Radar target detection method based on multi-dimensional time-frequency characteristics and multi-stage decision

Abstract

The application discloses a radar target detection method based on multi-dimensional time-frequency characteristics and multi-stage decision, which comprises the following steps: constructing a Hankel matrix singular value decomposition for radar echo data, projecting a target subspace, suppressing clutter preprocessing, reconstructing two-dimensional echo data into one-dimensional echo data after preprocessing; calculating a normalized smooth pseudo-Vignette-Wille distribution, extracting four-dimensional time-frequency characteristics of ridge accumulation, connected region number, maximum connected region size and time-frequency distribution flux; calculating the mean and standard deviation of each characteristic, adaptively determining a decision threshold according to the 2.5sigma rule, and determining a distance unit with at least three time-frequency characteristics as a target distance unit. The application can be used for low, slow and small target detection in a complex ground clutter background, has the advantages of reducing missed detection and false alarm caused by a single characteristic, and improving detection accuracy.

Description

Technical Field

[0001] This invention belongs to the field of electronic digital data processing technology, and further relates to a radar target detection method based on multi-dimensional time-frequency characteristics and multi-level decision-making in the field of radar signal data processing technology. This invention can be used to accurately distinguish weak targets from radar echo signals in environments with complex and variable ground clutter backgrounds. Background Technology

[0002] The core task of radar target detection is to determine whether a target exists in the received echo, which is essentially a binary hypothesis testing problem. Actual radar received echoes typically consist of target echoes, ground clutter, and system noise. Ground clutter exhibits significant non-stationarity, spatial correlation, and temporal correlation. Especially in complex environments such as mountainous terrain, building clusters, and vegetated areas, ground clutter energy is strong and fluctuates dramatically, easily masking weak target echoes and thus degrading the detection performance of low-profile, slow-moving, and small weak targets.

[0003] Xi'an University of Electronic Science and Technology disclosed a sea clutter suppression method based on time-spectrum energy distribution and singular value decomposition in its patent application document "A sea clutter suppression method based on time-spectrum energy distribution and singular value decomposition" (application number CN 202310241578.5, application publication number CN 116224 277 A). This method first performs a short-time Fourier transform on the echo signal containing clutter to obtain the time spectrum of the echo signal and calculates the instantaneous energy at each time frequency point. Then, it divides the frequency interval of the time spectrum into equal parts, calculates the energy proportion of each frequency sub-interval to the total frequency interval at each time point, and selects the average amplitude of the frequency sub-interval with the largest energy proportion as the filtering threshold to filter the corresponding time frequency point. Next, it performs an inverse short-time Fourier transform on the filtered time spectrum and performs singular value decomposition on the resulting echo signal matrix. From the singular value decomposition results, it selects the set of singular values ​​representing clutter and constructs a signal subspace matrix orthogonal to the clutter subspace. Finally, it projects the echo signal matrix onto this signal subspace matrix to obtain the clutter-suppressed echo signal. However, this method still has shortcomings. The above method focuses on strong clutter suppression of the clutter echo, aiming to obtain a clutter-suppressed echo signal. For target detection problems in complex environments, simply performing clutter suppression preprocessing is still insufficient to utilize the differences between the target and background clutter in terms of time-frequency energy distribution, regional connectivity structure, and other characteristics. Therefore, it is still necessary to further develop multi-dimensional time-frequency feature extraction and multi-level decision methods based on clutter suppression processing, so as to improve the accuracy and robustness of weak target detection under complex ground clutter backgrounds. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of the prior art by proposing a radar target detection method based on multi-dimensional time-frequency features and multi-level decision-making. This method aims to solve the problems of high false alarm probability, high missed detection probability, and poor detection robustness of the prior art in complex ground clutter environments, which rely solely on a single time-frequency feature to make decisions on targets and background clutter.

