Pulse Doppler radar target detection method and system based on background contrast attention mechanism

Abstract

The invention discloses a pulse Doppler radar target detection method and system based on a background contrast attention mechanism, and relates to the field of radar target detection. The method comprises the following steps: preprocessing data to generate a distance Doppler image; constructing a neural network based on a background contrast attention mechanism; the neural network is trained, and the trained neural network is obtained; and obtaining a detection result of a pulse Doppler image to be detected by using the trained neural network. Based on a background contrast attention mechanism, the method focuses on learning clutter distribution characteristics of a background in a distance Doppler map and inhibiting background features from a feature level, thereby improving the significance of target features. A multi-scale background learner module based on a random mask and an attention mechanism module based on background contrast are designed in a targeted manner, and JS divergence is introduced into a loss function, so that background interference can be better suppressed from a feature level, and the detection performance in a low signal-to-clutter ratio scene is improved.

Description

Technical Field

[0001] The present invention belongs to the field of signal processing and target detection, and particularly relates to a pulse Doppler radar target detection method and system based on a background contrast attention mechanism. Background Art

[0002] Radar is one of the effective tools for detecting unmanned aerial vehicle targets. Radar-based target detection refers to detecting and positioning a target according to the echo signal received by the radar through preprocessing and target detection algorithms to obtain relevant information about the target. With its advantages of working all day and in all climates, radar can capture the distance and speed information of the target, providing an effective way for target detection and tracking.

[0003] Current radar target detection methods include statistical model-based methods and feature-based methods. Statistical model-based methods analyze whether there is a target in the signal by hypothesizing and modeling clutter and targets, with the constant false alarm rate algorithm as a representative. Since such methods require prior hypothesis modeling of background clutter, their performance will decrease sharply in current complex and changeable scenarios. Feature-based methods use the feature differences between targets and backgrounds to distinguish targets, and the features include fractal features, chaotic features, time-frequency features, and micro-Doppler features, etc. Deep learning can automatically extract potential features of targets through training and then detect the targets, with strong robustness and generalization ability. Therefore, deep neural networks are gradually used in radar target detection. The input of the neural network can be the original radar echo signal, the time-frequency diagram of the signal, the range-Doppler diagram of the signal, etc., and the range-Doppler diagram is a relatively common data format. When current deep learning algorithms based on the range-Doppler diagram perform target detection, they often directly extract target feature information while ignoring the learning of the background, which leads to an increase in the false alarm rate of the network and a decrease in detection performance in scenarios with a low signal-to-clutter ratio.

[0004] Radar target detection algorithms model radar echo signals and detect targets, but their traditional algorithms have poor robustness, and the detection performance of current deep learning-based algorithms is greatly affected in scenarios with a low signal-to-clutter ratio. Summary of the Invention

[0005] To solve the problems in the prior art, the present invention proposes a pulse Doppler radar target detection method and device system based on a background contrast attention mechanism.

[0006] The technical solution adopted by the present invention is as follows:

[0007] In a first aspect, the present invention discloses a pulse Doppler radar target detection method based on a background contrast attention mechanism, including the following steps:

[0008] Step 1): Obtain the pulsed Doppler image of the echo according to the echo signal of the radar;

[0009] Step 2): Construct a neural network based on the background contrast attention mechanism, which is used to obtain the distance and speed information of the target in the range-Doppler image. The neural network includes a multi-scale background learner module based on random masks, a Transformer encoder module based on the background contrast attention mechanism, a Transformer decoder, and a feed-forward network module;

[0010] Step 3): Train the neural network based on the background contrast attention mechanism to obtain a trained neural network;

[0011] Step 4): Use the trained neural network to obtain the target category information and position information in the to-be-detected range-Doppler image, so as to realize the detection of radar targets.

