Airborne radar moving target detection method based on distance-Doppler characteristics

Abstract

The invention discloses an airborne radar moving target detection method and system based on distance-Doppler characteristics, and the method comprises the steps: building an airborne pulse Doppler radar echo signal model, and generating a large number of echo signals through simulation; performing pulse compression processing and frequency domain filtering processing on the machine echo signal to obtain a two-dimensional distance-Doppler image; processing the two-dimensional distance-Doppler image to obtain a processed distance-Doppler image; a generalized target is defined, the generalized target is marked on the processed distance-Doppler image, and a marked data set is generated; training a deep neural network model by using the labeled data set to obtain a trained deep neural network model; and by using the trained deep neural network model, target detection is performed on an actual airborne pulse Doppler radar distance-Doppler image, so that the working performance of an airborne radar system under clutter and interference conditions is greatly improved, and the minimum detectable speed limitation of the system is broken through.

Description

Technical Field

[0001] The present invention relates to the technical field of radar target detection, and more specifically, to an airborne radar moving target detection method based on range-Doppler characteristics. Background Art

[0002] Moving target detection is an important research direction in the field of radar, and is widely used in military and civilian fields such as penetration target warning, battlefield situation awareness, road condition monitoring, and automatic driving of automobiles. Compared with ground-based radars, airborne radars based on aerial moving platforms such as early warning aircraft and drones can overcome the shielding effect caused by the curvature of the earth and mountain occlusion, as well as the influence of low-altitude blind spots. They have the advantages of long detection distance, long warning time, strong maneuverability, and high flexibility. However, when detecting ground / sea or low-altitude moving targets, airborne radars are usually in a downward working state and will receive a large number of electromagnetic waves scattered by the ground / sea surface, namely clutter. The clutter intensity is large, the Doppler spectrum is broadened and changes with the beam incidence angle. Moving targets are often submerged by clutter. Traditional MTD methods and spatial beamforming methods generally cannot effectively suppress clutter. Although space-time adaptive processing methods can be used to combine spatial and temporal two-dimensional information to adaptively suppress clutter, they usually have problems such as insufficient training distance units, high computational complexity, and minimum detectable speed limits. Moreover, as the electromagnetic environment becomes more and more complex, airborne radars often receive intentional or unintentional electromagnetic interference in actual work, which seriously affects the target detection performance of airborne radars.

[0003] In view of this, the present invention proposes an airborne radar moving target detection method based on range-Doppler characteristics to solve the above problems. Summary of the invention

[0004] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solution: a method for detecting moving targets by airborne radar based on range-Doppler characteristics, comprising:

[0005] Step 1: Establish an airborne pulse Doppler radar echo signal model, and simulate and generate a large number of airborne pulse Doppler radar echo signals;

[0006] Step 2: performing pulse compression processing and frequency domain filtering processing on the airborne pulse Doppler radar echo signal to obtain a two-dimensional range-Doppler image;

[0007] Step 3, normalizing, threshold filtering and resizing the two-dimensional range-Doppler image to obtain a processed range-Doppler image;

[0008] Step 4: define all scattering points and multipath echoes of the target as generalized targets, and mark the generalized targets on the processed range-Doppler image to generate a marked data set;

[0009] Step 5: Using the labeled data set, training a deep neural network model to learn the features of the generalized target on the range-Doppler image, and obtaining a trained deep neural network model;

[0010] Step 6: Use the trained deep neural network model to perform target detection on the actual airborne pulse Doppler radar range-Doppler image.

[0011] Furthermore, the method of establishing an airborne pulse Doppler radar echo signal model and simulating and generating a large number of airborne pulse Doppler radar echo signals includes:

[0012] According to the airborne radar detection geometric model, the target echo signal component model is established, and the target is modeled as consisting of P+1 scattering points, each of which has a different normalized Doppler frequency f_p and complex amplitude a_p, p = 0, 1, ..., P, where p = 0, the scattering point represents the target body;

[0013] A clutter component model is established, and the clutter is divided into L range units. Each range unit contains Nc clutter scattering units evenly distributed in azimuth. Each clutter scattering unit has a different normalized Doppler frequency and complex amplitude.

[0014] Establish a multipath signal component model, including specular reflection multipath echo component and diffuse reflection multipath echo component;

[0015] A noise component model is established, and the noise is modeled as additive complex Gaussian white noise with a specific signal-to-noise ratio;

[0016] The target echo signal component model, the clutter component model, the multipath signal component model and the noise component are added together to obtain the airborne pulse Doppler radar echo signal model;

[0017] Under the condition of given airborne radar parameters, the pitch angle, azimuth, signal-to-noise ratio, signal-to-clutter ratio, altitude, speed, number of scattering points, normalized Doppler frequency of scattering points, complex amplitude of scattering points, and amplitude of mirror reflection multipath echo and amplitude of diffuse reflection multipath echo of the target are assumed to obey random distribution within a specific range. According to the airborne pulse Doppler radar echo signal model, simulation is performed to generate a large number of airborne pulse Doppler radar echo signals.