[0005] The technical approach to achieving the objective of this invention is as follows: First, the radar echo data is subspaced using Hankel matrix singular value decomposition, thereby performing strong ground clutter background filtering preprocessing. Then, the processed two-dimensional echo data is reconstructed into one-dimensional echo data. This invention calculates a normalized smooth pseudo-Wigner-Willi distribution on the clutter-suppressed data and extracts four complementary time-frequency features: ridge accumulation, number of connected regions, and time-frequency distribution flux of the largest connected region size. This solves the problem in existing technologies where complex clutter may only be prominent in one or two time-frequency features, such as strong clutter easily causing ridge accumulation or anomalies in the largest connected region size, but it is usually difficult for it to simultaneously present target characteristics in multiple aspects such as trajectory continuity, regional structure, and dynamic changes. This invention performs anomaly detection on each feature based on the 2.5σ criterion and uses a comprehensive decision criterion of "at least three features being simultaneously abnormal" to determine the target range cell. This invention overcomes the shortcomings of existing technologies, such as susceptibility to interference from single features and poor adaptability of fixed thresholds, by using multi-dimensional time-frequency features and performing multi-level joint decision-making. This significantly improves the detection probability and robustness in complex terrain backgrounds.

[0006] To achieve the above objectives, the specific implementation steps of the present invention include the following:

[0007] Step 1: Subspace projection is performed using singular value decomposition of the Hankel matrix to preprocess the radar echo data by filtering out strong ground clutter. Then, the processed two-dimensional echo data is reconstructed into one-dimensional echo data.

[0008] Step 2: Calculate the normalized smooth pseudo-Wigner-Willi distribution of the clutter-suppressed signal and extract the time-frequency features in four dimensions;

[0009] Step 3: Based on the 2.5σ criterion, anomaly judgment thresholds are automatically constructed according to the mean and variance of the time-frequency features in the current data. Candidate target feature judgments are made for each time-frequency feature, and the final target distance unit is determined according to the multi-level judgment logic.

[0010] Step 4: Mark the detection results of different time-frequency features on each distance cell.

[0011] Furthermore, the step of performing singular value decomposition on the radar echo data matrix is ​​as follows:

[0012] The first step is to analyze the radar echo data matrix. Extracting the first The total number of pulses in each distance cell constitutes the one-dimensional echo data: Based on this, construct the Hankel matrix. This makes all elements on every antidiagonal equal; that is:

[0013] ;

[0014] in, This represents the total number of pulses in the radar echo data matrix. This represents the total number of range cells in the radar echo data matrix. This indicates the number of columns in the Hankel matrix, which has 10 rows. ;

[0015] The second step is to perform singular value decomposition on the Hankel matrix, resulting in: ;in, Describes a left singular matrix. This represents a singular value matrix sorted from largest to smallest. This represents a right singular matrix, and the superscript H indicates the matrix complex conjugate transpose operation.

[0016] Furthermore, the steps for constructing the target subspace projection matrix to achieve ground clutter suppression are as follows:

[0017] The first step is to select the first segment of the radar echo data matrix after singular value decomposition. The singular left singular vectors constitute the ground clutter subspace matrix: ;in, This represents the left singular vector corresponding to ground clutter. The dimension of the clutter subspace of ground features;

[0018] The second step is to construct the target's subspace matrix based on the aforementioned ground clutter subspace matrix: ;in, The subspace matrix representing the target. Represents the identity matrix;

[0019] The third step is to project the original Hankel matrix A onto the target subspace to obtain the projected target subspace: ;

[0020] The fourth step is to apply the Hankel matrix reconstruction equation to... Reconstruction is performed to obtain one-dimensional data after ground clutter suppression:

[0021] ;

[0022] in, The first dimension of the one-dimensional data after ground clutter suppression represents the first dimension of the data. One value, Representation matrix The Line 1 The data in the column, , .

[0023] Furthermore, the steps for calculating the normalized smooth pseudo-Wigner-Willy distribution of the signal after clutter suppression are as follows:

[0024] The first step is to analyze the radar echo data after clutter suppression. Extract the first The total number of pulses in each distance unit constitutes a slow-time echo sequence: ;

[0025] The second step is to process the slow-time echo sequence. Perform a smooth pseudo-Gnahnau-Willi transform to obtain the first... Time-frequency distribution of each distance cell:

[0026] ;

[0027] in, Indicates the first The distance unit is the The time position and the first Time-frequency distribution values ​​under the joint coordinates of each frequency location Indicates the direction of time The smoothing window function, Indicates time delay The frequency-direction smoothing window function, with the superscript * indicating the complex conjugate sign;

[0028] The third step is to calculate the global mean of the smooth pseudo-Gnahnau-Willi distribution for all distance cells:

[0029] ;

[0030] in, This represents the global mean of a smooth pseudo-Gnahner-Willi distribution across all distance cells. Indicates the number of time sampling points. Indicates the number of frequency sampling points;

[0031] Step 4: Calculate the global standard deviation:

[0032] ;

[0033] in, This represents the global mean of a smooth pseudo-Gnahnau-Willi distribution across all distance cells;

[0034] The fifth step is to normalize the smooth pseudo-Gnahnau-Willi distribution for each distance cell:

[0035] ;

[0036] in, Indicates the first The distance unit of the first The time position and the first Normalized time-frequency distribution values ​​under the joint coordinates of each frequency location.

