A method for detecting marine radar targets

Abstract

The present invention discloses a method for detecting a target by a marine radar. The method comprises the following steps: selecting a radar echo data sample from offline radar data, selecting radar echo feature parameters and performing feature fusion to obtain fused echo feature parameters and a sample training set, calculating the fused echo feature parameter value and inputting it into an RBF neural network for training to obtain an RBF classifier; obtaining echo data to be tested, and calculating the echo feature value of the echo data to be tested according to the selected echo feature parameters; performing feature fusion on the echo feature parameters to obtain a feature matrix; inputting the feature matrix into an RBF classifier model to finally obtain an output y value; and judging whether the scanning line has a target: comparing the output y value with the expected output value, if y is greater than the expected output value, then there is a target; otherwise, there is no target. The present invention solves the problem of large amount of high-dimensional data calculation in target detection, and can not only quickly perform target detection, but also improve the accuracy of target detection.

Description

Technical Field

[0001] The invention belongs to the technical field of marine remote sensing under marine environmental conditions, and relates to a marine radar target detection method, in particular to a marine radar target rapid detection method based on multi-feature fusion. Background Art

[0002] In the complex sea clutter background, radar image data has too many features. In rapid target detection, the large number of features in the data set leads to a large amount of computation for classifier training, which has always been a challenge for marine radar target detection.

[0003] Classic marine radar target detection mostly uses point-by-point detection CFAR algorithms. Due to the correlation between sea clutter pulses, Watts analyzed the performance of the cell averaging (CA) CFAR detector in terms of sea clutter correlation in 1996 for the sea clutter model of the K distribution model. In 2002, Li Lan et al. realized constant false alarm rate target detection for marine targets. When the waveform of sea clutter changes greatly, the constant false alarm rate technology has a strong performance for marine target detection, but this method is not suitable for target detection in other sea conditions. In 2007, Xu Hui et al. found through a large number of radar image experiments that the double threshold CFAR technology can improve the performance of target detection in different angle domains, but it increases the complexity of the CFAR algorithm, resulting in low detection accuracy. In 2011, Chen Huajie et al. proposed an adaptive CFAR algorithm for clutter in multiple background areas based on the multi-characteristic distribution of sea clutter. This method uses different modeling methods to achieve target detection in areas with different sea clutter distribution characteristics. In 2021, Zou Chengxiao and others made a detailed review of the constant false alarm detection algorithm in the field of radar target detection, but the CFAR algorithm has a high false alarm rate, long detection time and low detection efficiency in marine radar target detection. Summary of the invention

[0004] In view of the above-mentioned prior art, the technical problem to be solved by the present invention is to provide a marine radar target detection method based on multi-feature fusion, which solves the problem of reducing the influence of too many features of the radar echo data set under the background of strong and complex sea clutter signals, improves the detection performance of the classifier, and quickly distinguishes pure sea clutter from target echoes, thereby improving the target detection efficiency of radar echo data with multi-features.

[0005] In order to solve the above technical problems, a marine radar target detection method of the present invention comprises:

[0006] Step 1: Select radar echo data samples on a single scan line in the offline radar data, the samples include samples containing target radar echoes and samples of pure sea clutter radar echoes without targets; select radar echo feature parameters, calculate feature parameter values, perform feature fusion on the selected radar echo feature parameters to obtain fused echo feature parameters and sample training sets, calculate fused echo feature parameter values, input the fused echo feature parameter values ​​into the RBF neural network for training, and obtain the RBF classifier after feature fusion;

[0007] Step 2: Acquire the echo data to be measured on a single scanning line in the radar image, and calculate the echo characteristic value of the echo data to be measured according to the echo characteristic parameters selected in step 1;

[0008] Step 3: perform feature fusion on the echo feature parameters in step 2 according to step 1 to obtain a feature matrix;

[0009] Step 4: Input the feature matrix calculated in step 3 into the RBF classifier model obtained in step 1, and finally obtain the output y value;

[0010] Step 5: Determine whether there is a target on the scan line: compare the output y value with the expected output value. If y is greater than the expected output value, there is a target; otherwise, there is no target.