[0012] In a second aspect, the present invention discloses a pulsed Doppler radar target detection system based on the background contrast attention mechanism for the above method, including:

[0013] A signal preprocessing module, which is used to generate a pulsed Doppler image corresponding to the radar echo;

[0014] A neural network construction module, which is used to construct a neural network based on the background contrast attention mechanism. The neural network is used to obtain the category information and position information of the target in the range-Doppler image. The neural network includes a multi-scale background learner module based on random masks, a Transformer encoder module based on the background contrast attention mechanism, a Transformer decoder, and a feed-forward network module;

[0015] A neural network training module, which is used to train the neural network based on the background contrast attention mechanism to obtain a trained neural network;

[0016] A range-Doppler image detection module, which is used to use the trained neural network to obtain the target category information and position information in the to-be-detected range-Doppler image, so as to realize the target detection of the range-Doppler image.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0018] 1) The present invention proposes a DETR detection network for range-Doppler images in the field of radar (Detection Transformer detection network, an end-to-end object detection network based on the Transformer architecture). For the first time, the DETR framework is introduced into the field of radar target detection, and background interference in the feature dimension is suppressed by learning background features during the target detection process, rather than directly extracting target information. This method effectively solves the problem of high false alarm rate in low signal-to-clutter ratio scenarios.

[0019] 2) The present invention proposes a multi-scale background learning module based on random masking, which uses random masking to eliminate the influence of the target on the background when learning background features, and learns to fuse multi-scale information to effectively obtain background features.

[0020] 3) The present invention proposes an attention mechanism module based on background contrast, and incorporates the JS divergence into the loss function to measure the difference degree between target features and background features. This module can suppress background interference in the feature dimension and extract purer target information. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a basic step flowchart of an embodiment of the radar target detection method of the present invention;

[0022] Figure 2 is a schematic structural diagram of the attention mechanism module based on background contrast of the present invention;

[0023] Figure 3 is a schematic structural diagram of the radar target detection device of the present invention;

[0024] Figure 4 is a partial schematic structural diagram of the Transformer decoder and feed-forward network module of the present invention;

[0025] Figure 5 is a detection result diagram of the experimental simulation range-Doppler image dataset and the detection results after detection using the embodiment of the present invention and different methods. DETAILED DESCRIPTION OF THE INVENTION

[0026] The following further elaborates and explains the present invention in conjunction with the specific embodiments. The embodiments are only examples of the present disclosure content and do not delimit the scope of limitation. Without conflict, the technical features of each embodiment of the present invention can be combined accordingly.

[0027] The object processed by the present invention is the down-converted radar echo signal. As a well-known technology in the art, down-conversion is the process of converting a high-frequency signal into a lower-frequency signal. By reducing the frequency and retaining the bandwidth and phase information, the purpose is to facilitate subsequent processing and analysis.

[0028] AsFigure 1 As shown in the figure, it is the basic step flow chart of the pulsed Doppler radar target detection method based on the background contrast attention mechanism of the present invention. The method of the present invention mainly includes:

[0029] Step 1): Signal preprocessing

[0030] For the down-converted radar echo signal to be processed, in this embodiment, pulse compression and Doppler processing methods are used to obtain the range-Doppler image of the target. Let S r1 (t) represent the one-dimensional echo sequence of the radar, t represents the time, and the calculation formula for pulse compression is as follows:

[0031] S rp (t) = S r1 (t) * h(t)

[0032] where h(t) is the signal obtained by taking the conjugate after folding the radar transmit signal backwards, and S rp (t) represents the signal after pulse compression; for multiple pulse echoes, the time t is written as where t m represents the slow time, m represents the number of received pulse echoes, T represents the pulse repetition interval, represents the fast time, that is, the time series within one pulse repetition interval. Then, the one-dimensional target echo S rp (t) is expressed as a two-dimensional target echo

[0033] Perform a fast Fourier transform along the fast time dimension on to obtain the range-Doppler image matrix of the target echo The calculation formula is as follows:

[0034]

[0035] where N represents the number of sampling points. In this embodiment, N = 256, k represents the frequency; the range-Doppler image matrix is used as the range-Doppler image I RD .