[0037] Furthermore, the extraction of the four-dimensional time-frequency features refers to:

[0038] A. At each time point of the normalized smooth pseudo-Wigner-Williams distribution, search for the maximum amplitude corresponding to the frequency dimension, sum up the maximum amplitude corresponding to all time points, and use the sum as the first feature dimension: Ridge Cumulative RI;

[0039] B. In the normalized smooth pseudo-Wigner-Willi distribution, the Otsu method is used to determine the significant time-frequency points, which are assigned a value of 1, and the remaining time-frequency points are assigned a value of 0, thus constructing a binary image STFP of significant time-frequency points; the 8-neighborhood connectivity rule is used to label the significant time-frequency points of STFP with connected components, and the statistically independent connected regions are defined as the second feature: the number of connected regions NR;

[0040] C. In the salient time-frequency binary image STFP, count the number of pixels with a value of 1 in each connected region, and take the maximum value as the third feature: the maximum connected region size MS;

[0041] D. In the normalized smooth pseudo-Wigner-Willi distribution, calculate the difference between adjacent time-frequency points in the time and frequency directions, square all the differences and sum them up, and use the result as the fourth dimension feature: time-frequency distribution flux.

[0042] Furthermore, the 2.5σ criterion refers to extracting the four-dimensional time-frequency features of each distance cell separately, and summing the values ​​of the same type of time-frequency features across all distance cells to form a set of time-frequency features: ;in, Indicates the first The distance unit of the first Calculate the time-frequency characteristic value. A set of time-frequency features mean and variance Meanwhile, with the mean as the center and 2.5 times the variance as the deviation range, the first... Normal range of time-frequency characteristic values: .

[0043] Furthermore, the specific exception detection threshold is as follows:

[0044] For the RI feature, the detection threshold condition is:

[0045] ;

[0046] in, Indicates the first The RI feature value of each distance cell, and These represent the mean and standard deviation of the RI features for all distance cells, respectively.

[0047] For the NR feature, the detection threshold condition is:

[0048] ;

[0049] in, Indicates the first The NR feature value of each distance cell, and These represent the mean and standard deviation of the NR features for all distance cells, respectively.

[0050] For MS features, the detection threshold condition is:

[0051] ;

[0052] in, Indicates the first MS eigenvalues ​​of distance cells, and Let represent the mean and standard deviation of the MS feature for all distance cells, respectively;

[0053] For the Flux feature, the detection threshold condition is:

[0054] ;

[0055] in, Indicates the first Flux eigenvalues ​​of distance cells, and These represent the mean and standard deviation of the Flux feature for all distance cells, respectively.

[0056] Furthermore, the step of making a candidate target feature decision for each time-frequency feature refers to: judging the time-frequency features of each distance unit. Does it meet the requirements? The 2.5σ criterion will determine the distance element that satisfies the 2.5σ criterion. Each time-frequency feature is labeled as a candidate target time-frequency feature.

[0057] Furthermore, the step of determining the final target distance unit based on multi-level decision logic is as follows:

[0058] The first step is to count the number of time-frequency features of each distance cell that are labeled as candidate target features based on their four dimensions of RI, NR, MS, and Flux.

[0059] The second step is to determine whether a distance cell contains a target if the number of candidate target features is greater than or equal to 3.

[0060] Furthermore, the steps for indicating the detection results of different time-frequency features at each distance cell are as follows:

[0061] The first step is to plot the distribution curves of RI, NR, MS, and Flux four-dimensional time-frequency across all distance cells;

[0062] The second step is to label the mean of each time-frequency characteristic on each time-frequency characteristic distribution curve. ,variance as well as Abnormal judgment threshold;

[0063] The third step is to mark the distance cells that meet the multi-level decision conditions as distance cells containing the target.

