target but is difficult to accurately measure in combat. Therefore, how to achieve complementary fusion of radar data from these two modalities to improve the accuracy and robustness of radar target recognition tasks is a pressing issue in the RATR field.

[0004] Meanwhile, traditional RATR uses template matching to extract and match statistical and mathematical features in radar signals to identify different targets. However, this method achieves low recognition rates. In recent years, deep learning-based methods have adopted a data-driven approach to capture and extract implicit features in radar signals, ultimately achieving higher recognition rates. However, deep learning methods are less robust and do not consider the fusion of cross-modal radar data. Summary of the Invention

[0005] Based on this, it is necessary to provide a target recognition method and device for single-platform radar multi-modal data fusion that can improve recognition accuracy in response to the above technical problems.

[0006] A target recognition method for multi-modal data fusion of a single-platform radar, the method comprising:

[0007] Acquire radar echo data of the target to be identified, wherein the radar echo data is radar scattering cross-section data and high-resolution one-dimensional range image obtained by single-platform detection;

[0008] Preprocessing the radar echo data, dividing the preprocessed radar echo data by Monte Carlo sampling, and then encoding the subsequences obtained after the division to obtain a coded sequence;

[0009] The coding sequence is input into a trained target recognition network. In the target recognition network, a cross-domain fusion feature extractor based on the self-attention mechanism extracts the fusion features of the radar scattering cross-section data and the high-resolution one-dimensional range image according to the coding sequence. The classification head then classifies the target according to the fusion features to obtain the target recognition result.

[0010] In one embodiment, when the radar echo data is preprocessed, normalization processing and center-of-gravity alignment processing are sequentially performed on the radar echo data to obtain the preprocessed radar echo data.

[0011] In one embodiment, the normalizing and aligning the radar echo data in sequence to obtain the pre-processed radar echo data includes:

[0012] Before preprocessing the radar scattering cross section data in the radar echo data, amplitude processing is first performed on the radar scattering cross section data to obtain real part data in the radar scattering cross section data, and subsequent preprocessing is performed based on the real part data;

[0013] Processing the radar echo data using modulo-2 norm normalization to obtain amplitude-normalized radar echo data;

[0014] The center alignment method is used to perform a circular shift operation on the amplitude-normalized radar echo data so that its center is located at the center of the range window, and the preprocessed radar echo data is obtained.

[0015] In one embodiment, dividing the preprocessed radar echo data by Monte Carlo sampling and then encoding the subsequences obtained after the division to obtain the coded sequence includes:

[0016] Divide the pre-processed radar echo data according to a preset length to obtain multiple subsequences with the same length;

[0017] Encoding each of the subsequences into subsequence encoding data represented by a D-dimensional vector, and adding a category hint vector to the front end of each of the subsequence encoding data and a segmentation vector to the back end;

[0018] The coding sequence is constructed based on the subsequence coding data, data sequence segment coding and time sequence position coding.

[0019] In one embodiment, the cross-domain fusion feature extractor based on the self-attention mechanism includes multiple layers of feature extraction layers connected in sequence;

[0020] The coding sequence is input into the first feature extraction layer to obtain the output data of the current layer, and then the output data of the current layer is used as the input of the next feature extraction layer, and so on, and the output data of the last feature extraction layer is used as the fusion feature.

[0021] In one embodiment, each feature extraction layer includes a first normalization layer, a multi-head attention layer, a second normalization layer and a feedforward layer connected in sequence.

[0022] In one embodiment, after the coding sequence or the output data of the previous feature extraction layer is input into the feature extraction layer:

[0023] The first normalization layer uses a quasi-norm to perform layer normalization on the coding sequence or the output data of the previous feature extraction layer to obtain first output data;

[0024] In the multi-head attention layer, the first output data is linearly projected three times to obtain a query matrix, a key matrix, and a value matrix, respectively. These three matrices are projected into multiple feature subspaces, and the attention results calculated in each subspace are concatenated to obtain the second output data.