[0018] Furthermore, the method of performing pulse compression processing and frequency domain filtering processing on the airborne pulse Doppler radar echo signal to obtain a two-dimensional range-Doppler image includes:

[0019] Construct the transfer function h(t) and amplitude weighting vector w_r of the matched filter, where t is the fast time vector, perform pulse compression processing on the k-th airborne pulse Doppler radar echo signal within a coherent processing interval (CPI), and obtain the pulse compression output y(r,k), where r=ct / 2 is the distance vector, c is the speed of light, k=1,2,...,K, and K is the number of pulses in a coherent processing interval;

[0020] The K pulse compression outputs corresponding to each range unit are subjected to frequency domain filtering using fast Fourier transform, and the absolute value is taken to obtain a two-dimensional range-Doppler image RD(r,f), where f is the Doppler vector.

[0021] Furthermore, the method of performing normalization, threshold filtering and size conversion processing on the two-dimensional range-Doppler image to obtain the processed range-Doppler image includes:

[0022] The two-dimensional range-Doppler image RD(r, f) is normalized. Specifically, the maximum value RD_Max of the two-dimensional range-Doppler image RD(r, f) is calculated, and the two-dimensional range-Doppler image RD(r, f) is divided by RD_Max to obtain a normalized range-Doppler image RD_norm(r, f);

[0023] Set the threshold Th and perform hard threshold filtering on RD_norm(r,f). Specifically, if RD_norm(r,f) is greater than or equal to Th, the range-Doppler image after threshold filtering RD_filt(r,f)=RD_norm(r,f); otherwise RD_filt(r,f)=0;

[0024] The range-Doppler image RD_filt(r,f) after threshold filtering is resized to obtain a processed range-Doppler image RD_final.

[0025] Furthermore, the method of defining all scattering points and multipath echoes of the target as a generalized target, annotating the generalized target on the processed range-Doppler image, and generating the annotated data set includes:

[0026] On the processed range-Doppler image RD_final, all scattering points and multipath echoes of the target are defined as generalized targets, and the area of ​​the generalized target is manually marked; the generalized target area is marked with a bounding box, each bounding box includes the target's range [rmin, rmax] and Doppler frequency range [fmin, fmax]; where rmin is the minimum distance, rmax is the maximum distance, fmin is the minimum Doppler frequency, and fmax is the maximum Doppler frequency;

[0027] A corresponding annotation file is created for each processed range-Doppler image RD_final; in the annotation file, the coordinates of all annotated bounding boxes in the image are recorded, and the coordinates include the range range and the Doppler frequency range;

[0028] All processed range-Doppler images RD_final and their corresponding annotation files form an annotation dataset.

[0029] Furthermore, the method of using the labeled data set to train a deep neural network model so that it learns the features of the generalized target on the range-Doppler image and obtaining the trained deep neural network model includes:

[0030] Define the architecture of the deep neural network model using a fully convolutional network; initialize the hyperparameters of the deep neural network model, including network weights, network bias, input image size, batch size, maximum number of iterations, and learning rate; define the loss function of the deep neural network model; define the optimizer;

[0031] Perform data enhancement on the labeled data set; obtain the enhanced data set, and divide the enhanced data set into a training set, a validation set, and a test set;

[0032] Read the processed range-Doppler image RD_final and the annotated bounding box from the training set in batches; use the processed range-Doppler image RD_final as training data and the annotated bounding box as a label;

[0033] Input the training data into the deep neural network model and use forward propagation to calculate the value of the loss function; back propagate according to the value of the loss function to update the network weights; evaluate the deep neural network model on the validation set at fixed iteration times and save the optimal model weights; repeat until the deep neural network model converges or reaches the preset maximum number of iterations;

[0034] Evaluate the target detection performance of the optimal model on the test set and calculate the evaluation results, which are precision, recall or mAP; further tune the model based on the evaluation results; and obtain a trained deep neural network model.