[0064] Compared with the prior art, the present invention has the following advantages:

[0065] First, this invention extracts four complementary time-frequency features from a normalized smoothed pseudo-Wigner-Willi distribution. This overcomes the shortcomings of existing technologies that rely solely on a single time-frequency feature to distinguish between targets and background clutter in complex terrain environments, leading to a high probability of missed detections of weak targets. This invention utilizes a combination of multi-dimensional time-frequency features, which, compared to a single feature, reduces the probability of false alarms. Furthermore, although weak targets may not be obvious with a single feature, they can be detected through the comprehensive anomaly detection of multiple features, thereby improving the target detection probability.

[0066] Second, this invention designs an adaptive multi-level decision method based on the 2.5σ criterion. This method automatically constructs anomaly decision thresholds based on the mean and variance of time-frequency features in the current data, avoiding threshold dependence on human experience. Compared to the traditional 3σ criterion, 2.5σ appropriately lowers the anomaly threshold, increasing the probability of weak target detection; simultaneously, it employs a comprehensive decision logic of "at least three features being simultaneously abnormal," significantly reducing the false alarm probability and improving the robustness and accuracy of detection. Attached Figure Description

[0067] Figure 1 This is an overall flowchart of an embodiment of the present invention;

[0068] Figure 2 This is a diagram showing the distribution of cumulative eigenvalues ​​of the ridge under different distance units and a schematic diagram of the decision result of a single feature target in the simulation experiment of this invention;

[0069] Figure 3 This is a distribution map of the number of connected regions under different distance units and a schematic diagram of the decision result of a single feature target in the simulation experiment of this invention;

[0070] Figure 4 This is a diagram showing the distribution of the maximum connected region size feature value under different distance units in the simulation experiment of this invention, and a schematic diagram of the decision result of a single feature target.

[0071] Figure 5 This is a diagram showing the distribution of time-frequency flux characteristic values ​​and the decision result of a single characteristic target under different distance units in the simulation experiment of this invention;

[0072] Figure 6 This is a distribution map of the cumulative eigenvalues ​​of the ridge under different distance units and a schematic diagram of the joint target decision results in the simulation experiment of this invention;

[0073] Figure 7 This is a distribution map of the number of connected regions under different distance units and a schematic diagram of the joint target decision results of multiple features in the simulation experiment of this invention;

[0074] Figure 8 This is a diagram showing the distribution of the maximum connected region size feature value under different distance units in the simulation experiment of this invention, and a schematic diagram of the joint target decision results of multiple features;

[0075] Figure 9 This is a diagram showing the distribution of time-frequency flux characteristic values ​​and the joint target decision results of multiple features under different distance units in the simulation experiment of this invention. Detailed Implementation

[0076] The specific implementation steps of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0077] To effectively detect low, slow-moving, and small targets in complex terrain clutter and improve target detection accuracy, this invention provides a radar target detection method based on multi-dimensional time-frequency features and multi-level decision-making, such as... Figure 1 As shown, the method includes the following steps:

[0078] Step 1: Subspace projection is performed using singular value decomposition of the Hankel matrix to preprocess the radar echo data by filtering out strong ground clutter. Then, the processed two-dimensional echo data is reconstructed into one-dimensional echo data.

[0079] Specifically, the Hankel matrix is ​​constructed as follows:

[0080] ;

[0081] in, Indicates the first [number] in the radar echo received data The first pulse data of each distance unit, This represents the total number of pulses. Representation matrix The number of columns, This represents the Hankel matrix.

[0082] The singular value decomposition of the Hankel matrix is ​​implemented as follows:

[0083] ;

[0084] in, Describes a left singular matrix. This represents a singular value matrix sorted from largest to smallest. This represents a right singular matrix, and the superscript H indicates the matrix complex conjugate transpose operation.

[0085] The following is a method for constructing the target subspace projection matrix to achieve ground clutter suppression:

[0086] Select the first part after singular value decomposition of the radar echo data matrix The singular left singular vectors constitute the ground clutter subspace matrix: ;in, This represents the left singular vector corresponding to ground clutter. This represents the dimension of the clutter subspace of ground features.

[0087] Construct the target's subspace matrix based on the aforementioned ground clutter subspace matrix: ;in, The subspace matrix representing the target. Represents the identity matrix.