[0025] After inputting the second output data and the encoding sequence into the second normalization layer, third output data is obtained;

[0026] The feedforward layer processes the third output data through two layers of fully connected multilayer perceptrons to obtain fourth output data;

[0027] The current layer feature extraction layer or the fusion feature is obtained according to the fourth output data and the second output data.

[0028] In one embodiment, a loss function based on a cross entropy formula is used when training the target recognition network.

[0029] A target recognition device for multi-modal data fusion of a single-platform radar, the device comprising:

[0030] A radar echo data acquisition module is used to acquire radar echo data of the target to be identified. The radar echo data includes radar scattering cross-section data and high-resolution one-dimensional range image obtained by single-platform detection;

[0031] a data preprocessing module, configured to preprocess the radar echo data, divide the preprocessed radar echo data into subsequences by Monte Carlo sampling, and then encode the subsequences obtained after the division to obtain a coded sequence;

[0032] The target recognition module is used to input the coding sequence into a trained target recognition network. In the target recognition network, a cross-domain fusion feature extractor based on the self-attention mechanism extracts the fusion features of the radar scattering cross-section data and the high-resolution one-dimensional range image according to the coding sequence, and then a classification head classifies the target according to the fusion features to obtain the target recognition result.

[0033] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0034] Acquire radar echo data of the target to be identified, wherein the radar echo data is radar scattering cross-section data and high-resolution one-dimensional range image obtained by single-platform detection;

[0035] Preprocessing the radar echo data, dividing the preprocessed radar echo data by Monte Carlo sampling, and then encoding the subsequences obtained after the division to obtain a coded sequence;

[0036] The coding sequence is input into a trained target recognition network. In the target recognition network, a cross-domain fusion feature extractor based on the self-attention mechanism extracts the fusion features of the radar scattering cross-section data and the high-resolution one-dimensional range image according to the coding sequence. The classification head then classifies the target according to the fusion features to obtain the target recognition result.

[0037] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0038] Acquire radar echo data of the target to be identified, wherein the radar echo data is radar scattering cross-section data and high-resolution one-dimensional range image obtained by single-platform detection;

[0039] Preprocessing the radar echo data, dividing the preprocessed radar echo data by Monte Carlo sampling, and then encoding the subsequences obtained after the division to obtain a coded sequence;

[0040] The coding sequence is input into a trained target recognition network. In the target recognition network, a cross-domain fusion feature extractor based on the self-attention mechanism extracts the fusion features of the radar scattering cross-section data and the high-resolution one-dimensional range image according to the coding sequence. The classification head then classifies the target according to the fusion features to obtain the target recognition result.

[0041] The above-mentioned target recognition method and device for single-platform radar multimodal data fusion preprocesses radar scattering cross-section data and high-resolution one-dimensional range profiles acquired by a single platform during target detection. The preprocessed radar echo data is then divided using Monte Carlo sampling, and the resulting subsequences are encoded to form a coded sequence. A trained target recognition network then extracts fused features from the two modal echo data based on the coded sequence, and a classification head then uses the fused features to identify the target. This method achieves cross-modal fusion of RCS and HRRP data, providing well-characterized data features for subsequent target recognition tasks, thereby improving recognition performance and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 1 is a flow chart of a target recognition method for multi-modal data fusion of a single-platform radar in one embodiment;

[0043] Figure 2 Schematic diagram of a multimodal unified representation of RCS and HRRP data in one embodiment;

[0044] Figure 3 FIG1 is a schematic diagram of radar echo data encoding based on Monte Carlo sampling in one embodiment;

[0045] Figure 4 Schematic diagram of the feature extraction layer structure in one embodiment;

[0046] Figure 5 This is a diagram showing the separability experimental verification results of this method in a simulation experiment;

[0047] Figure 6 1 is a block diagram of a target recognition device for multi-modal data fusion of a single-platform radar in one embodiment;

