Radar clutter map target detection method and device based on double-branch cross-attention neural network, equipment and medium

Abstract

The application discloses a radar clutter map target detection method and device based on a double-branch cross-attention neural network, equipment and a medium, comprising: preprocessing baseband echo data; obtaining true value information of a cooperative target, and converting and timestamp aligning the true value information; dividing a detection space into clutter units, recursively updating statistical characteristics of each clutter unit without a target, and constructing a historical frame statistical characteristic sequence matrix corresponding to each clutter unit; based on the historical frame statistical characteristic sequence matrix and corresponding current frame amplitudes, constructing a data set, and performing data enhancement and data set division on the data set; constructing a double-branch cross-attention deep learning network, and training the double-branch cross-attention deep learning network based on the data set. The application belongs to the field of radar signal processing. The application can improve the accuracy of radar target detection.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing, and in particular to a radar clutter map target detection method, apparatus, device, and medium based on a dual-branch cross-attention neural network. Background Technology

[0002] Radar target detection faces significant challenges in complex clutter environments. Traditional clutter map methods divide the space into multiple clutter cells and recursively update the mean and variance of each cell, setting a detection threshold based on statistical measures to achieve constant false alarm rate (CFAR) detection. However, these methods rely solely on scalar statistical features, making it difficult to fully utilize multi-frame historical information. Furthermore, their threshold calculations are typically based on pre-defined distribution assumptions such as Gaussian distributions, which are ill-suited to the prevalent non-stationary clutter environments in reality, resulting in insufficient overall adaptability and limited detection performance.

[0003] With the development of deep learning, CNNs, RNNs, and attention mechanisms have demonstrated strong capabilities in feature extraction and modeling. However, their direct application to radar target detection still has limitations: CNNs are limited by their local receptive field, making it difficult to model long-range dependencies; time-series models such as LSTM have low computational efficiency and limited ability to characterize long-term dependencies. Therefore, how to improve the accuracy of radar target detection based on deep learning is an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a radar clutter map target detection method, apparatus, device, and medium based on a dual-branch cross-attention neural network, solving the technical problem of how to improve the accuracy of radar target detection based on deep learning in the prior art, and achieving the technical effect of improving the accuracy of radar target detection based on deep learning.

[0005] In a first aspect, the present invention provides a radar clutter map target detection method based on a dual-branch cross-attention neural network, comprising: The baseband echo data and corresponding radar operating parameters in the target scene are acquired, and the baseband echo data is preprocessed. The preprocessing includes pulse compression and Doppler processing. The radar operating parameters include beam pointing, pulse repetition period, dwell time and corresponding timestamp. Obtain the truth information of the cooperation target, and transform and align the truth information with the timestamp; The probe airspace is divided into clutter units. The statistical features of each clutter unit without targets are recursively updated, and a historical frame statistical feature sequence matrix corresponding to each clutter unit is constructed. A dataset is constructed based on the historical frame statistical feature sequence matrix and the corresponding current frame amplitude, and the dataset is augmented and partitioned. A dual-branch cross-attention deep learning network was constructed and trained on a dataset. The dual-branch cross-attention deep learning network was then used for object detection.

[0006] Furthermore, it also includes: In the validation set of the dataset, the target probability output distribution of clutter samples is statistically analyzed, and the corresponding quantile value is determined according to the preset false alarm probability. Determine the detection threshold factor; The performance of the trained bi-branch cross-attention deep learning network is evaluated based on the unused test set partitioned by the dataset and the detection threshold factor.

[0007] Furthermore, the baseband echo data undergoes preprocessing, including: Pulse compression is performed on the baseband echo data to obtain the corresponding range profile; Slow-time Doppler processing is applied to the range image to obtain the range and Doppler matrices.

[0008] Furthermore, the probed airspace is divided into clutter units, and the statistical characteristics of each clutter unit without targets are recursively updated. A historical frame statistical characteristic sequence matrix corresponding to each clutter unit is then constructed, including: The detection airspace is divided into clutter units according to preset dimensions, including range dimension, azimuth dimension, elevation dimension and Doppler dimension; For each clutter unit, a corresponding historical frame statistical feature sequence is constructed. Each frame in the sequence contains several statistical features, forming a historical frame statistical feature sequence matrix, including:

[0009] in, For the length of historical frames, For the first Statistical eigenvectors of clutter cells at frame time. For frame index, For transpose, The mean of the amplitude. The standard deviation of the amplitude. For local contrast; When any clutter unit does not contain a target in any frame, the amplitude of that frame is used for recursive updating, resulting in an updated historical statistical sequence matrix. The update formula includes:

[0010]

[0011]

[0012] in, Forgetting factor, for The average amplitude of the frame. For the first The observed mean of the frames, for Frame variance This represents the number of amplitude observation points contained within the clutter cell. For the first The first frame The amplitude at each observation point; The differences between amplitude observations in clutter cells are characterized based on local contrast, including: .