[0088] Projecting the original Hankel matrix A onto the target subspace yields the projected target subspace: .

[0089] The following method is used to reconstruct the processed two-dimensional echo data into one-dimensional echo data:

[0090] Applying the Hankel matrix reconstruction equation to Reconstruction is performed to obtain one-dimensional data after ground clutter suppression:

[0091] ;

[0092] in, The first dimension of the one-dimensional data after ground clutter suppression represents the first dimension of the data. One value, Representation matrix The Line 1 The data in the column, , .

[0093] Step 2: Calculate the normalized smooth pseudo-Wigner-Willi distribution of the clutter-suppressed signal and extract the time-frequency features in four dimensions.

[0094] Specifically, the formula for the smoothed pseudo-Wigner-Wylie distribution after clutter suppression is as follows:

[0095] ;

[0096] in, Indicates the first The distance unit is the The time position and the first Time-frequency distribution values ​​under the joint coordinates of each frequency location Indicates the direction of time The smoothing window function, Indicates time delay The frequency-direction smoothing window function, with the superscript * indicating the complex conjugate symbol.

[0097] Calculate the global mean of the smooth pseudo-Gnahnau-Willi distribution for all distance cells:

[0098] ;

[0099] in, This represents the global mean of a smooth pseudo-Gnahner-Willi distribution across all distance cells. Indicates the number of time sampling points. Indicates the number of frequency sampling points;

[0100] Calculate the global standard deviation:

[0101] ;

[0102] in, This represents the global mean of the smooth pseudo-Gnahnau-Willi distribution of all distance cells.

[0103] Normalize the smooth pseudo-Gnahnau-Willi distribution for each distance cell:

[0104] ;

[0105] in, Indicates the first The distance unit of the first The time position and the first Normalized time-frequency distribution values ​​under the joint coordinates of each frequency location.

[0106] The specific implementation of extracting four-dimensional time-frequency features is as follows:

[0107] At each time point of the normalized smooth pseudo-Wigner-Willi distribution, search for the maximum amplitude corresponding to the frequency dimension, sum the maximum amplitudes corresponding to all time points, and use the sum as the first feature dimension: Ridge Cumulative (RI).

[0108] In a normalized smooth pseudo-Wigner-Willi distribution, the Otsu method is used to identify salient time-frequency points, which are assigned a value of 1, while other time-frequency points are assigned a value of 0, thus constructing a binary image of salient time-frequency points, STFP. The 8-neighborhood connectivity rule is used to label the salient time-frequency points of STFP with connected components, and the number of mutually independent connected regions is defined as the second feature: the number of connected components, NR.

[0109] In the salient time-frequency binary image STFP, the number of pixels with a value of 1 in each connected region is counted, and the maximum value is taken as the third feature: the maximum connected region size MS.

[0110] In the normalized smooth pseudo-Wigner-Willi distribution, the difference between adjacent time-frequency points in the time and frequency directions is calculated, all differences are squared and summed, and the result is used as the fourth dimension feature: time-frequency distribution flux.

[0111] Step 3: Based on the 2.5σ criterion, anomaly decision thresholds are automatically constructed according to the mean and variance of the time-frequency features in the current data. Candidate target feature decisions are made for each time-frequency feature, and the final target distance unit is determined according to the multi-level decision logic.

[0112] Specifically, the implementation method based on the 2.5σ criterion is as follows:

[0113] The four-dimensional time-frequency features of each distance cell are extracted separately, and the values ​​of the same type of time-frequency features across all distance cells are summarized to form a set of time-frequency features: ;in, Indicates the first The distance unit of the first Calculate the time-frequency characteristic value. A set of time-frequency features mean and variance Meanwhile, with the mean as the center and 2.5 times the variance as the deviation range, the first... Normal range of time-frequency characteristic values: .

[0114] The specific anomaly detection thresholds are as follows:

[0115] For the RI feature, the detection threshold condition is:

[0116] ;

[0117] in, Indicates the first The RI feature value of each distance cell, and These represent the mean and standard deviation of the RI characteristics for all distance cells, respectively.

[0118] For the NR feature, the detection threshold condition is:

[0119] ;

[0120] in, Indicates the first The NR feature value of each distance cell, and These represent the mean and standard deviation of the NR features for all distance cells, respectively.

