Marine target feature detection method applying dynamic time sequence weight

Abstract

The invention relates to a marine target feature detection method applying dynamic time sequence weight, and belongs to the technical field of radar detection algorithms. In order to overcome the defect that time sequence information is not fully utilized during AR model fitting in the prior art, the method comprises the following steps: step 1, carrying out feature extraction on original data to obtain observed values of features; 2, predicting the features obtained in the step 1 to obtain predicted values of the features; step 3, carrying out dynamic weighted fusion on the observation value in the step 1 and the predicted value in the step 2 to obtain a time sequence fusion value of the features; and 4, detecting the time sequence fusion value obtained in the step 3 by adopting a convex hull detector to obtain a detection probability. According to the method, the feature time sequence predicted value and the feature observed value are fused through dynamic calculation, the feature time sequence is fully utilized, the fusion feature variance is reduced, the time sequence feature separability is enhanced, and the target detection probability is improved.

Description

Technical Field

[0001] The invention relates to a method for detecting marine target features, in particular to a method for detecting marine target features using dynamic time series weights, and belongs to the technical field of radar detection algorithms. Background Art

[0002] The detection of maritime targets, especially small targets, has always been the focus and difficulty of research in the field of radar detection. In recent years, with the continuous optimization of radar parameter performance, the information received by radar echoes has become more and more abundant. With the continuous deepening of research on radar echo information characteristics, radar echo information has been mined and utilized more and more fully. At present, people have proposed a large number of features with good separability, such as fractal features, time-frequency features, polarization domain features, graph features, phase domain and singular spectrum domain features. However, these features can only represent the data of the current frame. In order to make better use of the data, people have proposed using the AR model to make a one-step prediction of the features: through Figure 1 The historical frame data shown is fused with the current moment data to add temporal information to the features, thereby making the features more separable.

[0003] The specific calculation method of the current frame feature fusion value using prior information is as follows: , in, is the feature fusion value of the current frame; is the feature observation value of the current frame; is the predicted value of the current frame feature, which is obtained by performing AR fitting and 1-step prediction on the historical frame feature sequence; , are weighted coefficients, which respectively represent the influence of the current frame data and the historical frame data on the current frame feature fusion value. At present, the fusion of the observed value and the predicted value is just to average the two, or adjust the parameters according to the sea condition level. When the sea condition is high, the predicted value is considered to be more valuable, and the corresponding When the sea state is lower, the observation value is considered to be more valuable. Bigger.

[0004] This treatment has the following problems: sea clutter data has time-varying characteristics, and the contributions of predicted values ​​and observed values ​​will also change. Simply averaging the two uses a static weighting algorithm, which does not take into account the time-varying nature of the data and thus cannot fully utilize the time series information. Summary of the invention

[0005] The purpose of the present invention is to solve the problem that the time series information is not fully utilized when fitting an AR model in the prior art, and to provide a method for detecting marine target features using dynamic time series weights.

[0006] In order to solve the above problems, the present application is implemented through the following technical solutions: A method for detecting marine target features using dynamic time series weights is special in that it includes the following steps: Step 1, extract features from the original data to obtain feature observation values: select radar echo data, calculate the relative average amplitude RAA, relative Doppler peak height RDPH, and spectrum peak to mean ratio FPAR to obtain RAA observation sequence, RDPH observation sequence, and FRAR observation sequence composed of M features respectively; Step 2: predict the features obtained in step 1 to obtain the predicted values ​​of the features: adopt the centralized AR(p) model and use q historical feature values ​​to predict each feature in the RAA observation sequence, RDPH observation sequence, and FRAR observation sequence extracted in step 1 to obtain the RAA prediction sequence, RDPH prediction sequence, and FRAR prediction sequence respectively; Step 3: Dynamically weight the observed value in step 1 and the predicted value in step 2 to obtain the time series fusion value of the feature: Select sliding window N, intercept the observation and prediction data of the same segment, calculate the variance of the two segments of data, and find the weighting coefficients of the observation and prediction values ​​according to the minimum variance of the fused data. Use the weighting coefficients as the weighting coefficients of the next step of sliding window to get the fused value. Calculate the fused value at each moment to get the fused value sequence. Step 4: Use a convex hull detector to detect the time series fusion value obtained in step 3 to obtain the detection probability.