[0035] Furthermore, the method of using the trained deep neural network model to perform target detection on an actual airborne pulse Doppler radar range-Doppler image includes:

[0036] The echo data received by the actual airborne pulse Doppler radar is subjected to pulse compression and frequency domain filtering to obtain a two-dimensional range-Doppler image RD_new; the range-Doppler image RD_new is normalized, threshold filtered and resized to obtain a processed range-Doppler image RD_final_new; RD_final_new is input into the trained deep neural network model for forward propagation operation; the deep neural network model outputs the detection results, which include the coordinates of the predicted target bounding box and the confidence score of each predicted target bounding box;

[0037] Set a confidence threshold to remove predicted target bounding boxes whose confidence scores are lower than the confidence threshold; perform non-maximum suppression to remove overlapping predicted target bounding boxes;

[0038] Render the final target detection results on RD_final_new, which include the target bounding box and confidence score.

[0039] An airborne radar moving target detection system based on range-Doppler characteristics, which is used to implement the airborne radar moving target detection method based on range-Doppler characteristics, comprising:

[0040] The echo signal generation module is used to establish an airborne pulse Doppler radar echo signal model and simulate and generate a large number of airborne pulse Doppler radar echo signals;

[0041] A range-Doppler imaging module is used to perform pulse compression processing and frequency domain filtering processing on the airborne pulse Doppler radar echo signal to obtain a two-dimensional range-Doppler image;

[0042] An image preprocessing module, used for normalizing, threshold filtering and size conversion processing the two-dimensional range-Doppler image to obtain a processed range-Doppler image;

[0043] A labeling data set generation module is used to define all scattering points and multipath echoes of the target as generalized targets, label the generalized targets on the processed range-Doppler image, and generate a labeling data set;

[0044] A deep neural network training module is used to train a deep neural network model using the labeled data set so that the model learns the characteristics of the generalized target on the range-Doppler image and obtains a trained deep neural network model;

[0045] The moving target detection module is used to perform target detection on an actual airborne pulse Doppler radar range-Doppler image using the trained deep neural network model.

[0046] The technical effects and advantages of the present invention's airborne radar moving target detection method based on range-Doppler characteristics are as follows:

[0047] The present invention uniformly defines all scattering points and multipath echoes of the target as generalized targets, comprehensively considers the complex extended features of the target on the range-Doppler image, thereby avoiding the performance loss caused by the traditional algorithm ignoring the secondary scattering points and multipath effects; the feature patterns of the generalized targets are automatically learned by the deep neural network, without the need to manually design a feature extractor, and the rich information contained in the range-Doppler image can be fully mined to achieve high-precision target detection; more importantly, the target detection process is endowed with a strong anti-interference ability; the deep learning model, with its excellent nonlinear mapping ability, can effectively resist complex environmental interference such as clutter and noise, and ensure the high robustness of target detection; even in a low signal-to-noise ratio and harsh clutter environment, the method can still achieve excellent detection performance, greatly improving the adaptability of the airborne radar system under harsh working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A schematic diagram of a method for detecting moving targets by airborne radar based on range-Doppler characteristics of the present invention;

[0049] Figure 2 The present invention is a schematic diagram of an airborne radar moving target detection system based on range-Doppler characteristics. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] Example 1

[0052] See also Figure 1 As shown, the method for detecting moving targets by airborne radar based on range-Doppler characteristics described in this embodiment includes:

[0053] Step 1: Establish an airborne pulse Doppler radar echo signal model, and simulate and generate a large number of airborne pulse Doppler radar echo signals;

[0054] Step 2: performing pulse compression processing and frequency domain filtering processing on the airborne pulse Doppler radar echo signal to obtain a two-dimensional range-Doppler image;

[0055] Step 3, normalizing, threshold filtering and resizing the two-dimensional range-Doppler image to obtain a processed range-Doppler image;

[0056] Step 4: define all scattering points and multipath echoes of the target as generalized targets, annotate the generalized targets on the processed range-Doppler image, and generate annotated data sets;

[0057] Step 5: Using the labeled data set, training a deep neural network model to learn the features of the generalized target on the range-Doppler image, and obtaining a trained deep neural network model;

[0058] Step 6: Use the trained deep neural network model to perform target detection on the actual airborne pulse Doppler radar range-Doppler image.

[0059] The airborne pulse Doppler radar echo signal model is established, and the methods of simulating and generating a large number of airborne pulse Doppler radar echo signals include:

[0060] According to the airborne radar detection geometric model, the target echo signal component model is established, and the target is modeled as consisting of P+1 scattering points, each of which has a different normalized Doppler frequency f_p and complex amplitude a_p, p = 0, 1, ..., P, where p = 0, the scattering point represents the target body;

[0061] A clutter component model is established, and the clutter is divided into L range units. Each range unit contains Nc clutter scattering units evenly distributed in azimuth. Each clutter scattering unit has a different normalized Doppler frequency and complex amplitude.