[0048] Figure 7 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0050] In view of the problem that the existing radar automatic target recognition technology has low recognition performance or high recognition performance but low recognition robustness, in this embodiment, Figure 1 As shown in the figure, a target recognition method for multi-modal data fusion of single-platform radar is proposed, including:

[0051] Step S100: Acquire radar echo data of the target to be identified, where the radar echo data includes radar scattering cross-section data and a high-resolution one-dimensional range image obtained by detection on a single platform;

[0052] Step S110, preprocessing the radar echo data, dividing the preprocessed radar echo data by Monte Carlo sampling, and then encoding the subsequences obtained after the division to obtain a coded sequence;

[0053] In step S120, the coding sequence is input into the trained target recognition network. In the target recognition network, a cross-domain fusion feature extractor based on the self-attention mechanism extracts the fusion features of the radar scattering cross-section data and the high-resolution one-dimensional range image according to the coding sequence. The classification head then classifies the target according to the fusion features to obtain the target recognition result.

[0054] In this method, we first construct a cross-domain data fusion model adapted to the RCS (Radar Cross Section) and HRRP (High Range Resolution Profile) data fusion tasks based on the self-attention mechanism commonly used in natural language processing problems. This is the cross-domain fusion feature extractor based on the self-attention mechanism in the target recognition model mentioned above. Then, in order to enable the cross-domain data fusion model to better extract cross-modal fusion features in the future, the radar echo data input to the model is preprocessed. At the same time, in order to fully utilize the local and global correlations in radar echo data under different modes, a coding method based on uniform sampling is also designed in this method to block and encode the preprocessed RCS and HRRP data. Finally, based on the task of context prediction commonly used in natural language processing problems, the task goal of matching the target categories of RCS and HRRP data is proposed to achieve a unified representation of RCS and HRRP data on a single platform. Specifically, Figure 2 shown.

[0055] In step S100 , the radar echo data includes radar cross-section data and a high-resolution one-dimensional range profile obtained by detecting the same target using the same radar (single platform).

[0056] In step S110, when pre-processing the radar echo data, normalization processing and center-of-gravity alignment processing are sequentially performed on the radar echo data to obtain pre-processed radar echo data.

[0057] Furthermore, the radar echo data is subjected to normalization and alignment processing to obtain preprocessed radar echo data, including: before preprocessing the radar cross-section data in the radar echo data, amplitude processing is performed on the radar echo data to obtain real data of the radar cross-section data, and subsequent preprocessing is performed based on the real data. Then, the radar echo data is processed using modulo-2 norm normalization to obtain amplitude-normalized radar echo data, and the amplitude-normalized radar echo data is subjected to a circular shift operation using a center alignment method so that its center is located at the center of the range window, thereby obtaining preprocessed radar echo data.

[0058] In order to effectively utilize the global and local information of RCS and HRRP data, in this embodiment, a data encoding method based on uniform sampling is proposed for radar echo data. However, before data encoding, the radar echo data needs to be preprocessed.

[0059] Radar echo data has problems of azimuth sensitivity, amplitude sensitivity, and translation sensitivity, which will have a great impact on the model's extraction of robust target features. Therefore, it is necessary to perform data preprocessing by normalizing and aligning the radar echo data (RCS and HRRP data) to reduce the impact of echo data sensitivity.

[0060] Specifically, let HRRP data be r=[r1,r2,…,r L ] T Indicates that r i represents the sub-echo of the ith range unit, L represents the total number of range units, and the RCS data is y=[y1,y2,…,y L ] T express.

[0061] Since the data type of HRRP data is in complex form, the input HRRP data is processed by amplitude processing to obtain real HRRP data:

[0062] x=|r|=[x1,x2,…,x L ] T . (1.)

[0063] In order to reduce the sensitivity of radar echo data to amplitude, in this embodiment, modulo 2 norm normalization is adopted to process the real radar echo data x, y to obtain amplitude-normalized HRRP data sample x norm , and the amplitude-normalized RCS data sample y norm , and their mathematical expressions are:

[0064]

[0065]

[0066] Furthermore, in order to reduce the translation sensitivity of radar echo data, in this embodiment, the center of gravity alignment method is used to align the x norm with y norm Perform a circular shift operation so that its center is located at the center of the range window.