[0007] Furthermore, the relative average amplitude RAA in step 1 is calculated as follows: , Where: Represents radar echo data; The RAA observation sequence is .

[0008] Furthermore, the relative Doppler peak height RDPH in step 1 is calculated as follows: ,in, , , c = [ − d 1 , d 2 ] ∪ [ d 2 , d 1 ] , Where: is to make The Doppler frequency with the largest amplitude, Determined by the average Doppler bandwidth of sea clutter, Determined by the guard interval of the Doppler peak, Represents in the interval The number of Doppler cells within; The RDPH observation sequence is .

[0009] Furthermore, the spectrum peak-to-average ratio FPAR in step 1 is calculated as follows: , Where: express through The amplitude spectrum obtained by point discrete Fourier transform; The FRAR observation sequence is .

[0010] Furthermore, in step 2, the centralized AR(p) model is used to analyze the RAA observation sequence obtained in step 1. The fitting results are as follows: , Where: Model order are known system parameters, is Gaussian white noise, model parameters is the parameter to be estimated, and its moment estimate The calculation is as follows: [ r ̂ 1 r ̂ 2 ⋮ r ̂ p ] = [ 1 r ̂ 1 ⋯ r ̂ p − 1 r ̂ 1 1 ⋯ r ̂ p − 2 ⋮ ⋮ ⋮ r ̂ p − 1 r ̂ p − 2 ⋯ 1 ] [ c ̂ 1 c ̂ 2 ⋮ c ̂ p ] , Where: is the estimated value of the autocorrelation coefficient of the characteristic sequence; Take p=5, predict the features at each moment and get the RAA prediction sequence composed of M features .

[0011] Furthermore, in step 2, the centralized AR(p) model is used to analyze the RDPH observation sequence obtained in step 1. The fitting results are as follows: , Where: Model order are known system parameters, is Gaussian white noise, model parameters is the parameter to be estimated, and its moment estimate The calculation is as follows: [ r ̂ 1 r ̂ 2 ⋮ r ̂ p ] = [ 1 r ̂ 1 ⋯ r ̂ p − 1 r ̂ 1 1 ⋯ r ̂ p − 2 ⋮ ⋮ ⋮ r ̂ p − 1 r ̂ p − 2 ⋯ 1 ] [ c ̂ 1 c ̂ 2 ⋮ c ̂ p ] , Where: is the estimated value of the autocorrelation coefficient of the characteristic sequence; Take p=5, predict the features at each moment and get the RDPH prediction sequence composed of M features .

[0012] Furthermore, in step 2, the centralized AR(p) model is used to analyze the FRAR observation sequence obtained in step 1. The fitting results are as follows: , Where: Model order are known system parameters, is Gaussian white noise, model parameters is the parameter to be estimated, and its moment estimate The calculation is as follows: [ r ̂ 1 r ̂ 2 ⋮ r ̂ p ] = [ 1 r ̂ 1 ⋯ r ̂ p − 1 r ̂ 1 1 ⋯ r ̂ p − 2 ⋮ ⋮ ⋮ r ̂ p − 1 r ̂ p − 2 ⋯ 1 ] [ c ̂ 1 c ̂ 2 ⋮ c ̂ p ] , Where: is the estimated value of the autocorrelation coefficient of the characteristic sequence; Take p=5, predict the features at each moment and get a FRAR prediction sequence consisting of M features .

[0013] Furthermore, the relative average amplitude RAA in step 3 is specifically as follows: select sliding window N, intercept the i-th segment data of the observation value , the i-th segment of the predicted value , calculate the variance of two segments of data , ; Assume that the radar echo is a stable signal in a short time, that is, its mean and variance are considered to be approximately unchanged. Take a small sliding window before the feature frame to calculate the weight, in which the mean and variance of the data are approximately unchanged, and the mathematical expectations of the predicted and observed data are the same, which conforms to the normal distribution, that is, , ,in is the mathematical expectation of the observed value and the predicted value, so that the variance of the fused data Minimum, that is, satisfying the following conditions: , Solving the above formula yields: , The solution to , , as the weighted coefficient of the sliding window step, the fusion value is obtained ; Calculate the fusion value at each moment to get the RAA fusion value sequence .