[0062] Establish a multipath signal component model, including specular reflection multipath echo component and diffuse reflection multipath echo component;

[0063] A noise component model is established, and the noise is modeled as additive complex Gaussian white noise with a specific signal-to-noise ratio;

[0064] The target echo signal component model, the clutter component model, the multipath signal component model and the noise component are added together to obtain the airborne pulse Doppler radar echo signal model;

[0065] Under the condition of given airborne radar parameters, the pitch angle, azimuth, signal-to-noise ratio, signal-to-clutter ratio, altitude, speed, number of scattering points, normalized Doppler frequency of scattering points, complex amplitude of scattering points, and amplitude of mirror reflection multipath echo and amplitude of diffuse reflection multipath echo of the target are assumed to obey random distribution within a specific range. According to the airborne pulse Doppler radar echo signal model, simulation is performed to generate a large number of airborne pulse Doppler radar echo signals.

[0066] Considering all possible components, the kth airborne pulse Doppler radar echo signal within a coherent processing interval (CPI):

[0067] X(t,t k )=S(t,t k )+C(t,t k )+M(t,t k )+N(t,t k ) (1)

[0068] In the formula, S(t,t k )、C(t,t k )、M(t,t k ) and N(t,t k ) represent the target echo signal component, clutter component, multipath signal component and noise component respectively, t represents the fast time, t k =(k-1)T r Indicates slow time, k = 1, 2, ..., K, T r represents the pulse repetition period, and K is the number of CPI pulses.

[0069] For the target signal component model, given a target consisting of P+1 scattering points, the three-dimensional coordinates, azimuth and elevation angles of the target body are [x0, y0, h], θ0 and Considering that the CPI is short and the target size is usually smaller than the radar resolution unit, the target signal model is:

[0070]

[0071] Where, τ0=2R0 / c, R0=||u rt ||2,u rt =[x0,y0,hH],ξ p is the complex amplitude of the pth target scattering point (the 0th scattering point is the target body), represents the normalized Doppler frequency of the pth target scattering point, H is the airborne radar altitude, represents the three-dimensional velocity of the pth target scattering point, T is the pulse width of the airborne radar, λ is the signal wavelength, V is the carrier speed, α is the angle between the airborne radar array and the radar flight direction, γ is the modulation frequency, and c is the speed of light.

[0072] For the clutter model, the clutter is divided into L range units, each of which contains Nc clutter scatterer patches (CSPs) uniformly distributed in the azimuth range. The azimuth and elevation angles of the i-th CSP in the l-th range unit are θi and Then the clutter signal model is:

[0073]

[0074] In the formula, ξ l,i represents the complex amplitude of the ith CSP in the lth distance unit, τ l =2R l / c, represents the lth clutter distance, f l,i =2Vcosφ l,i T r / λ represents the normalized Doppler frequency of the ith CSP in the lth range unit,

[0075] For the multipath signal model, the specular reflection multipath echo and the diffuse reflection multipath echo are comprehensively considered, among which the specular reflection multipath echo is modeled based on the four-path model, which is expressed as:

[0076]

[0077] In the formula, is the s-th mirror echo amplitude,

[0078] and denote the normalized Doppler frequencies of the mirror point relative to the radar and the target, respectively, represents the pitch angle of the mirror point, u st =[x0-x s ,y0-y s ,-h],x s and s represents the x- and y-coordinates of the mirror point,

[0079] The diffuse reflection multipath echo signal can be expressed as:

[0080]

[0081] In the formula, and represent the time delay, normalized Doppler frequency and complex amplitude of the (i, d)th diffuse reflection of the scintillation surface respectively.

[0082] Given the ground distance r between the (i, d)th diffuse reflection point on the glittering surface and the radar d , and its distance from the radar-target line of sight projection satisfies:

[0083]

[0084] In the formula, is the ground distance between the radar and the target, and β0 represents the maximum slope value of the glittering surface. d and After that, the three-dimensional coordinates of the (i, d)th diffuse reflection point can be calculated, and on this basis, and

[0085] The method of performing pulse compression processing and frequency domain filtering processing on the airborne pulse Doppler radar echo signal to obtain a two-dimensional range-Doppler image includes:

[0086] Construct the transfer function h(t) and amplitude weighting vector w_r of the matched filter, where t is the fast time vector, perform pulse compression processing on the k-th airborne pulse Doppler radar echo signal within a coherent processing interval (CPI), and obtain the pulse compression output y(r,k), where r=ct / 2 is the distance vector, c is the speed of light, k=1,2,...,K, and K is the number of pulses in a coherent processing interval;

[0087] The K pulse compression outputs corresponding to each range unit are subjected to frequency domain filtering using fast Fourier transform, and the absolute value is taken to obtain a two-dimensional range-Doppler image RD(r,f), where f is the Doppler vector.