[0067] Data sample x norm with y norm The center of gravity G x With G y They can be calculated using the following formulas:

[0068]

[0069]

[0070] In formula (4) and formula (5), Represents x norm with y norm Therefore, the radar echo data after centroid alignment can be obtained, that is, the radar echo data after preprocessing, which are expressed as x aligned with y aligned .

[0071] Next, in order to enable the subsequent feature extractor based on the self-attention mechanism to extract the features of the local substructure of the target, it is necessary to segment and map the radar echo data after data preprocessing and add position encoding information. The specific process can be referred to Figure 3 .

[0072] In this embodiment, after dividing the preprocessed radar echo data by Monte Carlo sampling, encoding the subsequences obtained after the division to obtain a coded sequence includes: dividing the preprocessed radar echo data according to a preset length to obtain multiple subsequences with the same length, encoding each subsequence into subsequence coded data represented by a D-dimensional vector, adding a category hint vector to the front end of each subsequence coded data and a segmentation vector to the back end, and constructing the coded sequence according to the subsequence coded data, the data sequence segmentation coding, and the time sequence position coding.

[0073] Specifically, let P be the preset length of each radar subsequence, and N = L / P be the number of divided radar subsequences. Therefore, the radar echo data x after centroid alignment is preprocessed. aligned with y aligned Can be divided into a combination of multiple subsequences of equal length: and where x′ align and y′ align It can be specifically expressed as:

[0074]

[0075]

[0076] In formula (15) and formula (16),

[0077] The divided subsequences can be regarded as the input radar echo sequence divided by Monte Carlo sampling, that is:

[0078]

[0079]

[0080] Then, for each subsequence after division and Perform input encoding and encode it into a D-dimensional vector to obtain the corresponding x MC,emb with y MC.emb ,Right now:

[0081]

[0082]

[0083] In formula (10) and formula (11), represents the encoder, represents the subsequence encoding, represents the encoded HRRP sample sequence, represents the subsequence encoding, Represents the encoded RCS sample sequence.

[0084] In this embodiment, a linear fully connected layer is used to encode the input sequence (multiple subsequences), and all subsequences share one input encoder. At the same time, in order to aggregate the information on each subsequence in the radar echo data without bias, a category hint vector is added before each subsequence. Implement the aggregation of radar echo data features in the deep attention feature fusion module. The category hint vector here is used to distinguish whether each subsequence belongs to RCS or HRRP. A segmentation vector is also added after each subsequence. Used to distinguish different input sequences and represent their temporal relationships.

[0085] Furthermore, the data sequence segment code E is added to the subsequence set processed above. seq, providing data sequence segmentation information for the model. Secondly, in order to make full use of the spatial position relationship of the subsequences in the radar echo data, the category prompt vector, segmentation vector, x MC,emp with y MC,emb Each radar sequence subsequence in adds a learnable temporal position code E pos ,Right now:

[0086] z in =[z cls ;x MS,emb ;z sep ;y MC.emb ;z sep ]+E seq +E pos , (12)

[0087] In formula (12), Represent the sequence segmentation encoding matrix and position encoding matrix respectively, z in It represents the output data of radar echo data after preprocessing, subsequence division and encoding, and also serves as the input of the target recognition model.

[0088] In step S120, the target recognition model includes a cross-domain fusion feature extractor based on a self-attention mechanism for extracting fusion features, and a classification head for identifying targets based on the fusion features.

[0089] Since the self-attention mechanism has excellent long-range dependency representation capabilities, it can capture local and global correlations between subsequences. In this embodiment, a cross-domain fusion feature extractor based on the self-attention mechanism is proposed for the unified representation of RCS data and HRRP data.