[0013] Furthermore, a dual-branch cross-attention deep learning network is constructed, including: Based on the input layer simultaneously receiving historical statistical sequence matrices and the corresponding current frame amplitude; Based on the historical frame feature extraction network, temporal features are extracted from the historical frame statistical feature sequence matrix, and the temporal features are mapped to generate key vector sequence and value vector sequence. Based on the current frame feature extraction network, amplitude features are extracted to generate a query vector; Based on a cross-attention fusion network, similarity calculation and weighted fusion are performed on query vectors, key vector sequences, and value vector sequences; Residual connections are performed based on residual fusion networks to generate fused feature vectors; Based on the classification output network, the fused feature vector is mapped to the probability of the target and the probability of clutter.

[0014] Furthermore, a dual-branch cross-attention deep learning network is trained based on the dataset, including: The dual-branch cross-attention deep learning network is trained on the training set of the dataset, where the cross-entropy loss function is used during training, and the network parameters of the dual-branch cross-attention deep learning network are optimized through backpropagation. Real-time acquisition of loss and accuracy of dual-branch cross-attention deep learning networks; When the preset requirements are met, the dual-branch cross-attention deep learning network is used for object detection.

[0015] Furthermore, online detection is performed based on the trained dual-branch cross-attention deep learning network, including: Real-time baseband echo data is acquired and preprocessed to obtain the historical frame feature statistical sequence matrix and the corresponding current frame amplitude. By inputting the real-time statistical sequence matrix and the corresponding current frame amplitude into the trained dual-branch cross-attention deep learning network, the probability of the target corresponding to each clutter unit is obtained. Based on the probability of a target in a clutter cell and the detection threshold, determine whether a target exists in that clutter cell; The statistical characteristics of clutter cells for which no target was detected are recursively updated, and the clutter cells containing the target are sent to the target receiver.

[0016] Secondly, the present invention provides a radar clutter map target detection device based on a dual-branch cross-attention neural network, comprising: The echo acquisition module is used to acquire baseband echo data and corresponding radar operating parameters in the target scene, and to preprocess the baseband echo data. The preprocessing includes pulse compression and Doppler processing. The radar operating parameters include beam pointing, pulse repetition period, dwell time and corresponding timestamp. The truth acquisition module is used to obtain the truth information of the cooperation target and convert the truth information to align with the timestamp. The matrix construction module is used to divide the probe airspace into clutter units, recursively update the statistical features of each clutter unit without targets, and construct the historical frame statistical feature sequence matrix corresponding to each clutter unit. The dataset module is used to construct a dataset based on the statistical feature sequence matrix of historical frames and the corresponding amplitude of the current frame, and to perform data augmentation and dataset partitioning on the dataset. The training module is used to construct a dual-branch cross-attention deep learning network and train the dual-branch cross-attention deep learning network based on the dataset. The dual-branch cross-attention deep learning network is used for object detection.

[0017] Thirdly, the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to execute a radar clutter map target detection method based on a dual-branch cross-attention neural network, as provided in the first aspect.

[0018] Fourthly, the present invention provides a non-transitory computer-readable storage medium that, when the instructions in the storage medium are executed by the processor of an electronic device, enables the electronic device to perform a radar clutter map target detection method based on a dual-branch cross-attention neural network as provided in the first aspect.

[0019] One or more technical solutions provided in this invention have at least the following technical effects or advantages: A radar clutter map target detection method based on a bi-branch cross-attention neural network includes: acquiring baseband echo data and corresponding radar operating parameters in the target scene, and preprocessing the baseband echo data, including pulse compression and Doppler processing; acquiring ground truth information of cooperative targets and aligning the ground truth information with the timestamps; dividing the detection airspace into clutter units, recursively updating the statistical features of each clutter unit without targets, and constructing a historical frame statistical feature sequence matrix for each clutter unit; constructing a dataset based on the historical frame statistical feature sequence matrix and the corresponding current frame amplitude, and performing data augmentation and dataset partitioning on the dataset; constructing a bi-branch cross-attention deep learning network, and training the bi-branch cross-attention deep learning network based on the dataset, which is then used for target detection. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the radar clutter map target detection method based on a dual-branch cross-attention neural network provided by the present invention. Figure 2 This is a schematic diagram of the dual-branch cross-attention network architecture provided by the present invention; Figure 3 A schematic diagram of the cross-attention fusion network architecture provided by this invention; Figure 4 The flowchart for the two-stage offline training and online detection provided by this invention; Figure 5 This is a diagram illustrating the traditional detection effect provided by the present invention; Figure 6 The image shows the detection results using the method provided by this invention. Detailed Implementation

[0022] This invention provides a radar clutter map target detection method based on a dual-branch cross-attention neural network, which solves the technical problem of how to improve the accuracy of radar target detection based on deep learning in the prior art.