[0036] Step 2): Construct a neural network based on the background contrast attention mechanism. The neural network includes a multi-scale background learner module based on random masking, a Transformer encoder module based on the background contrast attention mechanism, a Transformer decoder, and a feed-forward network module. Among them, the multi-scale background learner module based on random masking is used to learn the distribution characteristics of background clutter in the range-Doppler image. The Transformer encoder module based on the background contrast attention mechanism is used to extract the feature information of the target from the range-Doppler image and suppress background interference in the feature dimension. The Transformer decoder and the feed-forward network module parse the information provided by the encoder and output the category information and location information of the target. In this embodiment, random parameters are used as the initial weights of the network.

[0037] The specific working processes of each module in the neural network based on the background contrast attention mechanism are as follows:

[0038] Step 21): The multi-scale background learner module based on random masking aims to learn the distribution characteristics of background clutter, which is used to subsequently suppress the background information in the target from the feature dimension. Its structure is shown in Figure 1 . For the input range-Doppler image I RD , first perform a slicing operation to convert the range-Doppler image into range-Doppler image patches, which is convenient for random masking operations. Random masking reduces the contamination of the background by the target when learning the background as much as possible by randomly designing several blocks in the range-Doppler image patches. Its calculation process is as follows:

[0039] I patch = Slice(I RD )

[0040] T patch = Random_Mask(I patch )

[0041] where, I patch represents the range-Doppler image; Slice() represents the slicing function, Random_Mask() represents the random masking function, and T patch represents the processed sample.

[0042] Then, perform feature extraction and fusion on T patch through three convolution kernels of different sizes to achieve multi-scale background feature learning. Here, depthwise separable convolution is used for the convolution, which can reduce the number of parameters and the amount of calculation while maintaining performance compared to ordinary convolution. It is represented by "DWConv" in Figure 1 . The calculation process is as follows:

[0043] T1 = DWConv77(T paach)

[0044] T2 = DWConv55(T patch )

[0045] T3 = DWConv33(T patch )

[0046] T concat = Concat(T1, T2, T3)

[0047] T temp = DWConvBlock(T concat )

[0048] T B = LayerNorm(AvgPool(T temp ))

[0049] where DWConv77, DWConv55, DWConv33 represent depthwise separable convolution functions with convolution kernel sizes of 7×7, 5×5, and 3×3 respectively; Concat() represents the concatenation function; DWConvBlock() is a feature extraction function composed of three depthwise separable convolutions with a convolution kernel size of 3×3, and T1, T2, T3, T concat , T temp are intermediate results obtained from the corresponding operations; AvgPool() represents the average pooling function, and LayerNorm() represents the layer normalization function. The final feature obtained by the multi-scale background learner module based on the random mask

[0050] Step 22): The transformer encoder module based on the background contrast attention mechanism consists of a convolutional neural network, positional encoding, and a Transformer encoder. Its structure is shown in Figure 1 . In this embodiment, the convolutional neural network selects the structure of ResNet-50. Therefore, for the input range-Doppler image I RD , after passing through the CNN and positional encoding, a feature map F RD with positional encoding is obtained. The calculation formula is as follows:

[0051] F RD = ResNet(I RD ) + Positional Encoding

[0052] where, c represents the number of channels, h represents the height, w represents the width, ResNet() represents the deep residual network, and Positional Encoding is the positional encoding.

[0053] Then, flatten F RD and pass it through an embedding layer to obtain the feature T input to the self-attention module RD1 , and the calculation formula is as follows:

[0054] T RD1 = Embedding(Flatten(F RD ))

[0055] where Flatten() represents the flattening operation, and Embedding() represents the embedding layer operation, d model represents the number of channels in the embedding layer.