[0088] Specifically, let the amplitude weighted vector of matched filtering be w r , then the pulse compression result of the kth airborne pulse Doppler radar echo signal is:

[0089] y(r,k)=X(t,t k )*[w r eh(t)](7)

[0090] Where h(t) represents the transfer function of the matched filter, t represents the fast time vector, r = ct / 2 represents the distance vector, e represents the Hadamard product, * represents the convolution, and X(t,t k ) represents the kth airborne pulse Doppler radar echo signal vector.

[0091] After pulse compression processing, the K pulse compression outputs corresponding to each range unit are subjected to frequency domain filtering using Fast Fourier Transform, and then the absolute value is taken to obtain a two-dimensional range-Doppler image, which is expressed as:

[0092]

[0093] Where f is the Doppler frequency vector.

[0094] The method of normalizing, threshold filtering and size conversion processing the two-dimensional range-Doppler image to obtain the processed range-Doppler image includes:

[0095] The two-dimensional range-Doppler image RD(r, f) is normalized. Specifically, the maximum value RD_Max of the two-dimensional range-Doppler image RD(r, f) is calculated, and the two-dimensional range-Doppler image RD(r, f) is divided by RD_Max to obtain a normalized range-Doppler image RD_norm(r, f);

[0096] Set the threshold Th and perform hard threshold filtering on RD_norm(r,f). Specifically, if RD_norm(r,f) is greater than or equal to Th, the range-Doppler image after threshold filtering RD_filt(r,f)=RD_norm(r,f); otherwise RD_filt(r,f)=0, where the value of Th is determined according to the specific situation and can usually be set between 0.1-0.5;

[0097] The range-Doppler image RD_filt(r,f) after threshold filtering is resized to obtain the processed range-Doppler image RD_final for subsequent target detection.

[0098] All scattering points and multipath echoes of the target are defined as generalized targets, and the generalized targets are annotated on the processed range-Doppler image. The methods for generating the annotated data set include:

[0099] On the processed range-Doppler image RD_final, all scattering points and multipath echoes of the target are defined as generalized targets, and the area of ​​the generalized target is manually marked; the generalized target area is marked with a bounding box, each bounding box includes the target's range [rmin, rmax] and Doppler frequency range [fmin, fmax]; where rmin is the minimum distance, rmax is the maximum distance, fmin is the minimum Doppler frequency, and fmax is the maximum Doppler frequency;

[0100] A corresponding annotation file is created for each processed range-Doppler image RD_final; in the annotation file, the coordinates of all annotated bounding boxes in the image are recorded, and the coordinates include the range range and the Doppler frequency range;

[0101] All processed range-Doppler images RD_final and their corresponding annotation files form an annotation dataset.

[0102] It should be noted that since the generalized target includes all the scattering points and multipath echoes of the target, they need to be included in the annotation; the annotation quality is crucial to the subsequent training of the target detection model, so the annotation work needs to be very cautious and meticulous.

[0103] The method of using the labeled data set to train a deep neural network model so that it learns the features of the generalized target on the range-Doppler image and obtains the trained deep neural network model includes:

[0104] Define the architecture of the deep neural network model using a fully convolutional network, such as YOLO-v2 or YOLO-v3; initialize the hyperparameters of the deep neural network model, including network weights, network bias, input image size, batch size, maximum number of iterations, and learning rate; define the loss function of the deep neural network model, such as cross entropy loss, focalloss, etc.; define optimizers, such as SGD, Adam, etc.

[0105] Perform data augmentation on the labeled data set, such as flipping, rotating, scaling, etc., to expand the data and obtain an enhanced data set, which is then divided into a training set, a validation set, and a test set;

[0106] Read the processed range-Doppler images RD_final and annotated bounding boxes from the training set in batches; use the processed range-Doppler images RD_final as training data and the annotated bounding boxes as labels.

[0107] Input the training data into the deep neural network model and use forward propagation to calculate the value of the loss function; backpropagate according to the value of the loss function to update the network weights; evaluate the deep neural network model on the validation set every fixed number of iterations and save the optimal model weights; repeat until the deep neural network model converges or reaches the preset maximum number of iterations.

[0108] The target detection performance of the optimal model is evaluated on the test set, and the evaluation results are calculated. The evaluation results are indicators such as precision, recall or mAP. According to the evaluation results, the model is further tuned (adjusting hyperparameters). The constructed deep neural network model is obtained for target detection on range-Doppler images.

[0109] During the training process, various hyperparameters need to be adjusted according to the specific situation to obtain the best model performance.