[0011] Furthermore, the radar echo characteristic parameters are mean, variance, fitting correlation coefficient, variance coefficient, weighted filling ratio, dispersion coefficient and maximum fluctuation;

[0012] Furthermore, the selected feature parameters are subjected to feature fusion, and the fused echo feature parameters and sample training set include:

[0013] Step a: Calculate the optimal inter-feature covariance matrix:

[0014]

[0015] Where, T p represents the m×n training sample matrix of the test data, m represents the number of training set data, n represents the number of selected radar echo characteristic parameters, and R represents the n×n covariance matrix;

[0016] Step b: Calculate the R eigenvalue as:

[0017] |λ i IR|=0(i=1,2.....n)

[0018] R eigenvector calculation formula:

[0019] R i =λ i b i

[0020] Among them, b i represents the i-th eigenvector of R, λ i represents the i-th eigenvalue of R;

[0021] Step c: Calculate the projection matrix C:

[0022] C=[b a，..... b j ]

[0023] Where b a ......b j It represents the eigenvectors corresponding to the first x eigenvalues ​​in descending order of eigenvalues. C is an n×x matrix, where x represents the number of retained eigenvectors.

[0024] Step d: Calculate the sample training set after feature fusion:

[0025] T PF =T P C

[0026] The echo feature parameters after feature fusion are:

[0027] Q=[q a ,....,q j ]

[0028] In the formula, q a ......q j Denotes the eigenvalue λ a ......λ j The corresponding characteristic parameters.

[0029] Furthermore, the fused echo feature parameter values ​​are input into the RBF neural network for training, and the RBF classifier after feature fusion is obtained, including:

[0030] Step 1.1: Input the fusion echo characteristic parameter values ​​of the sample training set into the RBF neural network;

[0031] Step 1.2: Initialize the connection weight vector from the hidden layer to the output layer:

[0032] W m =[w m1 ,w m2 ,…w mq ] T ,(m=1,2,…,q)

[0033] In the formula, q is the number of hidden layer units, and m represents the mth hidden layer;

[0034] Initialize the weights from the hidden layer to the output layer:

[0035]

[0036] Where mink is the minimum value of the result predicted by the output of the k-th neuron in the training sample; maxk is the maximum value of the result predicted by the output of the k-th neuron in the training sample, and j represents the j-th neuron in the hidden layer;

[0037] Step 1.3: Initialize the central parameter vector of each neuron in the hidden layer:

[0038] C j =[c j1 ,c j2 ,…c jn ] T

[0039] In the formula, j represents the jth neuron in the hidden layer, and n represents the number of sampling points in a single scan line;

[0040] The initial values ​​of the central parameters of the RBF neural network are:

[0041]

[0042] Where min i is the minimum value of all input information of the i-th feature in the training set, max i is the maximum value of all input information of the i-th feature in the training set, and p represents the total number of neurons in the hidden layer;

[0043] Step 1.4: Initialize the width vector:

[0044] D j =[d j1 ,d j2 ,…d jn ] T

[0045] In the formula, j represents the jth neuron in the hidden layer, and n represents the number of sampling points in a single scan line;

[0046] The width vector is:

[0047]

[0048] Where, d f is the width adjustment coefficient, x ij represents the i-th eigenvalue of the j-th scan line, c ij represents the central parameter value of the i-th eigenvalue of the j-th scan line;

[0049] Calculate the output of the jth hidden layer neuron:

[0050]

[0051] In the formula, X represents the input feature vector, C represents its central parameter vector, and D represents the width vector;

[0052] Step 1.5: Input radar image feature data:

[0053]

[0054] Where N represents the number of scan lines;

[0055] Compute the radial basis function as the node activation function:

[0056]

[0057] Where R is the hidden layer neuron value;

[0058] Step 1.6: Calculate neuron output:

[0059]

[0060] In the formula, y k is a specific output element in the output vector Y, where the output vector is:

[0061] Y=[y1,y2,…y q ] T

[0062] The expected output vector is a zero vector, specifically:

[0063] O=[o1,o2,…o q ] T

[0064] Where q represents the number of hidden layer units;

[0065] Step 1.7: The RBF neural network uses the gradient descent method to adaptively adjust the RBF neural network parameters to the optimal value. The iterative calculation formulas of the connection weight vector, the center parameter vector and the width vector are:

[0066]

[0067]

[0068]

[0069] In the formula, W represents the training weight vector, C represents its central parameter vector, and D represents the width vector

[0070] Step 1.8: After the center, width and weight are updated, the root mean square error RMS value of the network output is calculated according to the given values ​​of η and α and the value of the iteration termination accuracy ε:

[0071]

[0072] Where O represents the output expected matrix, N is the number of scan lines, if RMS ≤ ε, the training ends, otherwise return to step 1.7 to iteratively calculate the connection weight vector, center parameter vector and width vector.