[0056] The Transformer encoder module based on the background contrast attention mechanism includes multiple cascaded Transformer encoder blocks, and each Transformer encoder block includes a multi-head self-attention module, a background contrast-based attention module, and a multi-layer perceptron module;

[0057] The Transformer encoder block performs multi-head self-attention mechanism calculation, residual connection, and normalization operations on T RD1 , and the process can be expressed as:

[0058] Q = T RD1 W q , K = T RD1 W k , V = T RD1 W v

[0059] head1,…,head h = Split([Q,K,V])

[0060]

[0061] T = Concat(head′ i )W0

[0062] T RD2 = LayerNorm(T + T RD1 )

[0063] where W q , W k , W v represent the projection matrices of the query matrix Q, the key matrix K, and the value matrix V respectively; head i represents the feature vector T obtained after the range-Doppler image undergoes preliminary feature extraction RDThe i-th attention head after being split, there are h attention heads in this embodiment; Split([Q, K, V]) is the attention head splitting function; Softmax is the activation function that can convert the input into weights with a sum of 1 and weight them to the value matrix V; head′ i represents the i-th attention head after attention weighting; Q i is the query matrix of the i-th attention head; is the transpose of the key matrix of the i-th attention head; V i is the value matrix of the i-th attention head; Concat(head′ i ) is the attention head concatenation function; T represents the intermediate result; W0 is a mapping matrix that makes the output T RD2 transformed into the same dimension as the input T RD1 ; LayerNorm() is the layer normalization function.

[0064] Then, T RD2 and the T B output by the multi-scale background learning module based on random masking are input into the attention mechanism module based on background contrast (BCA) together to suppress the background from the feature dimension by fusing the feature vectors of the target and the background. T RD2 First, it undergoes reshaping and average pooling operations to obtain the target feature tensor T T2 for subsequent calculations, and the calculation formula is as follows:

[0065] T T1 = Reshape(T RD2 )

[0066] T T2 = AvgPool(T T1 )

[0067] where Reshape() is the reshaping function used to change the shape of the tensor; AvgPool() is the average pooling function; The T B representing the background undergoes 1×1 convolution to obtain the same channel dimension as T T2 and the same spatial size through the replication operation, and the calculation formula is as follows:

[0068] T B1 = Conv(T B )

[0069] T B2 = Repeat(T B1 )

[0070] where Conv() is the convolution function; Repeat() is the replication function; feature feature

[0071] Concatenate the processed T T2 and T B2 to obtain T c , and fuse the target and background features through 1×1 convolution and max pooling in the pixel dimension to obtain T′ c , aiming to suppress the background at the feature level. The calculation formula is as follows:

[0072] T c =Concat(T T2 , T B2 )

[0073] T′ c =Conv(T c )+MaxPool(T c )

[0074] Then, upsample T′ c to the same spatial size as T T1 , and generate an attention weight map with stronger target saliency through the sigmoid activation function. Dot-multiply it with T T1 . After reshaping, and then through residual connection and normalization operations. The calculation formula is as follows:

[0075] T f1 =Reshape(σ(UpSample(T′ c ))·T T1 )

[0076] T f2 =LayerNorm(T f +T RD2 )

[0077] Among them, Reshape() is the reshaping function, UpSample() is the upsampling function; σ() represents the sigmoid activation function, LayerNorm() represents the layer normalization function, and T f1 , T f2 are features.

[0078] Finally, T f2 is mapped through a multi-layer perceptron. The calculation formula is as follows:

[0079] T M1 =MLP(T f2 )

[0080] T M2 =LayerNorm(T M1 +T f2 )

[0081] The output feature T of the upper cascaded Transformer encoder block M2 is input to the lower cascaded Transformer encoder block. After being processed by m cascaded Transformer encoder blocks, the Transformer encoder module based on the background contrast attention mechanism finally outputs the feature V encoder . In this embodiment, m = 4.

[0082] Step 23): The Transformer decoder and the feed-forward network module are designed to parse the target feature information output by the encoder module, and gradually convert the target queries into the category information and location information of the finally output target. The structure of the Transformer decoder is as Figure 4 shown. The Transformer decoder includes multiple cascaded Transformer decoder blocks;

[0083] The processing process of each level of Transformer decoder block includes: taking a single target query as an example, it is first initialized through positional encoding to obtain Q0, and then multi-head self-attention mechanism calculation, residual connection and normalization operations are performed on Q0 to obtain Q2. The process is the same as the multi-head self-attention mechanism in step 22), which will not be described here.