[0121] For MS features, the detection threshold condition is:

[0122] ;

[0123] in, Indicates the first MS eigenvalues ​​of distance cells, and represents the mean and standard deviation of the MS feature for all distance cells, respectively.

[0124] For the Flux feature, the detection threshold condition is:

[0125] ;

[0126] in, Indicates the first Flux eigenvalues ​​of distance cells, and These represent the mean and standard deviation of the Flux feature for all distance cells, respectively.

[0127] The method for determining candidate target features for each time-frequency feature is as follows:

[0128] Determine the time-frequency characteristics of each distance cell Does it meet the requirements? The 2.5σ criterion will determine the distance element that satisfies the 2.5σ criterion. Each time-frequency feature is labeled as a candidate target time-frequency feature.

[0129] The implementation method for determining the final target distance unit using multi-level decision logic is as follows:

[0130] For each distance cell, the number of time-frequency features of its four dimensions (RI, NR, MS, and Flux) that are labeled as candidate target features is counted.

[0131] Distance cells with a number of candidate target features greater than or equal to 3 are determined to contain a target.

[0132] Step 4: Mark the detection results of different time-frequency features on each distance cell.

[0133] Plot the distribution curves of RI, NR, MS and Flux four-dimensional time-frequency on all distance cells.

[0134] The mean of each time-frequency characteristic is marked on each time-frequency characteristic distribution curve. ,variance as well as Abnormal judgment threshold.

[0135] The distance cell that meets the multi-level decision conditions is marked as the distance cell containing the target.

[0136] The effectiveness of this invention can be further demonstrated through the following simulation.

[0137] 1. Simulation experimental conditions.

[0138] The hardware platform for the simulation experiment of this invention is: Intel i7 9750H CPU with a main frequency of 2.6GHz and 16GB of memory.

[0139] The software platform for the simulation experiment of this invention is: Windows 10 operating system and Matlab R2024a.

[0140] The simulation experiment of this invention uses measured radar echo data containing ground clutter.

[0141] 2. Simulation content and result analysis.

[0142] The simulation experiment of this invention uses the present invention and the 2.5σ criterion of single time-frequency feature to process the input measured radar echo data to obtain the target detection results.

[0143] In the simulation experiment, the single time-frequency feature and 2.5σ criterion target detection method refers to the use of singular value decomposition of radar echo data combined with subspace projection to suppress clutter, calculation of normalized smooth pseudo-Wigner-Willi distribution, and extraction of ridge accumulation feature RI, maximum connected region size feature MS, time-frequency distribution flux feature Flux, and connected region quantity feature NR as single decision criteria to detect targets in each range cell.

[0144] The simulation experiment of this invention uses the method of this invention and a prior art, selecting data from the 45-20-2021-07-29-14-37-03-09.mat file in the 45m20MHz file of the radar low-speed small UAV detection dataset (LSS-Ku-1.0) under ground clutter background. The results are shown in the figure below. Figure 2-9 As shown.

[0145] The following is combined with Figures 2-9 The simulation diagrams further illustrate the effects of the present invention.

[0146] Figure 2-9 The horizontal axis represents the number of distance cells, the vertical axis represents the time-frequency feature values, the blue lines represent the distribution curve of the time-frequency feature values ​​along the distance cells, the green lines represent the mean line of the time-frequency feature, the orange lines represent the sum and difference of the mean of the time-frequency feature and the unit standard deviation 1σ, the red lines represent the detection threshold line of the time-frequency feature, and the red circles represent the detected targets.

[0147] from Figure 2-5 It can be seen that the targets detected by ridge cumulative features and the number of connected regions are concentrated in the range of 50 to 100 distance cells, while the targets detected by the maximum connected region size and the time-frequency distribution flux are scattered. Ridge cumulative features, the number of connected regions, and the time-frequency distribution flux detect the targets that actually exist in the dataset, but at the same time misclassify many non-target distance cells as target distance cells.

