Radar target detection method and system based on multi-head attention mechanism

Abstract

The invention discloses a radar target detection method and system based on a multi-head attention mechanism, and relates to the radar technology, and the method comprises the steps: carrying out the screening of radar condensed plots, so as to generate candidate tracks; for the candidate track, adopting a two-layer bidirectional long-short time memory network Bi-LSTM to extract candidate track features; and for the candidate track, adopting a PointNet network to extract clutter point cloud features of an area where the candidate track is located; carrying out feature fusion on the extracted candidate track features and clutter point cloud features by using a multi-head attention mechanism MHA; and outputting a radar target detection result from the fused n attention channel data through a full connection layer. The problem that the radar target detection speed and the detection accuracy are difficult to balance in a traditional processing method can be solved, and the accuracy of target detection in the clutter environment is improved.

Description

Technical Field

[0001] This application relates to the field of radar technology, and in particular to a radar target detection method and system based on a multi-head attention mechanism. Background Technology

[0002] Target detection is the first step in multi-target tracking, and its main function is to automatically establish the trajectory of targets entering the radar detection range. In complex electromagnetic environments, the scale of radar tracks increases significantly. Traditional methods such as intuitive methods, logical methods, and Hough transforms use the target's motion characteristics such as heading, speed, and acceleration to make rule-based judgments. However, these methods suffer from problems such as high subjectivity in threshold setting, difficulty in achieving accuracy, and inability to adapt to all scenarios.

[0003] Existing deep learning-based path initiation methods, such as those based on SVM or LSTM, leverage the strong feature extraction capabilities of machine learning algorithms or neural networks to improve target detection performance to some extent. However, because the model focuses on the candidate path itself and does not fully utilize the features of the clutter environment surrounding the candidate path, the target detection performance drops significantly in environments with dense clutter. Summary of the Invention

[0004] This application provides a radar target detection method and system based on a multi-head attention mechanism to improve the accuracy of target detection in cluttered environments.

[0005] This application provides a radar target detection method based on a multi-head attention mechanism, including: The original tracks are filtered to obtain candidate tracks; For the candidate tracks, a two-layer bidirectional long short-term memory network (Bi-LSTM) is used to extract the candidate track features; and, For the candidate tracks, the PointNet network is used to extract the clutter point cloud features of the region where the candidate tracks are located; The extracted candidate track features and clutter point cloud features are fused using the multi-head attention mechanism (MHA). The fused data from the n attention channels is then passed through a fully connected layer to output radar target detection results.

[0006] This application also provides a radar target detection system based on a multi-head attention mechanism, including a processor and a memory. The memory stores a computer program, which, when executed by the processor, implements the steps of the radar target detection method based on the multi-head attention mechanism as described above.

[0007] This application embodiment utilizes a bidirectional long short-term memory network to extract the spatiotemporal features of candidate tracks, uses a PointNet network to extract clutter spot features in the region where the candidate tracks are located, and learns the correlation between tracks and regional clutter through a multi-head attention mechanism. This overcomes the problem of traditional processing methods struggling to balance target detection speed and accuracy, and improves the accuracy of target detection in cluttered environments.

[0008] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0009] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a flowchart illustrating the radar target detection method based on multi-head attention mechanism according to an embodiment of this application. Figure 2 This is a schematic diagram of the architecture of the radar target detection method based on multi-head attention mechanism according to an embodiment of this application; Figure 3 This application presents a Bi-LSTM track feature extraction architecture for a radar target detection method based on a multi-head attention mechanism, as described in this embodiment. Figure 4 This application presents an architecture for regional point feature extraction in a radar target detection method based on a multi-head attention mechanism. Detailed Implementation