[0084] Then, the output V of the encoder encoder is used to perform multi-head attention mechanism with Q2 to obtain Q3, and then Q4 is obtained through residual connection and normalization operations. The process is as follows:

[0085] Q = Q2W q , K = (V encoder + Positional Encoding)W k , V = V encoder W v

[0086] head1,…,head h = Split([Q, K, V])

[0087]

[0088] Q3 = Concat(head′ i )W0

[0089] Q4 = LayerNorm(Q3 + Q2)

[0090] Then, mapping is performed through a multi-layer perceptron, and the calculation formula is as follows:

[0091] Q5 = MLP(Q4)

[0092] Q6 = LayerNorm(Q5 + Q4)

[0093] The output Q6 of the upper-level Transformer decoder block is input into the lower-level Transformer decoder block, and the Transformer decoder finally outputs the feature vector V decoder , in this embodiment, the Transformer decoder includes 4 cascaded Transformer decoder blocks.

[0094] Then, the feature vector output by the Transformer decoder finally outputs the target category information through two feed-forward networks, namely a classification head and a regression head. The calculation formula is as follows:

[0095] V class = MLP_Class(V decoder )

[0096] V reg = MLP_Regression(V decoder )

[0097] where MLP_Class() represents the classification head function, and MLP_Regression() represents the regression head function; V class represents the class confidence. In this example, since there are only two categories, the target and the background, therefore V reg represents the position coordinates. In this example,

[0098] Step 3): Neural network model training

[0099] Using the entire range-Doppler image as the training sample, the network adopts a supervised training method. The sample is input into the neural network model. To ensure that the background in the newly generated target features is effectively suppressed, the JS divergence is introduced to measure the similarity between the background features and the newly generated target features, and its negative value is added to the loss function. Finally, the present invention uses the weighted sum of the Hungarian loss and the JS divergence loss of the target detection box as the loss function, and the total loss function is expressed as follows:

[0100]

[0101] where, represents the Hungarian loss, Denote the JS divergence loss as $L_{JS}$, and $\lambda$ as the weight coefficient. In this embodiment, $\lambda = 1.0$. Among them, the Hungarian loss is a well-known loss function in the art. It finds the optimal matching between the predicted bounding boxes and the ground truth bounding boxes through the Hungarian algorithm, minimizing the total cost of the matching. The matching cost is usually based on a certain distance metric between the predicted bounding box and the ground truth bounding box (such as IoU or L2 distance); for the Hungarian loss, reference can be made to: Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S. (2020). End-to-End Object Detection with Transformers. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds) Computer Vision – ECCV 2020. ECCV 2020. Lecture Notes in Computer Science (), vol 12346.

[0102] The calculation formula of the JS divergence loss is as follows:

[0103]

[0104] where $C$ is a constant used to ensure that is non - negative, and

[0105]

[0106]

[0107] where $P(x)=\text{softmax}(T f1 )$ is the probability distribution of the target, and $Q(x)=\text{softmax}(T B1 )$ is the probability distribution of the background.

[0108] The present invention updates the network weights based on the Adam gradient descent method with adaptive learning rate adjustment. In this embodiment, the training is iterated 400 times in total.

[0109] Step 4): For the down - converted radar echo signal to be processed, after obtaining the range - Doppler image through Step 1), input it into the neural network model trained in Step 3). The network outputs whether there is a target in the range - Doppler image and the position of the target, obtaining the detection result.