[0073] Furthermore, the radar is an X-band marine radar.

[0074] Beneficial effects of the present invention: The target rapid detection method based on multi-feature fusion that can be applied to pulse system X-band marine radar of the present invention can effectively solve the problem of excessive high-dimensional data calculation in multi-feature target detection. The method can perform overall detection on the radial data set of each scan line in the radar echo, screen out the scan lines where the target exists, and first perform feature fusion and then perform detection through a trained classifier to solve the problem of a large number of echo feature parameters. That is, the classifier trained by the multi-feature fusion method can perform rapid target detection on a single scan line data of the on-site radar image, solving the problem of too many features in the data set, resulting in large classifier training and test calculations. The method of feature fusion is used to reduce the number of features, which not only reduces the target detection time, but also improves the accuracy of target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 It is an image with co-channel interference removed;

[0076] Figure 2 is the target echo sample;

[0077] Figure 3 is a pure sea clutter sample;

[0078] Figure 4 It is a radial basis neuron model;

[0079] Figure 5 is a flow chart of implementation mode;

[0080] Figure 6 It is the scene feature fusion target detection result map;

[0081] Figure 7 It is a single scan line data curve graph. DETAILED DESCRIPTION

[0082] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0083] Combination Figure 5 , the present invention comprises the following steps:

[0084] Step 1: Get the RBF classifier model after feature fusion:

[0085] Through the offline radar echo data, the experiment was conducted, and two types of radar echo image data samples on a single scan line were selected, one type was samples containing target radar echoes, and the other type was samples of pure sea clutter radar echoes without targets; multiple echo data feature parameters were selected, and the feature parameter values ​​were calculated. The selected feature parameters were feature fused to obtain the feature fused echo feature parameters and sample training set, and the feature parameter values ​​were calculated. The feature parameter values ​​of the fused feature parameters of the fused samples were input into the RBF classifier model for training to obtain the RBF classifier after feature fusion.

[0086] Step 2, obtain the echo data to be measured on a single scanning line of a radar, select the characteristic parameters of the radar echo sample data before fusion according to step 1, and calculate the characteristic parameter values.

[0087] Step 3: perform feature fusion on the radar echo feature parameters in step 2 according to step 1 to obtain a feature matrix.

[0088] Step 4: Input the feature matrix finally calculated in step 3 into the RBF classifier model trained in step 1, and finally obtain an output y value.

[0089] Step 5: Determine whether there is a target on the selected scan line:

[0090] Use the final output y value to compare with the expected output value. If the output is greater than the expected output value, there is a target; if the output is less than the expected output value, there is no target.

[0091] The RBF classifier model obtained after feature fusion in step 1 specifically includes:

[0092] Step 1.1, obtain the feature subset and sample training set after feature fusion:

[0093] Step 1.1.1, the optimal feature covariance matrix calculation formula is:

[0094]

[0095] Where T p represents the m×n training sample matrix of the test data, m represents the number of training set data, n represents the number of features to be obtained, and R represents the n×n covariance matrix.

[0096] Step 1.1.2, R eigenvalue calculation formula:

[0097] |λ i IR|=0(i=1,2.....n)

[0098] R eigenvector calculation formula:

[0099] Ri =λ i b i

[0100] Step 1.1.3, projection matrix C calculation formula:

[0101] C=[b a，..... b j ]

[0102] Where b a ......b j Represents the eigenvectors corresponding to the largest eigenvalues. C is an n×x matrix, where x represents the number of retained eigenvectors.

[0103] Step 1.1.4, the calculation formula of the training sample set after feature fusion is:

[0104] T PF =T P C

[0105] Echo feature parameters after feature fusion:

[0106] Q=[q a ,....,q j ]

[0107] Where q a ......q j Represents λ a ......λ j The feature parameter corresponding to the feature.

[0108] Step 1.2, train the RBF classifier:

[0109] Step 1.2.1, input the fusion feature parameters of the sample training set finally obtained in step 1.1 into the classifier model.