[0010] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0011] This application proposes a method for establishing candidate radar target tracks using a logical approach, extracting features from these tracks using a Long Short-Term Memory (LSTM) network, and simultaneously using a PointNet network to extract features from the clutter environment surrounding the candidate tracks. A cross-attention module is then constructed to enable the model to learn the relationship between the candidate tracks and the clutter environment. Finally, a fully connected layer outputs the result determining whether a track is real or fake, thus improving radar target detection performance in dense clutter environments. Specifically, this application provides a radar target detection method based on a multi-head attention mechanism, such as... Figure 1 , Figure 2 As shown, it includes: In step S101, the radar converged points are filtered to obtain candidate tracks; in some embodiments, filtering the radar converged points to obtain candidate tracks includes: Definition of the first The number of radar points formed in each cycle is ,from The set of radar points formed over M cycles starting from time M ; Candidate tracks are selected using a logical method for the radar track set X: Define the target's maximum speed and target minimum speed If the radar detects the target twice, the target velocity is calculated based on the measured displacement and the scanning interval. ,but satisfy: Define the target's maximum acceleration Calculate the target acceleration, then the target acceleration satisfy: The yaw angle of a target is defined as follows: assuming the target's three adjacent positions are A, B, and C, then the vector... sum vector The angle between them is the yaw angle, and the maximum yaw angle of the target is defined. and minimum yaw angle Then the target's yaw angle satisfies: Candidate tracks were obtained through screening. .

[0012] In step S102, for the candidate tracks, a two-layer bidirectional long short-term memory network (Bi-LSTM) is used to extract the candidate track features. In some embodiments, such as... Figure 3As shown, the two-layer Bi-LSTM includes several LSTM units, each of which includes an input gate, a memory unit, an output gate, and a forget gate, satisfying the following: in, Output vector for the forget gate. The input gate output vector, The output gate outputs a vector. for The long-term memory vector at each moment, Candidate memory vectors, for The hidden state at all times These are the weight matrices for the forget gate, input gate, and output gate, respectively. These are the bias vectors for the forget gate, input gate, and output gate, respectively. The input vector at the current time step, for Activation function; Since LSTM only uses sequence information from previous time steps when predicting the output at the current time step, in order to improve the model's feature extraction capability, this application adopts a Bi-LSTM model, which can combine an LSTM that moves from the beginning of the sequence in time step and another LSTM that moves from the end of the sequence in time step. The output is obtained by connecting the states of the forward LSTM and the backward LSTM.

[0013] The candidate tracks are subjected to a random mask strategy, which means that during training, a specified number of position features are randomly selected and removed, and missing features are filled with 0, in order to simulate the radar's missed detection conditions, improve the network's generalization and detection capabilities.

[0014] In step S103, the PointNet network is used to extract clutter point cloud features of the region where the candidate track is located. This step corresponds to region feature extraction. For point-level target detection, the accuracy of detection is related to both the features of the candidate track itself and the clutter region it is currently in. In some embodiments of this application, the extraction of clutter point cloud features of the region where the candidate track is located using the PointNet network includes: The PointNet network, such as Figure 4 As shown, input alignment and spatial transformation are achieved through two T-Net modules to solve the feature bias problem caused by point cloud spatial rotation and translation, thereby improving the robustness of the model. High-dimensional local features are then extracted through a multilayer perceptron (MLP), and the features of all points are integrated through Max Pooling to generate a global description, ensuring that the model is insensitive to disorder.

[0015] In step S104, the extracted candidate track features and clutter point cloud features are fused using the multi-head attention mechanism (MHA). In step S105, the fused n attention channel data are output as radar target detection results through a fully connected layer.

[0016] In some embodiments, sample augmentation based on the candidate tracks is further performed by at least one of the following methods: While keeping the latitude and longitude of the candidate flight path constant, the entire path is raised or lowered in the altitude direction to simulate targets at different flight altitudes. For example, in some examples, the current flight path altitude sequence is set as follows: Translation amount Defined as: Considering that the altitude of civil aircraft is usually between 8,000 and 12,000 meters, the altitude after translation is limited to ensure that the sample conforms to the laws of physics.