[0110] Figure 3 is a block diagram of a pulsed - Doppler radar target detection system based on a background - contrast attention mechanism shown in the embodiment. As Figure 3 shown, the system includes:

[0111] A signal preprocessing module, which is used to generate a range-Doppler image corresponding to the radar echo;

[0112] A neural network construction module, which is used to construct a neural network based on a background contrast attention mechanism. The neural network is used to obtain the class information and position information of the target in the range-Doppler image. The neural network includes a multi-scale background learner module based on random masks, a Transformer encoder module based on the background contrast attention mechanism, a Transformer decoder, and a feed-forward network module;

[0113] A neural network training module, which uses the weighted sum of the Hungarian loss and the JS divergence loss of the target detection box as the loss function to train the neural network based on the background contrast attention mechanism to obtain a trained neural network;

[0114] A range-Doppler image detection module, which is used to use the trained neural network to obtain the class information and position information of the target in the range-Doppler image to be detected, and realize the target detection of the range-Doppler image.

[0115] Regarding the device in the above embodiments, the specific ways in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0116] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can refer to the partial descriptions of the method embodiments. The device embodiments described above are only illustrative. For example, the image preprocessing module can be a logical function division, and there can be other division methods in actual implementation. For example, multiple modules can be combined or integrated into another unit. Another point is that the connections between the displayed or discussed modules can be communication connections through some interfaces, which can be electrical or other forms. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present application. Those of ordinary skill in the art can understand and implement it without creative efforts. The following takes a simulated range-Doppler image as an example to illustrate the specific implementation manner to reflect the technical effects of the present invention, and the specific steps in the embodiments will not be repeated.

[0117] Embodiment

[0118] Next, taking the simulation dataset as the research object, the radar target detection method of the present invention is verified. In order to comprehensively compare the detection performance and display the detection results from the perspectives of visualization and quantification, the redefined detection rate p d , false alarm rate p f and average minimum distance d min are used for evaluation, and the formula is expressed as follows:

[0119]

[0120] Among them, is the set of targets to be detected, that is, the ground truth; is the set of detection results, that is, the predicted values; Calculate the Euclidean distance between two targets in the range-Doppler image, T dis represents the distance threshold.

[0121] The radar and target parameter settings of the simulation dataset are as follows: The transmitted signal is a chirp signal, the carrier frequency is 3 GHz, the bandwidth is 5 MHz, the pulse width is 5 μs, the pulse repetition frequency is 2 kHz, and the sampling frequency is 30 MHz; The target distance is between 0.5 km and 5.0 km, and the target speed is between 0 m / s and 40 m / s. Figure 5 The first row is the range-Doppler images and their target ground truth boxes obtained by simulation under signal-to-clutter ratios of -25 dB, -20 dB, -15 dB, and -10 dB respectively.

[0122] Figure 5 Rows 2 to 4 are the detection results obtained by using the embodiments of the present invention and different radar target detection algorithms for the simulation dataset.

[0123] Table 1 Evaluation Metrics for Detection Results of the Simulation Radar Dataset

[0124]

[0125]

[0126] The comparison method CFAR is from: S. Watts, "Cell-averaging CFAR gain in spatially correlated K-distributed clutter", IEE Proceedings-Radar, Sonar and Navigation, vol. 143, pp. 321–327, 1996.

[0127] The comparison method DCNN is from: L. Wang, J. Tang and Q. Liao, "A Study on Radar Target Detection Based on Deep Neural Networks," in IEEE Sensors Letters, vol. 3, no. 3, pp. 1-4, March 2019.

[0128] The comparison method DETR is from: Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S. (2020). End-to-End Object Detection with Transformers. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, JM. (eds) Computer Vision–ECCV 2020. ECCV2020. Lecture Notes in Computer Science(), vol 12346.

[0129] The detection results of the simulation dataset are as Figure 5 shown, and the quantization results are shown in Table 1. In the quantization results, the present invention has achieved the best performance under various signal-to-noise ratio conditions. In scenarios with a relatively high signal-to-noise ratio, each method has good performance, but when the signal-to-noise ratio decreases, this method can still maintain a good detection rate and a low false alarm rate; at the same time, the d min always has the best result, indicating that the target positioning is more accurate.