[0023] The technical solution of this invention is to solve the above-mentioned technical problems, and the overall idea is as follows: A radar clutter map target detection method based on a bi-branch cross-attention neural network includes: acquiring baseband echo data and corresponding radar operating parameters in the target scene, and preprocessing the baseband echo data, including pulse compression and Doppler processing; acquiring ground truth information of cooperative targets and aligning the ground truth information with the timestamps; dividing the detection airspace into clutter units, recursively updating the statistical features of each clutter unit without targets, and constructing a historical frame statistical feature sequence matrix for each clutter unit; constructing a dataset based on the historical frame statistical feature sequence matrix and the corresponding current frame amplitude, and performing data augmentation and dataset partitioning on the dataset; constructing a bi-branch cross-attention deep learning network, and training the bi-branch cross-attention deep learning network based on the dataset, which is then used for target detection.

[0024] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0025] First, it should be clarified that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0026] This invention provides, for example Figure 1 The radar clutter map target detection method shown includes steps S11-S15: Step S11: Obtain baseband echo data and corresponding radar operating parameters in the target scene, and preprocess the baseband echo data. The preprocessing includes pulse compression and Doppler processing. The radar operating parameters include beam pointing, pulse repetition period, dwell time and corresponding timestamp.

[0027] Preprocessing of the baseband echo data includes: pulse compression of the baseband echo data to obtain the corresponding range profile; and slow-time Doppler processing of the range profile to obtain the range and Doppler matrices.

[0028] Specifically: In target application scenarios (such as airports, sea surfaces, or test ranges), pulse Doppler radar is used to continuously observe the environment and collect baseband echo data. To ensure the stability and sufficiency of subsequent processing, the acquisition time needs to cover multiple coherent processing intervals (CPI) to reflect the temporal statistical characteristics of clutter and targets.

[0029] During data acquisition, radar operating parameters need to be recorded simultaneously, including beam pointing (characterizing the direction of space observation), pulse repetition period (determining the ambiguity characteristics of range and velocity), dwell time (affecting signal-to-noise ratio and Doppler resolution), and precise timestamps, for subsequent multidimensional data alignment and time series modeling.

[0030] The acquired baseband echo data is processed using a standard signal processing procedure. First, the range resolution is improved by pulse compression (such as matched filtering), converting the original echo into a high-resolution range image. Subsequently, Doppler processing (such as MTD) is performed on each range cell in the slow time dimension to obtain the range-Doppler matrix (RDM) that reflects the radial velocity information of the target. Its dimension is Nr×Nd, where Nr represents the number of range cells and Nd represents the number of Doppler channels. By further combining beam scanning information, the data can be expanded into a four-dimensional structure of Nr×Nd×Na×Ne (Na and Ne represent azimuth and elevation dimensions, respectively), and the RDMs of multiple consecutive CPIs can be organized into a time-series dataset (RDM sequence) in chronological order.

[0031] Step S12: Obtain the truth information of the cooperation target and convert the truth information to align with the timestamp.

[0032] Specifically, in order to obtain high-precision target reference information that can be used for supervised learning, differential GPS or BeiDou high-precision positioning equipment is used to equip cooperative targets (referring to observed targets that can actively provide their own precise location or status information (such as through GPS, ADS-B, AIS, etc.)) to record their longitude, latitude, altitude and corresponding timestamps in real time to ensure the accuracy of spatial location and time information. By synchronously receiving ADS-B or AIS broadcast data, the motion state of the target is verified and supplemented. Combined with manual interpretation and screening of data from time periods when no target appears, this data is used as pure clutter samples for negative sample construction.

[0033] Based on this, the true value information (longitude, latitude, altitude) in the acquired geographic coordinate system is mapped to the radar coordinate system through coordinate transformation to obtain the corresponding slant range, azimuth angle and elevation angle, thereby achieving a consistent expression of the target's spatial location and the radar observation space. The true timestamp is precisely aligned with the timestamp of the range-Doppler matrix sequence generated in step S11 to determine the appearance time of the target in a specific range-Doppler matrix sequence (RDM) frame, and then further mapped to the corresponding range cell and Doppler cell.