[0148] from Figure 6-9 It can be seen that the detected target at distance cells 54 and 55 has a distance cell consistent with the real target in the dataset, and its ridge cumulative feature value ( Figure 6 Features exceeding a threshold are considered candidate features; the number of connected regions ( Figure 8 If the flux is below the threshold, it is considered a candidate feature, and the time-frequency distribution flux ( Figure 9 If the value is higher than the threshold, it is considered a candidate feature. Combining the four features, distance cells 54 and 55 have three features that meet the criteria, and are therefore ultimately identified as target distance cells. Experimental results demonstrate that the method proposed in this invention successfully detects targets and filters out many target-free distance cells in complex terrain clutter, verifying its significant advantages in detection probability and robustness.

[0149] The simulation experiments above show that the present invention effectively suppresses ground clutter through subspace projection, and overcomes the shortcomings of existing technologies, such as susceptibility to interference and high false alarm rates due to single features, by using joint decision-making based on multi-dimensional time-frequency features. This significantly improves the detection probability and robustness in complex ground backgrounds.

Claims

1. A radar target detection method based on multi-dimensional time-frequency features and multi-level decision-making, characterized in that, The steps of this detection method include the following: Step 1: Subspace projection is performed using singular value decomposition of the Hankel matrix to preprocess the radar echo data by filtering out strong ground clutter. Then, the processed two-dimensional echo data is reconstructed into one-dimensional echo data. Step 2: Calculate the normalized smooth pseudo-Wigner-Willi distribution of the clutter-suppressed signal and extract the time-frequency features in four dimensions; Step 3: Based on the 2.5σ criterion, anomaly judgment thresholds are automatically constructed according to the mean and variance of the time-frequency features in the current data. Candidate target feature judgments are made for each time-frequency feature, and the final target distance unit is determined according to the multi-level judgment logic. Step 4: Mark the detection results of different time-frequency features on each distance cell.

2. The radar target detection method according to claim 1, characterized in that, The steps for performing singular value decomposition on the radar echo data matrix described in step 1 are as follows: The first step is to analyze the radar echo data matrix. Extracting the first The total number of pulses in each distance cell constitutes the one-dimensional echo data: Based on this, construct the Hankel matrix. This makes all elements on every antidiagonal equal; that is: ； in, This represents the total number of pulses in the radar echo data matrix. This represents the total number of range cells in the radar echo data matrix. This indicates the number of columns in the Hankel matrix, which has 10 rows. ; The second step is to perform singular value decomposition on the Hankel matrix, resulting in: ;in, Describes a left singular matrix. This represents a singular value matrix sorted from largest to smallest. This represents a right singular matrix, and the superscript H indicates the matrix complex conjugate transpose operation.

3. The radar target detection method according to claim 2, characterized in that, The steps described in step 1 for constructing the target subspace projection matrix to suppress ground clutter and reconstruct one-dimensional echo data are as follows: The first step is to select the first segment of the radar echo data matrix after singular value decomposition. The singular left singular vectors constitute the ground clutter subspace matrix: ;in, This represents the left singular vector corresponding to ground clutter. The dimension of the clutter subspace of ground features; The second step is to construct the target's subspace matrix based on the aforementioned ground clutter subspace matrix: ;in, The subspace matrix representing the target. Represents the identity matrix; The third step is to project the original Hankel matrix A onto the target subspace to obtain the projected target subspace: ; The fourth step is to apply the Hankel matrix reconstruction equation to... Reconstruction is performed to obtain one-dimensional data after ground clutter suppression: ； in, The first dimension of the one-dimensional data after ground clutter suppression represents the first dimension of the data. One value, Representation matrix The Line 1 The data in the column, , .

4. The radar target detection method according to claim 3, characterized in that, The steps for calculating the normalized and smoothed pseudo-Wigner-Willy distribution of the clutter-suppressed signal in step 2 are as follows: The first step is to analyze the radar echo data after clutter suppression. Extract the first The total number of pulses in each distance unit constitutes a slow-time echo sequence: ; The second step is to process the slow-time echo sequence. Perform a smooth pseudo-Gnahnau-Willi transform to obtain the first... Time-frequency distribution of each distance cell: ； in, Indicates the first The distance unit is the The time position and the first Time-frequency distribution values ​​under the joint coordinates of each frequency location Indicates the direction of time The smoothing window function, Indicates time delay The frequency-direction smoothing window function, with the superscript * indicating the complex conjugate sign; The third step is to calculate the global mean of the smooth pseudo-Gnahnau-Willi distribution for all distance cells: ； in, This represents the global mean of a smooth pseudo-Gnahner-Willi distribution across all distance cells. Indicates the number of time sampling points. Indicates the number of frequency sampling points; Step 4: Calculate the global standard deviation: ； in, This represents the global mean of the smooth pseudo-Gnahnau-Willi distribution of all distance cells; The fifth step is to normalize the smooth pseudo-Gnahnau-Willi distribution for each distance cell: ； in, Indicates the first The distance unit of the first The time position and the first Normalized time-frequency distribution values ​​under the joint coordinates of each frequency location.