[0130] The above-described embodiments merely represent several implementation manners of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. For those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A pulse Doppler radar target detection method based on background contrast attention mechanism, characterized in that: The following steps are involved: Step 1) obtaining a pulse Doppler image of the echo according to the echo signal of the radar; Step 2) constructing a neural network based on a background contrast attention mechanism, wherein the neural network is used to obtain the distance and speed information of the target in the range Doppler image, and the neural network includes a multi-scale background learner module based on a random mask, a Transformer encoder module based on a background contrast attention mechanism, a Transformer decoder and a feedforward network module; Step 3) training the neural network based on the background contrast attention mechanism to obtain a trained neural network; Step 4) For the radar echo signal to be processed, the pulse Doppler image is obtained by adopting step 1), and the target category information and position information in the range Doppler image are obtained by using the trained neural network to realize the detection of the radar target.

2. The pulse Doppler radar target detection method based on background contrast attention mechanism according to claim 1 is characterized in that: The step 1) comprises: For the radar echo signal, pulse compression and Doppler processing methods are used to obtain the range Doppler image of the target; represents the two-dimensional target echo sequence after pulse compression, where t m represents the slow time, m represents the number of received pulse echoes, represents fast time, i.e., the time series within a pulse repetition interval; along the fast time dimension Perform fast Fourier transform to obtain the range Doppler image matrix of the target echo The calculation formula is as follows: Where N represents the number of sampling points, k represents the frequency; the range Doppler image matrix As the range Doppler image I RD .

3. The pulse Doppler radar target detection method based on background contrast attention mechanism according to claim 1 is characterized in that: In step 2), the multi-scale background learner module based on random mask obtains the background clutter distribution characteristics from the range Doppler image; the Transformer encoder module based on the background contrast attention mechanism encodes and learns the target features to suppress background clutter interference from the feature level; the Transformer decoder and feedforward network module parse the output of the encoder, convert the initialized target queries into the category information and location information of the target and output them.

4. The pulse Doppler radar target detection method based on background contrast attention mechanism according to claim 3 is characterized in that: The multi-scale background learner module based on random mask obtains background clutter distribution characteristics from the range Doppler image; comprising: For the input range Doppler image I RD , first through the slicing operation, the range Doppler image is converted into a range Doppler image block I patch , Then, a random mask is used to randomly design several blocks in the range Doppler image block to reduce the pollution of the target to the background when learning the background, and the sample T is obtained. patch ; Through three convolution kernels of different sizes, T patch Feature extraction and fusion are performed to achieve multi-scale background feature learning; the process is as follows: T1=DWConv77(T patch ) T2=DWConv55(T patch ) <h2 style=";text-align:left;direction:ltr">T3=DWConv33(T<h2 style=";text-align:left;direction:ltr"> patch <h2 style=";text-align:left;direction:ltr"> ) T concat =Concat(T1,T2,T3) T temp =DWConvBlock(T concat ) T B =LayerNorm(AvgPool(T temp )) Among them, DWConv77, DWConv55, and DWConv33 represent the depth-separable convolution functions with convolution kernel sizes of 7×7, 5×5, and 3×3 respectively; Concat() represents the concatenation function; DWConvBlock() is a feature extraction function composed of three depth-separable convolutions with a convolution kernel size of 3×3, T1, T2, T3, T concat 、T temp is the intermediate result of the corresponding operation; AvgPool() represents the average pooling function, LayerNorm() represents the layer normalization function, is the output feature of the multi-scale background learner module.