[0110] Step 1.2.2, use RBF neural network training to initialize the connection weight vector from the hidden layer to the output layer:

[0111] W m =[w m1 ,w m2 ,…w mq ] T ,(m=1,2,…,q)

[0112] In the formula, q is the number of hidden layer units, and m represents the mth hidden layer;

[0113] The weight initialization formula from the hidden layer to the output layer is:

[0114]

[0115] Where mink is the minimum value of the result predicted by the output of the k-th neuron in the training sample; maxk is the maximum value of the result predicted by the output of the k-th neuron in the training sample, and j represents the j-th neuron in the hidden layer;

[0116] Step 1.2.3: Initialize the central parameter vector of each neuron in the hidden layer:

[0117] C j =[c j1 ,c j2 ,…c jn ] T

[0118] In the formula, j represents the jth neuron in the hidden layer, and n represents the number of sampling points in a single scan line;

[0119] The initial values ​​of the central parameters of the RBF neural network are:

[0120]

[0121] Where min i is the minimum value of all input information of the i-th feature in the training set, max i is the maximum value of all input information of the i-th feature in the training set, and p represents the total number of neurons in the hidden layer;

[0122] Step 1.2.4: Initialize the width vector formula:

[0123] D j =[d j1 ,d j2 ,…d jn ] T

[0124] In the formula, j represents the jth neuron in the hidden layer, and n represents the number of sampling points in a single scan line;

[0125] Width vector:

[0126]

[0127] Where, d f is the width adjustment coefficient, x ij represents the i-th eigenvalue of the j-th scan line, c ij represents the central parameter value of the i-th eigenvalue of the j-th scan line;

[0128] The calculation formula for the output of the jth hidden layer neuron is:

[0129]

[0130] Where X represents the input feature vector, C represents its central parameter vector, and D represents the width vector;

[0131] Step 1.2.5: Input radar image feature data:

[0132]

[0133] Where N represents the number of scan lines.

[0134] Compute the radial basis function as the node activation function:

[0135]

[0136] Where R is the hidden layer neuron value;

[0137] Step 1.2.6: Calculate neuron output:

[0138]

[0139] In the formula, y k is a specific output element in the output vector Y, where the output vector is:

[0140] Y=[y1,y2,…y q ] T

[0141] The expected output vector is a zero vector, specifically:

[0142] O=[o1,o2,…o q ] T

[0143] Where q represents the number of hidden layer units;

[0144] Step 1.2.7: The RBF neural network uses the gradient descent method to adaptively adjust the RBF neural network parameters connection weight vector, center parameter vector and width vector to the optimal value; the connection weight vector, center parameter vector and width vector calculation formula are:

[0145]

[0146]

[0147]

[0148] In the formula, W represents the training weight vector, C represents its central parameter vector, and D represents the width vector;

[0149] Step 1.2.8: After the center, width and weight are updated, the root mean square error RMS value of the network output is calculated according to the given values ​​of η and α and the value of the iteration termination accuracy ε:

[0150]

[0151] Where O represents the output expected matrix, N is the number of scan lines, if RMS ≤ ε, the training ends, otherwise return to step 1.2.7 to iteratively calculate the connection weight vector, center parameter vector and width vector.

[0152] The feature fusion in step 3 specifically includes:

[0153] Step 3.1, the optimal feature covariance matrix calculation formula is:

[0154]

[0155] Where T p represents the m×n training sample matrix of the test data, m represents the number of training set data, n represents the number of features to be obtained, and R represents the n×n covariance matrix.

[0156] Step 3.2, R eigenvalue calculation formula:

[0157] |λ i IR|=0(i=1,2.....n)

[0158] R eigenvector calculation formula:

[0159] R i =λ i b i

[0160] Step 3.3, projection matrix C calculation formula:

[0161] C=[b a，..... b j ]

[0162] Where b a ......b j Represents the eigenvectors corresponding to the largest eigenvalues. C is an n×x matrix, where x represents the number of retained eigenvectors.

[0163] Step 3.4, the calculation formula of the training sample set after feature fusion is:

[0164] T PF =T P C

[0165] Echo feature parameters after feature fusion:

[0166] Q=[q a ,....,q j ]

[0167] Where q a ......q j Represents λa ......λ j The feature parameter corresponding to the feature.

[0168] The following is an example with specific parameters:

[0169] The marine radar used in the embodiment of the present invention is an X-band navigation radar, which works in short pulse mode. The radar antenna scans one circle in about 2.5 seconds. The total number of lines in a radar image is 2048, and there are 2048 pixel points on each line. The azimuth resolution is about 0.1° and the radial resolution is about 2.5m.