[0034] Step S13: Divide the probe airspace into clutter units, recursively update the statistical features of each clutter unit that does not contain a target, and construct the historical frame statistical feature sequence matrix corresponding to each clutter unit.

[0035] The probed airspace is divided into clutter units. The statistical features of each clutter unit without targets are recursively updated, and a historical frame statistical feature sequence matrix corresponding to each clutter unit is constructed, including: The detection airspace is divided into clutter units according to preset dimensions, including range dimension, azimuth dimension, elevation dimension and Doppler dimension; For each clutter unit, a corresponding historical frame statistical feature sequence is constructed. Each frame in the sequence contains several statistical features, forming a historical frame statistical feature sequence matrix, including:

[0036] in, For the length of historical frames, For the first Statistical eigenvectors of clutter cells at frame time. For frame index, For transpose, The mean of the amplitude. The standard deviation of the amplitude. For local contrast; When any clutter unit does not contain a target in any frame, the amplitude of that frame is used for recursive updating, resulting in an updated historical statistical sequence matrix. The update formula includes:

[0037]

[0038]

[0039] in, Forgetting factor, for The average amplitude of the frame. For the first The observed mean of the frames, for Frame variance This represents the number of amplitude observation points contained within the clutter cell. For the first The first frame The amplitude at each observation point; The differences between amplitude observations in clutter cells are characterized based on local contrast, including: .

[0040] Specifically: The radar detection airspace is divided into grids according to four dimensions: range, azimuth, elevation, and Doppler. Each grid corresponds to a clutter unit, thereby realizing the modeling of clutter in the space-velocity domain. For each clutter unit, a historical statistical sequence of length L is constructed to characterize the statistical characteristics of the unit's evolution over time. Several statistical features are extracted from each frame and combined into a feature vector, which are stacked in chronological order to form an M×L historical statistical sequence matrix.

[0041] Combining the truth information obtained in step 12, it is determined whether each clutter unit contains a target in each frame. Only when the unit is marked as targetless in the current frame is its amplitude observation data used to participate in the recursive update of the statistics, thereby avoiding the contamination of clutter statistical characteristics by the target signal. Among them, the mean amplitude and variance are updated by an exponential recursion method that introduces a forgetting factor, so that the model can adapt to environmental changes while retaining historical information, and the observation mean is obtained by averaging the K amplitude observations in the unit. Meanwhile, local contrast is introduced to characterize the non-uniformity of amplitude distribution within the unit. It is calculated by the ratio of the maximum amplitude to the average amplitude, which can effectively reflect the peakness or dispersion of clutter.

[0042] Through the above process, a stable and continuous temporal statistical feature representation is established for each clutter unit, providing temporal input data for subsequent clutter modeling, anomaly detection, and target detection.

[0043] Step S14: Based on the historical frame statistical feature sequence matrix and the corresponding current frame amplitude, construct a dataset, and perform data augmentation and dataset partitioning on the dataset.

[0044] Based on the ground truth information, target and clutter discrimination is performed on each cell in the current frame range-Doppler matrix (RDM); For cells marked as "containing targets", the corresponding clutter map index is calculated based on the spatial-Doppler position of the cell. The historical statistical features of the most recent L frames are extracted from the clutter cell corresponding to the index to construct an M×L clutter map statistical sequence matrix H. At the same time, the amplitude value y of the cell in the current RDM is extracted as the instantaneous observation feature, and the sample is labeled as a positive class (label 1). For pure clutter cells labeled as "no target", the corresponding H and y are extracted in the same way and labeled as negative (label 0). Finally, all positive and negative samples are collected to construct a dataset.

[0045] To improve the model's adaptability and generalization performance to complex environmental changes, the original samples are enhanced in various ways, such as superimposing small random noise on the clutter plot statistics sequence H to simulate measurement disturbances and performing time dimension reversal to enhance temporal robustness. Simultaneously, the amplitude value y of the current frame is randomly scaled or perturbed to simulate the observation changes under different signal-to-noise ratio conditions, and the enhanced samples are merged with the original samples to expand the data scale.

[0046] After completing the sample construction and enhancement, the overall dataset is randomly divided into training set, validation set and test set according to a preset ratio. The training set is used for model parameter learning, the validation set is used for model structure optimization, hyperparameter tuning and detection threshold selection, and the test set is used to objectively evaluate the final detection performance of the model on independent data.

[0047] Step S15: Construct a dual-branch cross-attention deep learning network and train it based on the dataset. The dual-branch cross-attention deep learning network is used for object detection.