[0090] In this embodiment, the cross-domain fusion feature extractor based on the self-attention mechanism includes multiple layers of feature extraction layers connected in sequence. The encoding sequence is input into the first feature extraction layer to obtain the output data of the current layer, and then the output data of the current layer is used as the input of the next feature extraction layer, and so on, and the output data of the last feature extraction layer is used as the fusion feature.

[0091] Specifically, such as Figure 4 As shown in the figure, each feature extraction layer includes a first normalization layer, a multi-head attention layer, a second normalization layer, and a feedforward layer connected in sequence.

[0092] Furthermore, after the output data of the coding sequence or the previous feature extraction layer is input into the feature extraction layer: the first normalization layer uses the quasi-norm to perform layer normalization on the output data of the coding sequence or the previous feature extraction layer to obtain the first output data; in the multi-head attention layer, the first output data is linearly projected three times to obtain the query matrix, the key matrix and the value matrix respectively; these three matrices are projected into multiple feature subspaces respectively, and the attention results calculated in each subspace are spliced ​​to obtain the second output data; after the second output data and the coding sequence are input into the second normalization layer, the third output data is obtained; and then the third output data is processed by the feedforward layer through two layers of fully connected multilayer perceptrons to obtain the fourth output data; finally, the current layer feature extraction layer or fusion feature is obtained according to the fourth output data and the second output data.

[0093] Specifically, the first normalization layer can prevent problems such as gradient vanishing and gradient exploding when the network is deep. Specifically, this algorithm uses a quasi-norm input sequence for layer normalization, which can be expressed as:

[0094]

[0095] In formula (13), The normalized output of the representation layer is the first output data or the output data of the previous feature extraction layer, z in Represents the input sequence data or, LN represents the layer normalization operation.

[0096] The multi-head self-attention layer can capture the local and global relationship of the input sequence, and can enhance the model's ability to understand the fluctuations of RCS and HRRP echoes and the structural correlation between subsequences.

[0097] Specifically, the multi-head self-attention layer first outputs the first data Perform three linear projections respectively to obtain d q The query matrix Q of dimension d k The bond matrix K and d v The value matrix V of dimension is:

[0098]

[0099]

[0100]

[0101] In formulas (14) to (16),

[0102] Then the three matrices Q, K, and V are obtained through the self-attention mechanism based on cosine similarity measurement and the scaling factor The softmax function is used to calculate the attention weights between each subsequence, and the value matrix V is reweighted to obtain the output of the attention layer:

[0103]

[0104] In order to enable the feature extractor to learn more projection modes and better adapt to the feature modes of different modal data, the multi-head self-attention layer projects the query matrix Q, key matrix K and value matrix V into multiple feature subspaces respectively, and splices the attention results calculated in each subspace (the calculation result of formula (17)) to obtain the final multi-head attention layer output, which is the second output data, namely:

[0105] f MSA =Concat(head1,head2,...,head M )W O , (18)

[0106] In formula (18), M represents the number of attention heads, Concat represents the concatenation operation, Represents the recovery matrix, which is used to ensure the consistency of input and output dimensions.

[0107] Afterwards, the first output data and the second output data are output to the second normalization layer to obtain the third output data, and then the third output data is input to the feedforward layer.

[0108] The feedforward layer includes a multi-layer perceptron, which can map the input features to a high-dimensional latent space, and then map the high-dimensional latent space back to the original space to extract and filter the input features.

[0109] Specifically, the specific process of the feedforward layer can be defined as:

[0110]

[0111] In formula (18), the third output data is MLP(·) represents a multilayer perceptron with two fully connected layers.

[0112] The feature extraction layer can be expressed as:

[0113] T(z in )=f FFN +f MSA (19)

[0114] In formula (19), T(·) represents a feature extraction layer based on the self-attention mechanism.

[0115] In this embodiment, the cross-domain fusion feature extractor based on the self-attention mechanism includes multiple feature extraction layers, that is, the output of the final model can be expressed as:

[0116] T n (z in )=T(T(...T(z in ))) (20)

[0117] In formula (20), n represents the number of feature extraction layers contained in the model.

