CN-121978645-A - Radar target detection method and system based on multi-head attention mechanism

CN121978645ACN 121978645 ACN121978645 ACN 121978645ACN-121978645-A

Abstract

The application discloses a radar target detection method and a radar target detection system based on a multi-head attention mechanism, which relate to radar technology and comprise the steps of screening points after radar condensation to generate candidate tracks; the method comprises the steps of extracting candidate flight path features of the candidate flight path by adopting a Bi-directional long-short-time memory network Bi-LSTM of two layers, extracting clutter point cloud features of an area where the candidate flight path is located by adopting a PointNet network for the candidate flight path, carrying out feature fusion on the extracted candidate flight path features and the clutter point cloud features by utilizing a multi-head attention mechanism MHA, and outputting radar target detection results by using n attention channel data after fusion through a full-connection layer. The method can solve the problem that the traditional processing method is difficult to balance in radar target detection speed and detection accuracy, and improves the accuracy of target detection in clutter environment.

Inventors

JIN ZHONGQIAN
Zhang Xiushe
HAN CHUNLEI
LU YAO
ZHANG YANG
YANG DI
ZHAO WANG
XU HAOYANG

Assignees

中国电子科技集团公司第二十研究所

Dates

Publication Date: 20260505
Application Date: 20251223

Claims (10)

1. A radar target detection method based on a multi-head attention mechanism, comprising: Selecting radar original points to obtain candidate tracks, and extracting candidate track features by adopting a two-layer Bi-directional long-short-time memory network Bi-LSTM for the candidate tracks; Extracting clutter point cloud characteristics of the region where the candidate flight path is located by adopting PointNet networks for the candidate flight path; Performing feature fusion on the extracted candidate track features and clutter point cloud features by utilizing a multi-head attention mechanism MHA; And outputting radar target detection results through the full connection layer by the n fused attention channel data.
2. The multi-headed attentiveness-mechanism-based radar target detection method in accordance with claim 1, wherein filtering radar condensation tracks to obtain candidate tracks comprises: Definition of the first embodiment The number of radar traces formed in each period is From the slave Radar measuring track set formed by M periods from moment to moment ; And screening candidate tracks by adopting a logic method for the radar measurement track set X: Defining maximum speed of a target And a target minimum speed In the case of radar scanning a target twice, the target speed is calculated from the measured displacement and the scan interval Then The method meets the following conditions: Defining the maximum acceleration of the target Calculating the target acceleration, then the target acceleration The method meets the following conditions: the yaw angle of the object is defined as the vector assuming three adjacent positions of the object are A, B, C Sum vector The included angle between the two is the yaw angle, and the maximum yaw angle of the target is defined And a minimum yaw angle The yaw angle of the target satisfies: Screening to obtain candidate flight path as 。
3. The multi-headed gaze mechanism based radar target detection method of claim 2, further comprising sample enhancement based on said candidate tracks by at least one of: keeping the longitude and latitude of the candidate flight path unchanged, and integrally lifting or lowering the whole flight path in the height direction so as to simulate targets with different flying heights; and rotating the candidate tracks to simulate targets in different directions, and rotating the whole track around a designated center point during rotation.
4. The multi-headed attentiveness-based radar target detection method in accordance with claim 2, further comprising normalizing using a max-min normalization method based on candidate tracks.
5. The multi-head attention mechanism based radar target detection method of claim 1, wherein the Bi-LSTM of the two layers comprises a plurality of LSTM cells, each LSTM cell comprising an input gate, a memory cell, an output gate, and a forget gate, satisfying: Wherein, the For the forget gate to output a vector, For the input gate to output a vector, In order to output the gate output vector, Is that The long-term memory vector of the moment, As a result of the candidate memory vector, Is that The hidden state of the moment of time, The weight matrix of the forgetting gate, the input gate and the output gate are respectively arranged, The bias vectors of the forgetting gate, the input gate and the output gate are respectively, The vector is input for the current moment of time, Is that Activating a function; And adopting a random Mask strategy for the candidate tracks, namely randomly selecting a specified number of position features to remove during training, wherein the missing features are filled with 0 so as to simulate the missing detection condition of the radar.
6. The method for radar target detection based on multi-head attention mechanism of claim 1, wherein extracting clutter point cloud features of an area where the candidate track is located by adopting PointNet networks for the candidate track comprises: The PointNet network realizes input alignment and space transformation through 2T-Net modules, extracts high-dimensional local features through a multi-layer perceptron MLP, integrates the features of all points through Max Pooling, and generates global description.
7. The multi-head attention mechanism based radar target detection method of claim 5, wherein performing feature fusion on the extracted candidate track features and clutter point cloud features by using a multi-head attention mechanism MHA comprises: Each attention weight is calculated as: Wherein, the Q is a query vector, K is a key vector, V is a value vector, Is the first The output of the individual attention heads, Is a linear transformation matrix used for carrying out linear transformation on the spliced attention head output.
8. The multi-headed gaze mechanism based radar target detection method of claim 7, further comprising: model training is performed with elastic network regularization (ELASTIC NET), in combination with L1 and L2 regularization, by the following loss function: where N is the number of samples, Is the first The true labels of the individual samples are then displayed, Is the first The probability of prediction of the individual samples is determined, As a parameter of the weight-bearing element, , Is a regularized intensity superparameter.
9. The method for radar target detection based on a multi-head attention mechanism of claim 8, wherein outputting the radar target detection result from the n fused attention channel data through the full connection layer comprises: After training is completed, the input original track is processed, the probability that the full connection layer outputs a real target is utilized, if the probability is larger than a preset threshold value, the probability is converted into a confirmation track, and target detection is completed.
10. A multi-headed-attention-mechanism-based radar target detection system comprising a processor and a memory, the memory having stored thereon a computer program which, when executed by the processor, implements the steps of the multi-headed-attention-mechanism-based radar target detection method of any one of claims 1 to 9.