[0048] A schematic diagram of the structure of a two-branch cross-attention deep learning network is shown below. Figure 2 As shown.

[0049] Constructing a dual-branch cross-attention deep learning network, including: The input layer simultaneously receives historical statistical sequence matrices and the corresponding current frame amplitude.

[0050] The input layer receives two types of data: one is the clutter graph statistics sequence: the history stored in each clutter cell. L Frame statistics, each frame contains M Features (such as mean, standard deviation, local contrast) constitute M × L The matrix is ​​used as input to the historical feature extraction network; Second, the current frame RDM amplitude: the amplitude value (scalar) of the unit in the current RDM frame, which is used as the input of the current feature extraction network.

[0051] Based on the historical frame feature extraction network, the temporal features in the historical frame statistical feature sequence matrix are extracted, and the temporal features are mapped to generate key vector sequences and value vector sequences.

[0052] The historical feature extraction network extracts temporal features from the clutter graph statistics sequence and generates keys and values ​​for attention. Its internal structure is as follows: A hybrid CNN+LSTM network is used to process the input. M × L Feature extraction is performed on the sequence.

[0053] CNN (Convolutional Neural Network) is used to capture local features of clutter maps, while LSTM is used to obtain temporal feature sequences containing long-term and short-term dependencies.

[0054] Key projection layer: Maps the temporal feature sequence output by the CNN+LSTM hybrid network into a key vector sequence.

[0055] Value projection layer: maps the temporal feature sequence output by the CNN+LSTM hybrid network into a value vector sequence.

[0056] Based on the current frame feature extraction network, amplitude features are extracted to generate a query vector.

[0057] The current feature extraction network is used to extract features from the current RDM magnitude and generate a query vector for attention, with the following structure: Amplitude feature extraction network: This network processes the input scalar amplitude value to obtain an amplitude feature vector. f .

[0058] Query Feature Extraction Network: The input scalar magnitude value is passed through a fully connected network to obtain the query feature vector.

[0059] Query projection layer: maps the sequence output by the query feature extraction network to a query vector sequence.

[0060] Based on a cross-attention fusion network, similarity calculation and weighted fusion are performed on query vectors, key vector sequences, and value vector sequences.

[0061] The cross-attention fusion network calculates attention weights using query vectors and key-value sequences, and performs a weighted summation of the value sequences to extract the historical information most relevant to the current observation. Its internal structure is as follows: Cross-attention layer: Calculates the similarity between the query vector and the key vector, performs a weighted summation of the value vectors, and generates attention features. The calculation method is as follows: Figure 3 .

[0062] Feature-weighted pooling layer: utilizing attention weights The context vector Z is obtained by performing a weighted summation on the value sequence v.

[0063] When the current observation comes from the target, its amplitude or statistical characteristics often deviate significantly from the historical clutter characteristics, resulting in a low similarity between the query vector and the key vector of most historical frames, and the attention score is generally small and evenly distributed. When the current observation comes from clutter, the similarity with the historical key vector is high and stable, the attention weight distribution tends to be uniform, and the context vector approximates the historical mean, thus suppressing random fluctuations.

[0064] Based on the differential perception mechanism of attention weights, the network can adaptively amplify the statistical anomalies caused by the target, while smoothing the inherent fluctuations of clutter, thus achieving accurate clutter and target discrimination.

[0065] Residual connections are performed based on residual fusion networks to generate fused feature vectors.

[0066] The residual fusion network fuses the context vector output by the attention function with the magnitude feature of the current branch, enhancing its feature representation capability. Its internal structure is as follows: Residual connection layer: connects context vectors Z With amplitude eigenvector f Residual connections are performed, thereby preserving the original amplitude information while introducing historical information from attention filtering.

[0067] Feature fusion network: The residual connected features are deeply fused through a fully connected network to finally output a fused feature vector.

[0068] Based on the classification output network, the fused feature vector is mapped to the probability of the target and the probability of clutter.

[0069] The classification output network maps the fused features to probabilities of the target and clutter, and its internal structure is as follows: Classification network: The fused feature vectors are mapped to 2-dimensional output classes (corresponding to the target class and clutter class) through a fully connected network.

[0070] Softmax classification output: Apply softmax to the 2D output class to obtain the probability p of the target's presence and the clutter probability 1-p, which are used as the final output of the network.

[0071] Training a dual-branch cross-attention deep learning network based on a dataset includes: training the dual-branch cross-attention deep learning network on the training set of the dataset, wherein the cross-entropy loss function is used during training, and the network parameters of the dual-branch cross-attention deep learning network are optimized through backpropagation; the loss and accuracy of the dual-branch cross-attention deep learning network are obtained in real time; and when the preset requirements are met, the dual-branch cross-attention deep learning network is used for object detection.