[0110] In the specific process, according to the established airborne pulse Doppler radar echo signal model, 500 radar echo signal matrices and their corresponding two-dimensional range-Doppler images were simulated. After the generalized targets were labeled, these two-dimensional range-Doppler images were randomly divided into two data sets, 80% for training and 20% for testing. The network input size was 672×672, the number of classes was 1, the number of anchor boxes was 5, the batch size was 20, the maximum number of iterations was 200, and the initial learning rate was 10. -4 , using the Adam optimizer to train YOLO-v2, the loss (Loss) and root mean square error (RMSE) of training and testing are obtained. Both Loss and RMSE gradually decrease, and the test value remains almost unchanged after about 100 rounds, indicating that YOLO-v2 has reached convergence.

[0111] The trained YOLO-v2 algorithm is used to detect moving targets on the test data set. Under different clutter distribution, signal-to-noise ratio and target parameter conditions, the method of the present invention can always detect moving targets because it can effectively learn the characteristics of the two-dimensional range-Doppler image of the target. In order to quantitatively evaluate the target detection performance of the method of the present invention under different SNR conditions, SCR=-30dB is kept unchanged, and the SNR is increased from 0dB to 20dB in steps of 5dB. 100 test two-dimensional range-Doppler images are randomly generated under each SNR condition. When SNR≥5dB, the method of the present invention has higher target detection performance, and the average detection accuracy is higher than 95%. When SNR=0dB, the target may be submerged by noise, and the performance of the method of the present invention is reduced, with an average detection accuracy of 85.94%.

[0112] Using the trained deep neural network model, target detection methods on actual airborne pulse Doppler radar range-Doppler images include:

[0113] The echo data received by the actual airborne pulse Doppler radar is subjected to pulse compression and frequency domain filtering to obtain a two-dimensional range-Doppler image RD_new; the range-Doppler image RD_new is normalized, threshold filtered and resized to obtain a processed range-Doppler image RD_final_new; RD_final_new is input into the trained deep neural network model for forward propagation operation; the deep neural network model outputs the detection results, which include the coordinates of the predicted target bounding box and the confidence score of each predicted target bounding box;

[0114] Set a confidence threshold to remove predicted target bounding boxes whose confidence scores are lower than the confidence threshold; perform non-maximum suppression to remove overlapping predicted target bounding boxes;

[0115] Render the final target detection results on RD_final_new, which include the target bounding box and confidence score.

[0116] It should be noted that in practical applications, the deep neural network model may need to be improved and optimized to meet the requirements of real-time and computing resources.

[0117] In this embodiment, all scattering points and multipath echoes of the target are uniformly defined as generalized targets, and the complex extended features of the target on the range-Doppler image are fully considered, thereby avoiding the performance loss caused by the traditional algorithm ignoring secondary scattering points and multipath effects; the characteristic pattern of the generalized target is automatically learned by the deep neural network, without the need to manually design a feature extractor, and the rich information contained in the range-Doppler image can be fully mined to achieve high-precision target detection; more importantly, the target detection process is endowed with a strong anti-interference ability; the deep learning model, with its excellent nonlinear mapping ability, can effectively resist complex environmental interference such as clutter and noise, and ensure high robustness of target detection; even in a low signal-to-noise ratio and harsh clutter environment, the method can still achieve excellent detection performance, greatly improving the adaptability of the airborne radar system under harsh working conditions.

[0118] Example 2

[0119] See also Figure 2 As shown, the part not described in detail in this embodiment is described in Example 1, which provides an airborne radar moving target detection system based on range-Doppler characteristics, including:

[0120] The echo signal generation module is used to establish an airborne pulse Doppler radar echo signal model and simulate and generate a large number of airborne pulse Doppler radar echo signals;

[0121] A range-Doppler imaging module is used to perform pulse compression processing and frequency domain filtering processing on the airborne pulse Doppler radar echo signal to obtain a two-dimensional range-Doppler image;

[0122] An image preprocessing module, used for normalizing, threshold filtering and size conversion processing the two-dimensional range-Doppler image to obtain a processed range-Doppler image;

[0123] A labeling data set generation module is used to define all scattering points and multipath echoes of the target as generalized targets, label the generalized targets on the processed range-Doppler image, and generate a labeling data set;

[0124] A deep neural network training module is used to train a deep neural network model using the labeled data set so that the model learns the characteristics of the generalized target on the range-Doppler image and obtains a trained deep neural network model;

[0125] The moving target detection module is used to perform target detection on the actual airborne pulse Doppler radar range-Doppler image using the trained deep neural network model; each module is connected by wire and / or wireless means to achieve data transmission between modules.

[0126] Example 3

[0127] The present embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the operation mode of the above-mentioned method for detecting moving targets by an airborne radar based on range-Doppler characteristics is implemented.