[0170] The main technical parameters of the marine radar are shown in Table 1:

[0171] Table 1 Technical parameters of marine radar

[0172]

[0173]

[0174] Combined with Figure 5 , the specific experimental steps of the present invention are:

[0175] Step 1: Use feature fusion to obtain the RBF classifier:

[0176] Step 1.1, conduct the experiment offline, first perform co-frequency interference suppression processing on the selected original radar image, such as Figure 1 . Perform a preliminary screening of the preprocessed radar image. Select two types of radar echo data samples on a single scan line through offline radar data. In this example, two types of radar echo data samples on 5559 scan lines of radar data are selected for training. One type is 2704 samples containing target radar echoes, such as Figure 2 The other category is 2855 samples of pure sea clutter radar echoes without targets, such as Figure 3 .

[0177] Step 1.2: In this example, 7 radar echo sample data features are selected, including mean, variance, fitting correlation coefficient, variance coefficient, weighted filling ratio, dispersion coefficient and maximum fluctuation as feature parameters and the feature values ​​are calculated.

[0178] Step 1.2.1, the mean, variance, and fitting correlation coefficient characteristic parameters are calculated as follows:

[0179]

[0180]

[0181]

[0182] In this example, n is 550, which indicates the number of sampling points, σ∈[0,2.5] indicates the grayscale value, θ∈[0°,360°] indicates the azimuth angle, and r∈[0,4500](m) indicates the radial distance. The result representing the radar echo intensity after fitting; represents the mean value of the radar echo intensity of the selected line; z∈[0,5] represents the radar echo intensity value before fitting.

[0183] Step 1.2.2, the variance coefficient and weighted filling ratio are:

[0184]

[0185]

[0186] In this example, σ∈[0,2.5] represents the grayscale value, θ∈[0°,360°] represents the azimuth angle, and r∈[0,4500](m) represents the radial distance. n is 550, which represents the number of sampling points in a single scan line, and μ is a 5559×1 column vector representing the mean.

[0187] Step 1.2.3, the dispersion coefficient and the maximum fluctuation are:

[0188]

[0189] F1=max(σ(θ,r i ))-μ

[0190] In this example, σ∈[0,2.5] represents the grayscale value, θ∈[0°,360°] represents the azimuth angle, and r∈[0,4500](m) represents the radial distance. μ is a 5559×1 column vector representing the mean, and s is a 5559×1 column vector representing the standard deviation.

[0191] Step 1.3, obtain the feature subset and sample training set after feature fusion:

[0192] Step 1.3.1, the optimal feature covariance matrix calculation formula is:

[0193]

[0194] In this example, T P It represents the 5559×7 test data training sample matrix selected in this example, 5559 represents the number of training set data, 7 represents the number of features obtained, and the dimension of the covariance matrix R is 7×7.

[0195] Step 1.3.2, R eigenvalue calculation formula:

[0196] |λ i IR|=0(i=1,2.....n)

[0197] The eigenvalues ​​calculated in this example are 1.5628, 0.0801, 0.0361, 0.0095, 0.0028, 1.2262×10 -4 , 9.1261×10 -7 .

[0198] R eigenvector calculation formula:

[0199] R i =λ i b i

[0200] Step 1.3.3, projection matrix C calculation formula:

[0201] C=[b a ,.....b j ]

[0202] In this example, the eigenvectors corresponding to the three eigenvalues ​​λ1=1.5628, λ2=0.0801, and λ3=0.036 are selected to form a 7×3 projection matrix C.

[0203] Step 1.3.4, the calculation formula of the training sample set after feature fusion is:

[0204] T PF =T P C

[0205] Echo feature parameters after feature fusion:

[0206] Q=[q a ,....,q j ]

[0207] In this example, Q is a 1×3 matrix, which means that the eigenvectors corresponding to the eigenvalues ​​λ1, λ2, and λ3 constitute the feature parameters corresponding to the projection features. In this example, 2382 fused sample sets and 3 fused feature parameters are finally obtained.

[0208] Step 1.4, train the RBF classifier:

[0209] Step 1.4.1, input the fusion feature parameters of the sample training set finally obtained in step 1 into the RBF model.

[0210] Step 1.4.2, in this example, RBF neural network is used for training. Figure 4 , initialize the connection weight vector from the hidden layer to the output layer:

[0211] W m =[w m1 ,w m2 ,…w mq ]T ,(m=1,2,…,q)

[0212] In this example, q is 300, which indicates the number of hidden layer units. m indicates the mth hidden layer.