[0072] It also includes: statistically analyzing the target probability output distribution of clutter samples in the validation set of the dataset partitioning, and determining the corresponding quantile value according to the preset false alarm probability; determining the detection threshold factor; and evaluating the performance of the trained dual-branch cross-attention deep learning network based on the unused test set of the dataset partitioning and the detection threshold factor.

[0073] After training, threshold calibration is performed using all clutter samples (labeled 0) in the validation set. The specific steps are as follows: Input the historical sequence H and the current RDM amplitude y of each clutter sample into the network to obtain the target probability p; Collect the target probabilities of all clutter samples to obtain the set. ,in To verify the total number of concentrated clutter samples.

[0074] Arrange the probabilities in descending order: .

[0075] Based on the preset false alarm probability Calculate the threshold index .in, This is for rounding down.

[0076] The probability value of the k-th clutter sample is taken as the detection threshold factor: .

[0077] The final model was evaluated using a test set that was not involved in training and threshold determination: For each sample in the test set (containing both the target and clutter), calculate the target probability p; The detection results are obtained by applying the detection threshold factor T: ,otherwise ; Statistical detection rate and false alarm probability And compare it with the design value.

[0078] Plot receiver operating characteristic (ROC) curves and calculate the area under the curve (AUC) to evaluate the overall performance of the model.

[0079] Online detection is performed based on a trained bi-branch cross-attention deep learning network, including: acquiring real-time baseband echo data and preprocessing it to obtain a historical frame feature statistical sequence matrix and the corresponding current frame amplitude; inputting the real-time statistical sequence matrix and the corresponding current frame amplitude into the trained bi-branch cross-attention deep learning network to obtain the probability of the target corresponding to each clutter unit; determining whether the clutter unit contains a target based on the target probability and the detection threshold; recursively updating the statistical features of clutter units that do not detect targets, and sending the traces of clutter units containing targets to the target receiver.

[0080] The online detection phase involves deploying the trained model and detection threshold factor into the real-time radar signal processing system to detect continuously input radar echoes. Specifically, it includes the following steps: The radar system acquires baseband echo data in real time, forming a data stream.

[0081] The real-time echo undergoes the same preprocessing steps as the offline stage, including pulse compression and Doppler processing, to generate the RDM for the current CPI. This RDM has the same dimensional structure as the RDM generated in the offline stage.

[0082] For each clutter cell corresponding to the current RDM, the stored historical statistics of L frames are read from the clutter map and formed into an M×L matrix H.

[0083] The extracted historical statistical sequence H and the current unit amplitude y are combined to form the input format required by the network. Specifically, H is directly used as the input to the time-series branch, and y is used as the scalar input to the current branch.

[0084] The combined input is fed into the deployed neural network for forward computation to obtain the target probability p of the unit.

[0085] Compare the target probability p with the detection threshold T determined in the offline phase: like If the target is detected, the cell is determined to contain a target and the target location (including distance, azimuth, elevation, Doppler information and probability value) is output. like If the signal is not clear, it is considered noise and no output is made.

[0086] After detecting all units in the current frame, the historical statistical sequence of each unit in the clutter map needs to be updated. For units where a target was detected, no update is selected to avoid the target contaminating the historical statistics. For units where no target was detected, the mean and standard deviation are updated using a recursive formula: Then shift the old sequence left by one position, discard the oldest frame, and use the newly calculated... Save the latest frame.

[0087] The system can send the detected points to the radar terminal for display or to the subsequent tracking module in real time, and can also record the detection results for performance monitoring.

[0088] Figure 4 The flowchart for the two-stage offline training and online detection provided by this invention. Figure 5 This is a diagram showing the traditional detection effect of the present invention. Figure 6 The image shows the detection results using the method provided by this invention.

[0089] In summary, this invention addresses the performance limitations of traditional clutter pattern detection methods in low signal-to-clutter ratio and non-stationary clutter environments by proposing a target detection method based on a dual-branch cross-attention mechanism, achieving a significant improvement in detection performance and robustness. This invention constructs a dual-branch structure consisting of a historical statistical sequence branch and a current amplitude branch, extracting the temporal evolution features of clutter and the instantaneous features of the current observation, respectively. The cross-attention mechanism is then used to deeply correlate and model the differences between these two aspects, enabling adaptive focusing on anomalous changes inconsistent with historical clutter patterns, thus achieving accurate target detection. Under the same false alarm probability conditions, the detection rate and accuracy for weak targets and targets in complex backgrounds are higher, and false alarm control is more stable.