[0118] In this embodiment, when training the target recognition network, a loss function constructed based on the cross entropy formula is adopted.

[0119] Specifically, when training the target recognition network, RCS and HRRP sequences of random target category combinations are input, and then they are extracted and fused by a feature extractor based on the self-attention mechanism. Finally, the fused features are input into the classification head to obtain the final prediction of whether RCS and HRRP belong to the same target category.

[0120] Assume that the target category of the RCS sequence is l RCS , the target category of HRRP sequence is l HRRP , then in each batch, the sequence labels are set as:

[0121]

[0122] In formula (21), m represents the total number of training sequences. Then, the loss function of the algorithm is constructed based on the cross entropy formula, namely:

[0123]

[0124] In formula (22), C(·) represents the classification head, which outputs the probability of classification prediction. This loss is used to iteratively optimize the learnable parameters of the target recognition network, learning the global and local structural features between RCS and HRRP data, and achieving matching alignment between RCS and HRRP sequences.

[0125] Furthermore, the training steps of the target recognition network are as follows: first, the RCS and HRRP signals are encoded into radar echo data based on Monte Carlo sampling to obtain the corresponding encoding vector Secondly, the obtained encoding vector is input into the cross-domain fusion feature extractor based on the self-attention mechanism to obtain the final fusion feature vector Finally, the obtained fusion feature vector is input into the classification head of the downstream task to calculate the final loss function And use this to perform gradient updates on the learnable parameters in the overall algorithm.

[0126] When testing the trained object recognition network, the trained cross-domain fusion feature extractor based on the self-attention mechanism is first connected to the recognition and classification head of the downstream task and the parameters are adjusted (fixing the fusion model parameters) on the downstream dataset. After the parameter adjustment is completed, the parameters of the cross-domain fusion feature extractor based on the self-attention mechanism and the classification head are fixed, and the recognition test is performed on the downstream dataset.

[0127] In this paper, we also conduct experimental simulation based on the method in this paper, and the results are as follows: Figure 5 As shown in Table 1. Figure 5 As shown in the figure, (a) represents the RCS data feature, (b) represents the HRRP data feature, (c) represents the fused data feature, and (d) represents the grayscale corresponding to each target. Compared with the original RCS and HRRP data, the fused feature vector obtained by the cross-domain data fusion model has good separability, proving that the present invention can obtain well-characterized radar signal fusion features. As shown in Table 1, the radar signal fusion features obtained by this method can effectively improve the accuracy of downstream recognition tasks, increasing the accuracy by approximately 12% compared to the original FCN method.

[0128] Table 1 Recognition accuracy of downstream classification tasks

[0129]

[0130]

[0131] In the above-mentioned target recognition method for single-platform radar multimodal data fusion, a cross-domain data fusion model adapted to the RCS and HRRP data fusion tasks is first constructed based on the self-attention mechanism commonly used in natural language processing problems. Afterwards, the radar echo data is preprocessed using the normalization and center of gravity alignment algorithm to eliminate the sensitivity of the echo data to azimuth, amplitude and translation. Then, in order to fully utilize the local and global correlations in radar echo data under different modes, the present invention designs a coding method based on uniform sampling to perform block and encoding operations on the input RCS and HRRP data. Finally, based on the task of context prediction commonly used in natural language processing problems, the task goal of matching the target categories of RCS and HRRP data is proposed to achieve a unified characterization result of RCS and HRRP data on a single platform. This method can realize the feature-level fusion of RCS and HRRP signals, and the obtained fusion features have good characterization. At the same time, this method can improve the performance of the classification and recognition algorithm.