[0017] The candidate track is rotated to simulate targets in different directions. The rotation occurs around a specified center point, and the entire track is rotated, for example, by defining a heading rotation angle. The rotation matrix is ​​defined as: Apply rotation: In some embodiments, normalization is further performed using a max-min normalization method based on the candidate tracks. Specifically, normalization can be performed according to the value range of each feature, compressing the data to the [0,1] interval. The specific formula is as follows: in, For the first input data One characteristic, and Features The maximum and minimum values ​​are determined. Normalization eliminates dimensional differences, improves the convergence speed and accuracy of the algorithm, and provides reliable feature support for subsequent feature extraction.

[0018] In some embodiments, the extracted candidate track features and clutter point cloud features are fused using a multi-head attention (MHA) mechanism. MHA is the core feature fusion mechanism of the Transformer architecture, capable of capturing multi-level relationships in sequence data and significantly improving feature representation capabilities. Each attention weight in MHA is calculated as follows: in, Q is the query vector, K is the key vector, and V is the value vector. For the first The output of each attention head, This is a linear transformation matrix used to perform a linear transformation on the spliced ​​attention head output.

[0019] This application utilizes a multi-head attention mechanism to project (Q,K,V) onto... n Different subspaces, generating n Multiple attention channels enable the model to obtain information from different aspects in different representation subspaces, thereby improving the ability of feature representation.

[0020] The loss function design primarily considers the accuracy and complexity of the network model's predictions. Accuracy loss is measured by calculating the cross-entropy loss between the predicted and true labels, as follows: Considering network complexity, this application employs regularization and other methods to control network complexity, prevent overfitting of the training data, improve the network's generalization ability, and enable the model to effectively detect targets in various clutter environments. In some embodiments, it further includes: employing Elastic Net regularization, combined with L1 and L2 regularization, and performing model training using the following loss function: Where N is the sample size, For the first The true label of each sample For the first The predicted probability of a sample. For weight parameters, , This is a regularization strength hyperparameter used to control the penalty applied to the weights.

[0021] In a specific example, ground truth data (such as ADS-B or AIS data) is used to identify potential target candidate tracks. The labeling process involves first interpolating the ground truth data, then associating the candidate tracks with the ground truth data using a spherical gate with a distance threshold of 500 meters. If more than 2 / 3 of the points in a candidate track are associated with the ground truth data of the same target, the candidate track sample is labeled as "1"; otherwise, it is labeled as "0".

[0022] In some embodiments, outputting radar target detection results through a fully connected layer from the fused n attention channel data includes: after training, processing the input raw trajectory, using the probability of the fully connected layer outputting the true target; if the probability is greater than a preset threshold, the target detection is considered successful, and the predicted value and confidence level are output. For example, after training, loading the trained neural network parameters, initializing the neural network, and then inputting the processed candidate trajectory and surrounding point cloud into the deep neural network for calculation to obtain the probability that the candidate trajectory is a true target. If the probability If the probability is greater than, for example, 80%, then the candidate track is considered to have a high probability of being the real target, and the predicted value 1 and confidence level are output. .

[0023] This application addresses the challenges of high clutter input, high track density, and difficulty in suppressing false tracks in radar system target detection. It innovatively proposes an intelligent target detection method based on multi-head attention, designing a bidirectional LSTM+PointNet multi-head attention network model structure. This enables the model to better extract candidate tracks and their surrounding clutter track features, solving the problem of difficulty in suppressing false tracks in dense clutter environments using traditional methods. This improves the accuracy of target detection and provides highly reliable situational information for subsequent command and decision-making.

[0024] This application also proposes a radar target detection system based on a multi-head attention mechanism, including a processor and a memory. The memory stores a computer program, which, when executed by the processor, implements the steps of the radar target detection method based on the multi-head attention mechanism as described above.

[0025] It should be noted that, in the embodiments of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0026] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0027] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0028] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims. All of these forms are within the protection scope of this application.

Claims

1. A radar target detection method based on a multi-head attention mechanism, characterized in that, include: The original radar points are filtered to obtain candidate tracks; For the candidate tracks, a two-layer bidirectional long short-term memory network Bi-LSTM is used to extract the candidate track features; For the candidate tracks, the PointNet network is used to extract the clutter point cloud features of the region where the candidate tracks are located; The extracted candidate track features and clutter point cloud features are fused using the multi-head attention mechanism (MHA). The fused data from the n attention channels is then passed through a fully connected layer to output radar target detection results.