[0090] This invention constructs multi-frame historical statistical sequences for each clutter unit and combines CNN and LSTM networks to effectively mine the temporal correlation and dynamic variation patterns of clutter, significantly outperforming traditional methods that rely solely on scalar features such as mean and variance. The attention mechanism further enables adaptive weighting of historical information, improving adaptability to time-varying clutter. In terms of model construction, an end-to-end learning approach is adopted, eliminating reliance on prior distribution assumptions such as Gaussian, Rayleigh, or Weibull distributions, resulting in stronger generalization capabilities. It can be quickly transferred to different radar scenarios using limited data and supports online incremental optimization.

[0091] The model of this invention outputs the probability of target presence, and the detection threshold can be adaptively determined in the validation set based on a preset false alarm rate, improving the flexibility and convenience of engineering applications. This invention also has good versatility and scalability, is suitable for multidimensional clutter diagram modeling, and can be extended to fields such as sonar, seismic, and industrial anomaly detection.

[0092] Based on the same inventive concept, this invention provides a radar clutter map target detection device based on a dual-branch cross-attention neural network, comprising: The echo acquisition module is used to acquire baseband echo data and corresponding radar operating parameters in the target scene, and to preprocess the baseband echo data. The preprocessing includes pulse compression and Doppler processing. The radar operating parameters include beam pointing, pulse repetition period, dwell time and corresponding timestamp. The truth acquisition module is used to obtain the truth information of the cooperation target and convert the truth information to align with the timestamp. The matrix construction module is used to divide the probe airspace into clutter units, recursively update the statistical features of each clutter unit without targets, and construct the historical frame statistical feature sequence matrix corresponding to each clutter unit. The dataset module is used to construct a dataset based on the statistical feature sequence matrix of historical frames and the corresponding amplitude of the current frame, and to perform data augmentation and dataset partitioning on the dataset. The training module is used to construct a dual-branch cross-attention deep learning network and train the dual-branch cross-attention deep learning network based on the dataset. The dual-branch cross-attention deep learning network is used for object detection.

[0093] Based on the same inventive concept, the present invention also provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to execute a radar clutter map target detection method based on a dual-branch cross-attention neural network, as described above.

[0094] Based on the same inventive concept, the present invention also provides a non-transitory computer-readable storage medium, which, when the instructions in the storage medium are executed by the processor of an electronic device, enables the electronic device to execute the radar clutter map target detection method based on the dual-branch cross-attention neural network provided above.

[0095] Since the electronic device described in this embodiment is an electronic device used to implement the information processing method in the embodiments of the present invention, those skilled in the art can understand the specific implementation methods and various variations of the electronic device in this embodiment based on the information processing method described in the embodiments of the present invention. Therefore, how the electronic device implements the method in the embodiments of the present invention will not be described in detail here. Any electronic device used by those skilled in the art to implement the information processing method in the embodiments of the present invention falls within the scope of protection of the present invention.

[0096] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0097] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0098] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0099] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0100] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0101] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A radar clutter map target detection method based on a dual-branch cross-attention neural network, characterized in that, include: The baseband echo data and corresponding radar operating parameters in the target scene are acquired, and the baseband echo data is preprocessed. The preprocessing includes pulse compression and Doppler processing. The radar operating parameters include beam pointing, pulse repetition period, dwell time and corresponding timestamp. Obtain the truth information of the cooperation target, and transform and align the truth information with the timestamp; The probe airspace is divided into clutter units. The statistical features of each clutter unit without targets are recursively updated, and a historical frame statistical feature sequence matrix corresponding to each clutter unit is constructed. A dataset is constructed based on the historical frame statistical feature sequence matrix and the corresponding current frame amplitude, and the dataset is augmented and partitioned. A dual-branch cross-attention deep learning network is constructed and trained on the dataset. The dual-branch cross-attention deep learning network is used for object detection.

2. The radar clutter map target detection method based on a dual-branch cross-attention neural network as described in claim 1, characterized in that, Also includes: In the validation set of the dataset, the target probability output distribution of clutter samples is statistically analyzed, and the corresponding quantile value is determined according to the preset false alarm probability. Determine the detection threshold factor; The performance of the trained bi-branch cross-attention deep learning network is evaluated based on the unused test set partitioned from the dataset and the detection threshold factor.

3. The radar clutter map target detection method based on a dual-branch cross-attention neural network as described in claim 1, characterized in that, Preprocessing of baseband echo data includes: The baseband echo data is pulse-compressed to obtain the corresponding range profile; Slow-time Doppler processing is applied to the range image to obtain the range and Doppler matrices.