Description

Radar target detection method and system based on multi-head attention mechanism Technical Field The application relates to the technical field of radars, in particular to a radar target detection method and system based on a multi-head attention mechanism. Background The target detection is the first step of multi-target tracking, and the main function is to automatically establish tracks for targets entering the radar detection range. In a complex electromagnetic environment, the scale of radar trace points can be greatly increased, and the traditional methods such as an intuitive method, a logic method, hough transformation and the like use the motion characteristics such as target heading, navigational speed, acceleration and the like to judge based on rule constraint, so that the problems that the subjectivity of threshold setting is large, the accuracy is difficult to achieve, and all scenes are adapted are solved. The existing track initiation method based on deep learning, such as SVM or LSTM, can improve the performance of target detection to a certain extent by utilizing the advantage of strong feature extraction capability of a machine learning algorithm or a neural network, but the feature of clutter environment around the candidate track is not fully utilized due to the fact that the model focuses on the candidate track, so that the target detection performance is greatly reduced in the environment with denser clutter. Disclosure of Invention The embodiment of the application provides a radar target detection method and a radar target detection system based on a multi-head attention mechanism, which are used for improving the accuracy of target detection in a clutter environment. The embodiment of the application provides a radar target detection method based on a multi-head attention mechanism, which comprises the following steps: screening the original tracks to obtain candidate tracks; extracting candidate track features by adopting a Bi-directional long-short-time memory network Bi-LSTM of two layers for the candidate tracks, and Extracting clutter point cloud characteristics of the region where the candidate flight path is located by adopting PointNet networks for the candidate flight path; Performing feature fusion on the extracted candidate track features and clutter point cloud features by utilizing a multi-head attention mechanism MHA; And outputting radar target detection results through the full connection layer by the n fused attention channel data. The embodiment of the application also provides a radar target detection system based on the multi-head attention mechanism, which comprises a processor and a memory, wherein the memory is stored with a computer program, and the computer program realizes the steps of the radar target detection method based on the multi-head attention mechanism when being executed by the processor. According to the embodiment of the application, the space-time characteristics of the candidate tracks are extracted by using the bidirectional long-short-term memory network, the clutter point track characteristics of the region where the candidate tracks are located are extracted by using the PointNet network, the relevance between the tracks and the regional clutter is learned by using a multi-head attention mechanism, the problem that the traditional processing method is difficult to balance in terms of target detection speed and detection accuracy is solved, and the accuracy of target detection in a clutter environment is improved. The foregoing description is only an overview of the present application, and is intended to be implemented in accordance with the teachings of the present application in order that the same may be more clearly understood and to make the same and other objects, features and advantages of the present application more readily apparent. Drawings Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are only for purposes of illustrating the preferred embodiments and are not to be construed as limiting the application. Also, like reference numerals are used to designate like parts throughout the figures. In the drawings: FIG. 1 is a flow chart of a radar target detection method based on a multi-head attention mechanism according to an embodiment of the present application; FIG. 2 is a schematic architecture of a radar target detection method based on a multi-head attention mechanism according to an embodiment of the present application; FIG. 3 is a Bi-LSTM track feature extraction architecture of a radar target detection method based on a multi-head attention mechanism according to an embodiment of the present application; Fig. 4 is a block diagram of a multi-head attention mechanism-based radar target detection method according to an embodiment of the present application. Detailed Description Exemplar