2. The radar target detection method based on multi-head attention mechanism as described in claim 1, characterized in that, The radar convergence points are filtered to obtain candidate tracks, including: Definition of the first The number of radar spots formed in each cycle is ,from The set of radar measurement points formed over M cycles starting from time M ; Candidate tracks are selected using a logical method for the radar measurement point set X: Define the target's maximum speed and target minimum speed If the radar detects the target twice, the target velocity is calculated based on the measured displacement and the scanning interval. ,but satisfy: Define the target's maximum acceleration Calculate the target acceleration, then the target acceleration satisfy: The yaw angle of a target is defined as follows: assuming the target's three adjacent positions are A, B, and C, then the vector... sum vector The angle between them is the yaw angle, and the maximum yaw angle of the target is defined. and minimum yaw angle Then the target's yaw angle satisfies: Candidate tracks were obtained through screening. .

3. The radar target detection method based on multi-head attention mechanism as described in claim 2, characterized in that, It also includes sample augmentation based on the candidate tracks through at least one of the following methods: Keeping the latitude and longitude of the candidate flight path unchanged, the entire flight path is raised or lowered in the altitude direction to simulate targets at different flight altitudes; The candidate track is rotated to simulate targets in different directions, rotating the entire track around a specified center point.

4. The radar target detection method based on multi-head attention mechanism as described in claim 2, characterized in that, It also includes normalization based on candidate tracks using the max-min normalization method.

5. The radar target detection method based on multi-head attention mechanism as described in claim 1, characterized in that, The two-layer Bi-LSTM comprises several LSTM units, each including an input gate, a memory unit, an output gate, and a forget gate, satisfying the following: in, Output vector for the forget gate. The input gate output vector, The output gate outputs a vector. for The long-term memory vector at each moment, Candidate memory vectors, for The hidden state at all times These are the weight matrices for the forget gate, input gate, and output gate, respectively. These are the bias vectors for the forget gate, input gate, and output gate, respectively. The input vector at the current time. for Activation function; The candidate tracks are subjected to a random mask strategy, which means that during training, a specified number of position features are randomly selected and removed, and missing features are filled with 0 to simulate radar miss detection conditions.

6. The radar target detection method based on multi-head attention mechanism as described in claim 1, characterized in that, For the candidate tracks, the PointNet network is used to extract clutter point cloud features of the region where the candidate tracks are located, including: The PointNet network achieves input alignment and spatial transformation through two T-Net modules, extracts high-dimensional local features through a multilayer perceptron (MLP), and integrates the features of all points through Max Pooling to generate a global description.

7. The radar target detection method based on multi-head attention mechanism as described in claim 5, characterized in that, The extracted candidate track features and clutter point cloud features are fused using a multi-head attention mechanism (MHA), including: Each attention weight is calculated as follows: in, Q is the query vector, K is the key vector, and V is the value vector. For the first The output of each attention head, This is a linear transformation matrix used to perform a linear transformation on the spliced ​​attention head output.

8. The radar target detection method based on multi-head attention mechanism as described in claim 7, characterized in that, Also includes: Elastic Net regularization, combined with L1 and L2 regularization, is used to train the model using the following loss function: Where N is the sample size. For the first The true label of each sample For the first The predicted probability of a sample. For weight parameters, , This is the hyperparameter for regularization intensity.

9. The radar target detection method based on multi-head attention mechanism as described in claim 8, characterized in that, The fused data from the n attention channels is passed through a fully connected layer to output radar target detection results, including: After training, the input raw trajectory is processed, and the probability of the fully connected layer outputting the true target is used. If the probability is greater than a preset threshold, it is converted into a confirmed trajectory, thus completing the target detection.

10. A radar target detection system based on a multi-head attention mechanism, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, which, when executed by the processor, implements the steps of the radar target detection method based on the multi-head attention mechanism as described in any one of claims 1 to 9.

Images