[0213] The weight initialization formula from the hidden layer to the output layer is:

[0214]

[0215] In this example, mink is the minimum value -1 of the result predicted by the kth output neuron in the training sample; maxk is the maximum value 1 of the result predicted by the kth output neuron in the training sample, and j represents the jth neuron in the hidden layer.

[0216] Step 1.4.3: Initialize the central parameter vector of each neuron in the hidden layer:

[0217] C j =[c j1 ,c j2 ,…c jn ] T

[0218] In this example, j represents the jth neuron in the hidden layer, and n is 550, which represents the number of sampling points in a single scan line.

[0219] The initial values ​​of the central parameters of the RBF neural network are:

[0220]

[0221] In this example, mini is the minimum value of all input information of the i-th feature in the training set, which is 0, and maxi is the maximum value of all input information of the i-th feature in the training set, which is 1. p is 3, which indicates the total number of neurons in the hidden layer.

[0222] Step 1.4.4: Initialize the width vector formula:

[0223] D j =[d j1 ,d j2 ,…d jn ] T

[0224] In this example, j represents the jth neuron in the hidden layer, and n is 550, which represents the number of sampling points in a single scan line.

[0225] Width vector:

[0226]

[0227] In this example, d f 0.8 represents the width adjustment coefficient, x ijrepresents the i-th eigenvalue of the j-th scan line, c ij Indicates the center parameter value. N is 550, which means the number of sampling points in a single scan line.

[0228] The calculation formula for the output of the jth hidden layer neuron is:

[0229]

[0230] In this example, X represents the 3×5559 input feature vector, C represents the 1×550 center parameter vector, and D represents the 1×550 width vector.

[0231] Step 1.4.5: Input radar image feature data:

[0232]

[0233] In this example, N is 5559, which represents the number of scan lines.

[0234] Compute the radial basis function as the node activation function:

[0235]

[0236] In this example, R∈[0,1] represents the hidden layer neuron value.

[0237] Step 1.4.6: Calculate neuron output:

[0238]

[0239] In this example, y k is the kth output element in the output vector Y, where the output vector is:

[0240] Y=[y1,y2,…y q ] T

[0241] The expected output vector is a zero vector, specifically:

[0242] O=[o1,o2,…o q ] T

[0243] In this example, q is 330, which represents the number of hidden layer units.

[0244] Step 1.4.7: The RBF neural network uses the gradient descent method to adaptively adjust the RBF neural network parameters connection weight vector, center parameter vector and width vector to the optimal value; the connection weight vector, center parameter vector and width vector calculation formula are:

[0245]

[0246]

[0247]

[0248] In this example, W represents a 1×300 training weight vector, C represents a 1×550 center parameter vector, and D represents a 1×550 width vector.

[0249] Step 1.4.8: After the center, width and weight are updated, the values ​​of η and α in this example are 1 and 3 respectively, and the value of the iteration termination accuracy ε is 0.005. The value of the root mean square error RMS of the network output is calculated as follows:

[0250]

[0251] In this example, O represents the output 0 matrix, and N is 5559, which represents the number of scan lines. If RMS ≤ ε, the training ends, otherwise the updated center, width, and weight are recalculated and updated again until the conditions are met.

[0252] Step 2: Get the echo data to be measured on a single scanning line of a radar as shown in the attached figure. Figure 7 According to step 1.2, this example selects 7 radar echo sample data features, including mean, variance, fitting correlation coefficient, variance coefficient, weighted filling ratio, discrete coefficient and maximum fluctuation as feature parameters and calculates the feature values.

[0253] Step 2.1.1, the mean, variance, and fitting correlation coefficient characteristic parameter calculations are:

[0254]

[0255]

[0256]

[0257] In this example, n is 550, which indicates the number of sampling points of a single scan line, σ∈[0,2.5] indicates the grayscale value, θ∈[0°,360°] indicates the azimuth angle, and r∈[0,4500](m) indicates the radial distance. The result representing the radar echo intensity after fitting; represents the mean value of the radar echo intensity of the selected line; z∈[0,5] represents the radar echo intensity value before fitting. The calculated mean value of this example is 0.8017, the variance is 0.5197, and the fitting correlation coefficient is 0.7829.

[0258] Step 2.1.2, the variance coefficient and weighted filling ratio are:

[0259]

[0260]

[0261] In this example, σ∈[0,2.5] represents the grayscale value, θ∈[0°,360°] represents the azimuth angle, and r∈[0,4500](m) represents the radial distance. n is 550, which represents the number of sampling points in a single scan line, and μ represents the mean. The calculated variance coefficient of this example is 0.5769, and the weighted fill ratio is 0.2363.