5. The radar target detection method according to claim 1, characterized in that, Step 2, which involves extracting the four-dimensional time-frequency features, refers to: A. At each time point of the normalized smooth pseudo-Wigner-Williams distribution, search for the maximum amplitude corresponding to the frequency dimension, sum up the maximum amplitude corresponding to all time points, and use the sum as the first feature dimension: Ridge Cumulative RI; B. In the normalized smooth pseudo-Wigner-Willi distribution, the Otsu method is used to determine the significant time-frequency points, which are assigned a value of 1, and the remaining time-frequency points are assigned a value of 0, thus constructing a binary image of significant time-frequency points STFP; The 8-neighborhood connectivity rule is used to label the significant time-frequency points of STFP with connected components, and the number of mutually independent connected regions is defined as the second feature: the number of connected regions NR. C. In the salient time-frequency binary image STFP, count the number of pixels with a value of 1 in each connected region, and take the maximum value as the third feature: the maximum connected region size MS; D. In the normalized smooth pseudo-Wigner-Willi distribution, calculate the difference between adjacent time-frequency points in the time and frequency directions, square all the differences and sum them up, and use the result as the fourth dimension feature: time-frequency distribution flux.

6. The radar target detection method according to claim 5, characterized in that, The 2.5σ criterion mentioned in step 3 refers to extracting the four-dimensional time-frequency features of each distance cell, and summing the values ​​of the same type of time-frequency features across all distance cells to form a set of time-frequency features: ； in, Indicates the first The distance unit of the first Calculate the time-frequency characteristic value. A set of time-frequency features mean and variance Meanwhile, with the mean as the center and 2.5 times the variance as the deviation range, the first... Normal range of time-frequency characteristic values: .

7. The radar target detection method according to claim 6, characterized in that, The specific anomaly detection threshold mentioned in step 3 is as follows: For the RI feature, the detection threshold condition is: ； in, Indicates the first The RI feature value of each distance cell, and These represent the mean and standard deviation of the RI characteristics for all distance cells, respectively. For the NR feature, the detection threshold condition is: ； in, Indicates the first The NR feature value of each distance cell, and These represent the mean and standard deviation of the NR features for all distance cells, respectively. For MS features, the detection threshold condition is: ； in, Indicates the first MS eigenvalues ​​of distance cells, and Let represent the mean and standard deviation of the MS feature for all distance cells, respectively; For the Flux feature, the detection threshold condition is: ； in, Indicates the first Flux eigenvalues ​​of distance cells, and These represent the mean and standard deviation of the Flux feature for all distance cells, respectively.

8. The radar target detection method according to claim 7, characterized in that, Step 3, which involves making a candidate target feature decision for each time-frequency feature, refers to: determining the time-frequency features of each distance cell. Does it meet the requirements? The 2.5σ criterion will determine the distance element that satisfies the 2.5σ criterion. Each time-frequency feature is labeled as a candidate target time-frequency feature.

9. The radar target detection method according to claim 8, characterized in that, The steps in step 3 to determine the final target distance unit based on the multi-level decision logic are as follows: The first step is to count the number of time-frequency features of each distance cell that are labeled as candidate target features based on their four dimensions of RI, NR, MS, and Flux. The second step is to determine whether a distance cell contains a target if the number of candidate target features is greater than or equal to 3.

10. The radar target detection method according to claim 9, characterized in that, The steps in step 4 for indicating the detection results of different time-frequency features in each distance cell are as follows: The first step is to plot the distribution curves of RI, NR, MS, and Flux four-dimensional time-frequency across all distance cells; The second step is to label the mean of each time-frequency characteristic on each time-frequency characteristic distribution curve. ,variance as well as Abnormal judgment threshold; The third step is to mark the distance cells that meet the multi-level decision conditions as distance cells containing the target.

Images