[0132] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0133] In one embodiment, Figure 6 As shown, a target recognition device for single-platform radar multi-modal data fusion is provided, comprising: a radar echo data acquisition module 200, a data preprocessing module 210 and a target recognition module 220, wherein:

[0134] A radar echo data acquisition module 200 is used to acquire radar echo data of a target to be identified, wherein the radar echo data includes radar scattering cross-section data and a high-resolution one-dimensional range image obtained by detection by a single platform;

[0135] The data preprocessing module 210 is used to preprocess the radar echo data, divide the preprocessed radar echo data into subsequences by Monte Carlo sampling, and then encode the subsequences obtained after the division to obtain a coded sequence;

[0136] The target recognition module 220 is used to input the coding sequence into a trained target recognition network. In the target recognition network, a cross-domain fusion feature extractor based on the self-attention mechanism extracts the fusion features of the radar scattering cross-section data and the high-resolution one-dimensional range image according to the coding sequence, and then the classification head classifies the target according to the fusion features to obtain the target recognition result.

[0137] Regarding the specific limitations of the target recognition device for single-platform radar multimodal data fusion, please refer to the limitations of the target recognition method for single-platform radar multimodal data fusion above and will not be repeated here. The various modules in the above-mentioned target recognition device for single-platform radar multimodal data fusion can be implemented in whole or in part through software, hardware, or a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of the above-mentioned modules.

[0138] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown. The computer device includes a processor, memory, network interface, display screen and input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a target recognition method for multi-modal data fusion of a single-platform radar is implemented. The display screen of the computer device can be a liquid crystal display or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a key, trackball or touchpad provided on the computer device housing, or an external keyboard, touchpad or mouse.

[0139] Those skilled in the art will understand that Figure 7 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0140] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0141] Acquire radar echo data of the target to be identified, wherein the radar echo data is radar scattering cross-section data and high-resolution one-dimensional range image obtained by single-platform detection;

[0142] Preprocessing the radar echo data, dividing the preprocessed radar echo data by Monte Carlo sampling, and then encoding the subsequences obtained after the division to obtain a coded sequence;

[0143] The coding sequence is input into a trained target recognition network. In the target recognition network, a cross-domain fusion feature extractor based on the self-attention mechanism extracts the fusion features of the radar scattering cross-section data and the high-resolution one-dimensional range image according to the coding sequence. The classification head then classifies the target according to the fusion features to obtain the target recognition result.

[0144] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0145] Acquire radar echo data of the target to be identified, wherein the radar echo data is radar scattering cross-section data and high-resolution one-dimensional range image obtained by single-platform detection;

[0146] Preprocessing the radar echo data, dividing the preprocessed radar echo data by Monte Carlo sampling, and then encoding the subsequences obtained after the division to obtain a coded sequence;

[0147] The coding sequence is input into a trained target recognition network. In the target recognition network, a cross-domain fusion feature extractor based on the self-attention mechanism extracts the fusion features of the radar scattering cross-section data and the high-resolution one-dimensional range image according to the coding sequence. The classification head then classifies the target according to the fusion features to obtain the target recognition result.

[0148] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0149] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0150] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A target recognition method for multimodal data fusion of a single-platform radar, characterized in that: The method comprises: Acquire radar echo data of the target to be identified, wherein the radar echo data is radar scattering cross-section data and high-resolution one-dimensional range image obtained by single-platform detection; The radar echo data is preprocessed, the preprocessed radar echo data is divided by Monte Carlo sampling, and then the subsequences obtained after the division are encoded to obtain a coded sequence. When the division is performed by Monte Carlo sampling, the following formula is used: In the above formula, , P Indicates the preset length of each radar subsequence; The coding sequence is input into a trained target recognition network. In the target recognition network, a cross-domain fusion feature extractor based on a self-attention mechanism extracts the fusion features of the radar scattering cross-section data and the high-resolution one-dimensional range image according to the coding sequence, and then a classification head classifies the target according to the fusion features to obtain the target recognition result. The cross-domain fusion feature extractor based on the self-attention mechanism includes multiple layers of feature extraction layers connected in sequence. The coding sequence is input into the first feature extraction layer to obtain the output data of the current layer, and the output data of the current layer is used as the input of the next feature extraction layer. And so on, the output data of the last feature extraction layer is used as the fusion feature. Each feature extraction layer includes a first normalization layer, a multi-head attention layer, a second normalization layer and a feedforward layer connected in sequence.