5. The pulse Doppler radar target detection method based on background contrast attention mechanism according to claim 3 is characterized in that: The Transformer encoder module based on background contrast attention mechanism includes multiple cascade-connected Transformer encoder blocks, each of which includes a multi-head self-attention module, a background contrast-based attention module, and a multi-layer perceptron module; the Transformer encoder module based on background contrast attention mechanism encodes and learns the target features to suppress background clutter interference from the feature level; specifically includes: For the input range Doppler image I RD , after ResNet-50 and position encoding, we get the feature map F with position encoding RD ; F RD Flattened and passed through an embedding layer to obtain the feature T input to the Transformer encoder block RD1 ; For feature T RD1 Perform multi-head self-attention mechanism calculation, residual connection and normalization operations to obtain the output feature T RD2 ; T RD2 and T output by the multi-scale background learning module based on random masks B Together they are input into the background contrast-based attention mechanism module, T RD2 First, the target feature tensor T for subsequent calculations is obtained through reshaping and average pooling operations. T2 ; T represents the background B After 1×1 convolution, we get T2 Features T with the same channel dimension and the same spatial size B2 ; After processing, the T T2 and T B2 Splicing is performed, and the target and background features are fused through 1×1 convolution and pixel-dimensional maximum pooling to obtain feature T′ c ; For T′ c Upsample to T T1 The same spatial size, and after the sigmoid activation function, generate an attention weight map with strong target significance, which is the same as T T1 After point multiplication and reshaping, residual connection and normalization are performed to obtain feature T f2 ; Last T f2 After mapping by a multi-layer perceptron, the output feature T of the Transformer encoder block is obtained M2 ; The calculation formula is as follows: T M1 =MLP(T f2 ) T M2 =LayerNorm(T M1 +T f2 ) After being processed by multiple cascaded Transformer encoder blocks, the Transformer encoder module based on the background contrast attention mechanism finally outputs the feature V encoder .

6. The pulse Doppler radar target detection method based on background contrast attention mechanism according to claim 5 is characterized in that: The Transformer decoder and feedforward network module parse the output of the encoder, convert the initialized target queries into target category information and position information and output them; wherein the Transformer decoder includes a plurality of cascaded Transformer decoder blocks; The processing of each level of Transformer decoder block includes: For a single target query, Q0 is initialized by position encoding, and then Q0 is calculated by multi-head self-attention mechanism, residual connection and normalization operations to obtain Q2; Using the encoder output V encoder A multi-head attention mechanism is performed with Q2 to obtain Q3, which is then subjected to residual connection and normalization to obtain Q4. After mapping by a multi-layer perceptron, Q6 is obtained; the calculation formula is as follows: Q5=MLP(Q4) Q6=LayerNorm(Q5+Q4) The output Q6 of the upper-level Transformer decoder block is input into the lower-level Transformer decoder block, and the Transformer decoder finally outputs the feature vector V decoder ; The feature vector output by the Transformer decoder finally outputs the target category information through two feedforward networks, namely a classification head and a regression head; the calculation formula is as follows: V class =MLP_Class(V decoder ) V reg =MLP_Regression(V decoder ) Among them, MLP_Class() represents the classification head function, MLP_Regression() represents the regression head function; V class Represents the category confidence, V reg Represents the location coordinates.

7. The pulse Doppler radar target detection method based on background contrast attention mechanism according to claim 5 is characterized in that: In the step 3), The weighted sum of the Hungarian loss and the JS divergence loss is used as the loss function, and the weights of the neural network are updated based on the gradient descent method with adaptively adjusted learning rate; The calculation formula of JS divergence loss function is as follows: Where C is a constant used to ensure is a non-negative number, and Where P(x) = softmax(T f1 ) is the probability distribution of the target, Q(x)=softmax(T B1 ) is the probability distribution of the background.

8. A pulse Doppler radar target detection system based on background contrast attention mechanism for implementing the method of claim 1, characterized in that: include: A signal preprocessing module, which is used to generate a range Doppler image corresponding to the radar echo; A neural network building module, which is used to construct a neural network based on a background contrast attention mechanism, wherein the neural network is used to obtain category information and position information of a target in a range Doppler image, and the neural network includes a multi-scale background learner module based on a random mask, a Transformer encoder module based on a background contrast attention mechanism, a Transformer decoder, and a feedforward network module; A neural network training module is used to train a neural network based on a background contrast attention mechanism to obtain a trained neural network; The range Doppler image detection module is used to obtain the target category information and position information in the range Doppler image to be detected by using the trained neural network, so as to realize the target detection of the range Doppler image.

Images