[0262] Step 2.1.3, the coefficient of dispersion and the maximum fluctuation are:

[0263]

[0264] F1=max(σ(θ,r i ))-μ

[0265] In this example, σ∈[0,2.5] represents the grayscale value, θ∈[0°,360°] represents the azimuth angle, and r∈[0,4500](m) represents the radial distance. μ represents the mean, and s represents the standard deviation. The calculated dispersion coefficient of this example is 0.7966, and the maximum fluctuation is 0.2530.

[0266] Step 3, the radar echo feature parameters in step 2 are subjected to feature fusion according to step 1.

[0267] Step 3.1, the optimal feature covariance matrix calculation formula is:

[0268]

[0269] In this example, Tp represents the m×n training sample matrix of the test data, m is 1, which represents the number of test data, n is 7, which represents the number of features to be obtained, and R represents the 7×7 covariance matrix.

[0270] Step 3.2, R eigenvalue calculation formula:

[0271] |λ i IR|=0(i=1,2.....n)

[0272] The calculated eigenvalues ​​of this example are 1.5728, 0.1801, 0.2361, 0.0195, 0.0038, 1.3262×10 -4 , 9.2261×10 -7 .

[0273] R eigenvector calculation formula:

[0274] R i =λ i b i

[0275] Step 3.3, projection matrix C calculation formula:

[0276] C=[b a..... b j ]

[0277] In this example, the eigenvectors corresponding to the three eigenvalues ​​λ1=1.5728, λ2=0.1801, and λ3=0.2361 are selected to form a 7×3 projection matrix C.

[0278] Step 3.4, the calculation formula of the training sample set after feature fusion is:

[0279] T PF =T P C

[0280] Echo feature parameters after feature fusion:

[0281] Q=[q a ,....,q j ]

[0282] In this example, Q is a 1×3 matrix, which means that the eigenvectors corresponding to the eigenvalues ​​λ1, λ2, and λ3 constitute the feature parameters corresponding to the projection features. The three fused feature parameters are obtained, and the input feature matrix of this example is [0.8017, 0.5197, 0.7829].

[0283] Step 4: Input the feature matrix finally calculated in step 3 into the RBF classifier model trained in step 1, and finally obtain an output y value. In this example, the output y value is 0.9.

[0284] Step 5: Determine whether there is a target on the selected scan line:

[0285] The final output y value is obtained by using ; it is compared with the expected output value 0. If the output is greater than the expected output value 0, there is a target; if the output is less than the expected output value 0, there is no target. In this example, the final output y value of the selected scan line is 0.9, which is greater than 0, so the selected scan line has a target.

[0286] During 2016-2018, the experimental ship acquired a large amount of radar data and sea condition information during the relevant period while sailing in the East China Sea. Figure 6 The performance indicators of the classifier and feature fusion classifier in this paper's method, CFAR, RBF neural network are shown in Table 1.

[0287] Table 1. Feature fusion classifier performance indicators

[0288]

[0289] It can be seen from Table 1 that the detection performance of this paper is better, the precision and accuracy are higher, and the detection time is shorter.

[0290] The marine radar target rapid detection technology based on multi-feature fusion proposed in the present invention is not only effective in actual measurement, but also has a shorter detection time, and can be widely promoted and applied in marine observation equipment.