[0128] Since the electronic device introduced in this embodiment is an electronic device used to implement an airborne radar moving target detection method based on range-Doppler characteristics in the embodiment of the present application, based on the airborne radar moving target detection method based on range-Doppler characteristics introduced in the embodiment of the present application, a person skilled in the art can understand the specific implementation of the electronic device of the present embodiment and its various variations, so how the electronic device implements the method in the embodiment of the present application is not described in detail here. As long as a person skilled in the art implements an electronic device used by an airborne radar moving target detection method based on range-Doppler characteristics in the embodiment of the present application, it belongs to the scope of protection of the present application.

[0129] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0130] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technical users in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A method for detecting moving targets by airborne radar based on range-Doppler characteristics, characterized in that: include: Step 1: Establish an airborne pulse Doppler radar echo signal model, and simulate and generate a large number of airborne pulse Doppler radar echo signals; Step 2: performing pulse compression processing and frequency domain filtering processing on the airborne pulse Doppler radar echo signal to obtain a two-dimensional range-Doppler image; Step 3, normalizing, threshold filtering and resizing the two-dimensional range-Doppler image to obtain a processed range-Doppler image; Step 4: define all scattering points and multipath echoes of the target as generalized targets, and mark the generalized targets on the processed range-Doppler image to generate a marked data set; Step 5: Using the labeled data set, training a deep neural network model to learn the features of the generalized target on the range-Doppler image, and obtaining a trained deep neural network model; Step 6: Utilize the trained deep neural network model to perform target detection on an actual airborne pulse Doppler radar range-Doppler image.

2. The method for detecting moving targets by airborne radar based on range-Doppler characteristics according to claim 1, characterized in that: The method of establishing an airborne pulse Doppler radar echo signal model and simulating and generating a large number of airborne pulse Doppler radar echo signals includes: According to the airborne radar detection geometric model, the target echo signal component model is established, and the target is modeled as consisting of P+1 scattering points, each of which has a different normalized Doppler frequency f_p and complex amplitude a_p, p = 0, 1, ..., P, where p = 0, the scattering point represents the target body; A clutter component model is established, and the clutter is divided into L range units. Each range unit contains Nc clutter scattering units evenly distributed in azimuth. Each clutter scattering unit has a different normalized Doppler frequency and complex amplitude. Establish a multipath signal component model, including specular reflection multipath echo component and diffuse reflection multipath echo component; A noise component model is established, and the noise is modeled as additive complex Gaussian white noise with a specific signal-to-noise ratio; The target echo signal component model, the clutter component model, the multipath signal component model and the noise component are added together to obtain the airborne pulse Doppler radar echo signal model; Under the condition of given airborne radar parameters, the pitch angle, azimuth, signal-to-noise ratio, signal-to-clutter ratio, altitude, speed, number of scattering points, normalized Doppler frequency of scattering points, complex amplitude of scattering points, and amplitude of mirror reflection multipath echo and amplitude of diffuse reflection multipath echo of the target are assumed to obey random distribution within a specific range. According to the airborne pulse Doppler radar echo signal model, simulation is performed to generate a large number of airborne pulse Doppler radar echo signals.

3. The method for detecting moving targets by airborne radar based on range-Doppler characteristics according to claim 2, characterized in that: The method of performing pulse compression processing and frequency domain filtering processing on the airborne pulse Doppler radar echo signal to obtain a two-dimensional range-Doppler image includes: Construct the transfer function h(t) and amplitude weighting vector w_r of the matched filter, where t is the fast time vector, perform pulse compression processing on the k-th airborne pulse Doppler radar echo signal within a coherent processing interval (CPI), and obtain the pulse compression output y(r,k), where r=ct / 2 is the distance vector, c is the speed of light, k=1,2,...,K, and K is the number of pulses in a coherent processing interval; The K pulse compression outputs corresponding to each range unit are subjected to frequency domain filtering using fast Fourier transform, and the absolute value is taken to obtain a two-dimensional range-Doppler image RD(r,f), where f is the Doppler vector.

4. The method for detecting moving targets by airborne radar based on range-Doppler characteristics according to claim 3, characterized in that: The method of performing normalization, threshold filtering and size conversion processing on the two-dimensional range-Doppler image to obtain the processed range-Doppler image includes: The two-dimensional range-Doppler image RD(r, f) is normalized. Specifically, the maximum value RD_Max of the two-dimensional range-Doppler image RD(r, f) is calculated, and the two-dimensional range-Doppler image RD(r, f) is divided by RD_Max to obtain a normalized range-Doppler image RD_norm(r, f); Set the threshold Th and perform hard threshold filtering on RD_norm(r,f). Specifically, if RD_norm(r,f) is greater than or equal to Th, the range-Doppler image after threshold filtering RD_filt(r,f)=RD_norm(r,f); otherwise RD_filt(r,f)=0; The range-Doppler image RD_filt(r,f) after threshold filtering is resized to obtain a processed range-Doppler image RD_final.