2. The target recognition method according to claim 1, characterized in that: When the radar echo data is preprocessed, normalization processing and center of gravity alignment processing are sequentially performed on the radar echo data to obtain preprocessed radar echo data.

3. The target recognition method according to claim 2, characterized in that: The radar echo data is subjected to normalization processing and alignment processing in sequence to obtain pre-processed radar echo data, including: Before preprocessing the radar scattering cross section data in the radar echo data, amplitude processing is first performed on the radar scattering cross section data to obtain real part data in the radar scattering cross section data, and subsequent preprocessing is performed based on the real part data; Processing the radar echo data using modulo-2 norm normalization to obtain amplitude-normalized radar echo data; The center alignment method is used to perform a circular shift operation on the amplitude-normalized radar echo data so that its center is located at the center of the range window, and the preprocessed radar echo data is obtained.

4. The target recognition method according to claim 1, characterized in that: The step of dividing the pre-processed radar echo data by Monte Carlo sampling and then encoding the subsequences obtained after the division to obtain the coded sequence comprises: Divide the pre-processed radar echo data according to a preset length to obtain multiple subsequences with the same length; Each subsequence is encoded as dimensional vector, and adding a category hint vector to the front end of each subsequence coded data and a segmentation vector to the back end; The coding sequence is constructed based on the subsequence coding data, data sequence segment coding and time sequence position coding.

5. The target recognition method according to claim 4, characterized in that: After the encoding sequence or the output data of the previous feature extraction layer is input into the feature extraction layer: The first normalization layer uses a quasi-norm to perform layer normalization on the coding sequence or the output data of the previous feature extraction layer to obtain first output data; In the multi-head attention layer, the first output data is linearly projected three times to obtain a query matrix, a key matrix, and a value matrix, respectively. These three matrices are projected into multiple feature subspaces, and the attention results calculated in each subspace are concatenated to obtain the second output data. After inputting the second output data and the encoding sequence into the second normalization layer, third output data is obtained; The feedforward layer processes the third output data through two layers of fully connected multilayer perceptrons to obtain fourth output data; The current layer feature extraction layer or the fusion feature is obtained according to the fourth output data and the second output data.

6. The target recognition method according to claim 5, characterized in that: When training the target recognition network, a loss function constructed based on the cross entropy formula is adopted.

7. A target recognition device for multi-modal data fusion of a single-platform radar, characterized in that: The device comprises: A radar echo data acquisition module is used to acquire radar echo data of the target to be identified. The radar echo data includes radar scattering cross-section data and high-resolution one-dimensional range image obtained by single-platform detection; The data preprocessing module is used to preprocess the radar echo data, divide the preprocessed radar echo data by Monte Carlo sampling, and then encode the subsequences obtained after the division to obtain a coded sequence. When dividing by Monte Carlo sampling, the following formula is used: In the above formula, , P Indicates the preset length of each radar subsequence; A target recognition module is used to input the coding sequence into a trained target recognition network. In the target recognition network, a cross-domain fusion feature extractor based on a self-attention mechanism extracts the fusion features of the radar scattering cross-section data and the high-resolution one-dimensional range image according to the coding sequence, and then a classification head classifies the target according to the fusion features to obtain the target recognition result. The cross-domain fusion feature extractor based on the self-attention mechanism includes multiple layers of feature extraction layers connected in sequence, the coding sequence is input into the first feature extraction layer to obtain the output data of the current layer, and the output data of the current layer is used as the input of the next feature extraction layer, and so on, the output data of the last feature extraction layer is used as the fusion feature, and each feature extraction layer includes a first normalization layer, a multi-head attention layer, a second normalization layer and a feedforward layer connected in sequence.

Images