Claims

1. A method for detecting a marine radar target, characterized in that: include: Step 1: Select radar echo data samples on a single scan line from the offline radar data, wherein the samples include samples containing target radar echoes and samples of pure sea clutter radar echoes without targets; Select radar echo feature parameters, calculate feature parameter values, perform feature fusion on the selected radar echo feature parameters to obtain fused echo feature parameters and a sample training set, calculate fused echo feature parameter values, input the fused echo feature parameter values ​​into the RBF neural network for training, and obtain a feature-fused RBF classifier; The fusion echo feature parameter value is input into the RBF neural network for training to obtain the RBF classifier after feature fusion, which includes: Step 1.1: Input the fusion echo characteristic parameter values ​​of the sample training set into the RBF neural network; Step 1.2: Initialize the connection weight vector from the hidden layer to the output layer: W m =[w m1 ,w m2 ,…w mq ] T ,(m=1,2,…,q) In the formula, q is the number of hidden layer units, and m represents the mth hidden layer; Initialize the weights from the hidden layer to the output layer: Where mink is the minimum value of the result predicted by the output of the k-th neuron in the training sample; maxk is the maximum value of the result predicted by the output of the k-th neuron in the training sample, and j represents the j-th neuron in the hidden layer; Step 1.3: Initialize the central parameter vector of each neuron in the hidden layer: C j =[c j1 ,c j2 ,…c jn ] T In the formula, j represents the jth neuron in the hidden layer, and n represents the number of sampling points in a single scan line; The initial values ​​of the central parameters of the RBF neural network are: Where min i is the minimum value of all input information of the i-th feature in the training set, max i is the maximum value of all input information of the i-th feature in the training set, and p represents the total number of neurons in the hidden layer; Step 1.4: Initialize the width vector: D j =[d j1 ,d j2 ,…d jn ] T In the formula, j represents the jth neuron in the hidden layer, and n represents the number of sampling points in a single scan line; The width vector is: Where, d f is the width adjustment coefficient, x ij represents the i-th eigenvalue of the j-th scan line, c ij represents the central parameter value of the i-th eigenvalue of the j-th scan line; Calculate the output of the jth hidden layer neuron: In the formula, X represents the input feature vector, C represents its central parameter vector, and D represents the width vector; Step 1.5: Input radar image feature data: Where N represents the number of scan lines; Compute the radial basis function as the node activation function: Where R is the hidden layer neuron value; Step 1.6: Calculate neuron output: In the formula, y k is a specific output element in the output vector Y, where the output vector is: Y=[y1,y2,…y q ] T The expected output vector is a zero vector, specifically: O=[o1,o2,…o q ] T Where q represents the number of hidden layer units; Step 1.7: The RBF neural network uses the gradient descent method to adaptively adjust the RBF neural network parameters to the optimal value. The iterative calculation formulas of the connection weight vector, the center parameter vector and the width vector are: In the formula, W represents the training weight vector, C represents its central parameter vector, and D represents the width vector; Step 1.8: After the center, width and weight are updated, the root mean square error RMS value of the network output is calculated according to the given values ​​of η and α and the value of the iteration termination accuracy ε: Where O represents the output expected matrix, N is the number of scan lines, if RMS ≤ ε, the training ends, otherwise return to step 1.7 to iteratively calculate the connection weight vector, center parameter vector and width vector; Step 2: Acquire the echo data to be measured on a single scanning line in the radar image, and calculate the echo characteristic value of the echo data to be measured according to the echo characteristic parameters selected in step 1; Step 3: perform feature fusion on the echo feature parameters in step 2 according to step 1 to obtain a feature matrix; Step 4: Input the feature matrix calculated in step 3 into the RBF classifier model obtained in step 1, and finally obtain the output y value; Step 5: Determine whether there is a target on the scan line: compare the output y value with the expected output value. If y is greater than the expected output value, there is a target; otherwise, there is no target.

2. A marine radar target detection method according to claim 1, characterized in that: The radar echo characteristic parameters are mean, variance, fitting correlation coefficient, variance coefficient, weighted filling ratio, dispersion coefficient and maximum fluctuation.

3. A marine radar target detection method according to claim 1, characterized in that: The feature fusion of the selected feature parameters to obtain the fused echo feature parameters and sample training set includes: Step a: Calculate the optimal inter-feature covariance matrix: Where, T p represents the m×n training sample matrix of the test data, m represents the number of training set data, n represents the number of selected radar echo characteristic parameters, and R represents the n×n covariance matrix; Step b: Calculate the R eigenvalue as: |λ i I-R|=0(i=1,2.....n) R eigenvector calculation formula: Rb i =λ i b i Among them, b i represents the i-th eigenvector of R, λ i represents the i-th eigenvalue of R; Step c: Calculate the projection matrix C: C=[b a ,.....b j ] Where b a ......b j It represents the eigenvectors corresponding to the first x eigenvalues ​​in descending order of eigenvalues. C is an n×x matrix, where x represents the number of retained eigenvectors. Step d: Calculate the sample training set after feature fusion: T PF =T P C The echo feature parameters after feature fusion are: Q=[q a ,....,q j ] In the formula, q a ......q j Denotes the eigenvalue λ a ......λ j The corresponding characteristic parameters.

4. A method for rapid detection of marine radar targets according to claim 1, characterized in that: The radar is an X-band marine radar.

Images