4. The radar clutter map target detection method based on a dual-branch cross-attention neural network as described in claim 1, characterized in that, The probed airspace is divided into clutter units. The statistical features of each clutter unit without targets are recursively updated, and a historical frame statistical feature sequence matrix corresponding to each clutter unit is constructed, including: The detection airspace is divided into clutter units according to preset dimensions, including range dimension, azimuth dimension, elevation dimension and Doppler dimension; For each clutter unit, a corresponding historical frame statistical feature sequence is constructed. Each frame in the sequence contains several statistical features, forming a historical frame statistical feature sequence matrix, including: in, For the length of historical frames, For the first Statistical eigenvectors of clutter cells at frame time. For frame index, For transpose, The mean of the amplitude. The standard deviation of the amplitude. For local contrast; When any clutter unit does not contain the target in any frame, the amplitude of that frame is used for recursive updating, and the updated historical statistical sequence matrix is ​​obtained. The update formula includes: in, Forgetting factor, for The average amplitude of the frame. For the first The observed mean of the frames, for Frame variance This represents the number of amplitude observation points contained within the clutter cell. For the first The first frame The amplitude at each observation point; The differences between amplitude observations in clutter cells are characterized based on local contrast, including: 。 5. The radar clutter map target detection method based on a dual-branch cross-attention neural network as described in claim 1, characterized in that, Constructing a dual-branch cross-attention deep learning network, including: Based on the input layer simultaneously receiving historical statistical sequence matrices and the corresponding current frame amplitude; Based on the historical frame feature extraction network, temporal features are extracted from the historical frame statistical feature sequence matrix, and the temporal features are mapped to generate key vector sequence and value vector sequence. Based on the current frame feature extraction network, amplitude features are extracted to generate a query vector; Based on a cross-attention fusion network, similarity calculation and weighted fusion are performed on query vectors, key vector sequences, and value vector sequences; Residual connections are performed based on residual fusion networks to generate fused feature vectors; Based on the classification output network, the fused feature vector is mapped to the probability of the target and the probability of clutter.

6. The radar clutter map target detection method based on a dual-branch cross-attention neural network as described in claim 1, characterized in that, Training the dual-branch cross-attention deep learning network based on the dataset includes: The dual-branch cross-attention deep learning network is trained on the training set of the dataset, wherein the cross-entropy loss function is used during training, and the network parameters of the dual-branch cross-attention deep learning network are optimized through backpropagation. The loss and accuracy of the dual-branch cross-attention deep learning network are obtained in real time. When preset requirements are met, the dual-branch cross-attention deep learning network is used for target detection.

7. The radar clutter map target detection method based on a dual-branch cross-attention neural network as described in claim 1, characterized in that, Online detection is performed based on a trained dual-branch cross-attention deep learning network, including: Real-time baseband echo data is acquired and preprocessed to obtain the historical frame feature statistical sequence matrix and the corresponding current frame amplitude. By inputting the real-time statistical sequence matrix and the corresponding current frame amplitude into the trained dual-branch cross-attention deep learning network, the probability of the target corresponding to each clutter unit is obtained. Based on the probability of a target in a clutter cell and the detection threshold, determine whether a target exists in that clutter cell; The statistical characteristics of clutter cells for which no target was detected are recursively updated, and the clutter cells containing the target are sent to the target receiver.

8. A radar clutter map target detection device based on a dual-branch cross-attention neural network, characterized in that, include: The echo acquisition module is used to acquire baseband echo data and corresponding radar operating parameters in the target scene, and to preprocess the baseband echo data. The preprocessing includes pulse compression and Doppler processing. The radar operating parameters include beam pointing, pulse repetition period, dwell time and corresponding timestamp. The truth value acquisition module is used to acquire the truth value information of the cooperation target and convert the truth value information and align it with the timestamp; The matrix construction module is used to divide the probe airspace into clutter units, recursively update the statistical features of each clutter unit without targets, and construct the historical frame statistical feature sequence matrix corresponding to each clutter unit. The dataset module is used to construct a dataset based on the historical frame statistical feature sequence matrix and the corresponding current frame amplitude, and to perform data augmentation and dataset partitioning on the dataset. The training module is used to construct a dual-branch cross-attention deep learning network and train the dual-branch cross-attention deep learning network based on the dataset. The dual-branch cross-attention deep learning network is used for object detection.

9. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the radar clutter map target detection method based on a dual-branch cross-attention neural network as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that, When the instructions in the storage medium are executed by the processor of the electronic device, the electronic device is able to perform the radar clutter map target detection method based on a dual-branch cross-attention neural network as described in any one of claims 1 to 7.

Images