5. The method for detecting moving targets by airborne radar based on range-Doppler characteristics according to claim 4, characterized in that: The method of defining all scattering points and multipath echoes of the target as a generalized target, annotating the generalized target on the processed range-Doppler image, and generating annotated data sets includes: On the processed range-Doppler image RD_final, all scattering points and multipath echoes of the target are defined as generalized targets, and the area of ​​the generalized target is manually marked; the generalized target area is marked with a bounding box, each bounding box includes the target's range [rmin, rmax] and Doppler frequency range [fmin, fmax]; where rmin is the minimum distance, rmax is the maximum distance, fmin is the minimum Doppler frequency, and fmax is the maximum Doppler frequency; A corresponding annotation file is created for each processed range-Doppler image RD_final; in the annotation file, the coordinates of all annotated bounding boxes in the image are recorded, and the coordinates include the range range and the Doppler frequency range; All processed range-Doppler images RD_final and their corresponding annotation files form an annotation dataset.

6. The method for detecting moving targets by airborne radar based on range-Doppler characteristics according to claim 5, characterized in that: The method of using the labeled data set to train the deep neural network model so that it learns the features of the generalized target on the range-Doppler image and obtains the trained deep neural network model includes: Define the architecture of the deep neural network model using a fully convolutional network; initialize the hyperparameters of the deep neural network model, including network weights, network bias, input image size, batch size, maximum number of iterations, and learning rate; define the loss function of the deep neural network model; define the optimizer; Perform data enhancement on the labeled data set; obtain the enhanced data set, and divide the enhanced data set into a training set, a validation set, and a test set; Read the processed range-Doppler image RD_final and the annotated bounding box from the training set in batches; use the processed range-Doppler image RD_final as training data and the annotated bounding box as a label; Input the training data into the deep neural network model and use forward propagation to calculate the value of the loss function; back propagate according to the value of the loss function to update the network weights; evaluate the deep neural network model on the validation set at fixed iteration times and save the optimal model weights; repeat until the deep neural network model converges or reaches the preset maximum number of iterations; Evaluate the target detection performance of the optimal model on the test set and calculate the evaluation results, which are precision, recall or mAP; further tune the model based on the evaluation results; and obtain a trained deep neural network model.

7. The method for detecting moving targets by airborne radar based on range-Doppler characteristics according to claim 6, characterized in that: The method of using the trained deep neural network model to perform target detection on an actual airborne pulse Doppler radar range-Doppler image includes: The echo data received by the actual airborne pulse Doppler radar is subjected to pulse compression and frequency domain filtering to obtain a two-dimensional range-Doppler image RD_new; the range-Doppler image RD_new is normalized, threshold filtered and resized to obtain a processed range-Doppler image RD_final_new; RD_final_new is input into the trained deep neural network model for forward propagation operation; the deep neural network model outputs the detection results, which include the coordinates of the predicted target bounding box and the confidence score of each predicted target bounding box; Set a confidence threshold to remove predicted target bounding boxes whose confidence scores are lower than the confidence threshold; perform non-maximum suppression to remove overlapping predicted target bounding boxes; Render the final target detection results on RD_final_new, which include the target bounding box and confidence score.

8. An airborne radar moving target detection system based on range-Doppler characteristics, which is used to implement an airborne radar moving target detection method based on range-Doppler characteristics as claimed in any one of claims 1 to 7, characterized in that: include: The echo signal generation module is used to establish an airborne pulse Doppler radar echo signal model and simulate and generate a large number of airborne pulse Doppler radar echo signals; A range-Doppler imaging module is used to perform pulse compression processing and frequency domain filtering processing on the airborne pulse Doppler radar echo signal to obtain a two-dimensional range-Doppler image; An image preprocessing module, used for normalizing, threshold filtering and size conversion processing the two-dimensional range-Doppler image to obtain a processed range-Doppler image; A labeling data set generation module is used to define all scattering points and multipath echoes of the target as generalized targets, label the generalized targets on the processed range-Doppler image, and generate a labeling data set; A deep neural network training module is used to train a deep neural network model using the labeled data set so that the model learns the characteristics of the generalized target on the range-Doppler image and obtains a trained deep neural network model; The moving target detection module is used to perform target detection on an actual airborne pulse Doppler radar range-Doppler image using the trained deep neural network model.

Images