[0014] Furthermore, the relative Doppler peak height RDPH in step 3 is specifically as follows: select a sliding window N, intercept the i-th segment of the observation value , the i-th segment of the predicted value , calculate the variance of two segments of data , ; Assume that the radar echo is a stable signal in a short time, that is, its mean and variance are considered to be approximately unchanged. Take a small sliding window before the feature frame to calculate the weight, in which the mean and variance of the data are approximately unchanged, and the mathematical expectations of the predicted and observed data are the same, which conforms to the normal distribution, that is, , ,in is the mathematical expectation of the observed value and the predicted value, so that the variance of the fused data Minimum, that is, satisfying the following conditions: , Solving the above formula yields: , The solution to , , as the weighted coefficient of the sliding window step, the fusion value is obtained ; Calculate the fusion value at each moment to get the FRAR fusion sequence .

[0015] Furthermore, the spectrum peak-to-average ratio FPAR in step 3 is specifically as follows: select a sliding window N, intercept the i-th segment of the observation value , the i-th segment of the predicted value , calculate the variance of two segments of data , ; Assume that the radar echo is a stable signal in a short time, that is, its mean and variance are considered to be approximately unchanged. Take a small sliding window before the feature frame to calculate the weight, in which the mean and variance of the data are approximately unchanged, and the mathematical expectations of the predicted and observed data are the same, which conforms to the normal distribution, that is, , ,in is the mathematical expectation of the observed value and the predicted value, so that the variance of the fused data Minimum, that is, satisfying the following conditions: , Solving the above formula yields: , The solution to , , as the weighted coefficient of the sliding window step, the fusion value is obtained ; Calculate the fusion value at each moment to get the RDPH fusion sequence .

[0016] Furthermore, the step 4 specifically includes the following steps: Step 41, initializing the convex hull detector: converting the calculated three eigenvalues ​​into eigenvectors in a three-dimensional space, and the three features of each data frame constitute three-dimensional coordinate points in the space; Step 42: Calculate the number of false alarms and remove sample points: Assume that the feature set of small targets and sea clutter is The number of eigenvectors of is , set the false alarm rate to , calculate the number of false alarms as ,turn up The point with the longest Euclidean distance from the mean center of the sea clutter sample point Sample points are removed; Step 43: Generate a convex hull decision area using the sea clutter feature points after removing the sample points from the total sea clutter sample points ; Step 44: Calculate the target feature point falling into the convex hull judgment area The number of outliers divided by the number of target feature points gives the detection probability P d .

[0017] This application fuses the feature time series prediction value and the feature observation value through dynamic calculation, fully utilizes the feature time series, reduces the variance of the fused features, enhances the separability of the time series features, and improves the probability of target detection; the optimized time series feature detection method has strong adaptability, and has been verified by multiple sets of data sets with different sea conditions and different polarizations. The optimized time series detection method has improved detection performance under different conditions compared with the unoptimized time series detection method. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a fusion method of feature space with temporal information; Figure 2 is a flow chart of the present invention; Figure 3 This is a comparison chart of data variance of three methods under HH polarization; Figure 4 This is a comparison chart of data variance of three methods under VV polarization; Figure 5 This is a comparison of the characteristic Bhattacharyya distances of the three methods under HH polarization; Figure 6 This is a comparison of the characteristic Bhattacharyya distances of the three methods under VV polarization; Figure 7 Schematic diagram of convex hull detection; Figure 8 This is a comparison chart of detection probabilities of the three methods under HH polarization; Figure 9 This is a comparison chart of detection probabilities of the three methods under VV polarization; Figure 10 is the curve of detection probability changing with sliding window width under HH polarization; Figure 11 This is the curve of detection probability changing with sliding window width under VV polarization. DETAILED DESCRIPTION

[0019] The following is a specific implementation of the present invention with reference to the accompanying drawings to further illustrate the composition of the present invention.

[0020] Example 1. A method for detecting marine target features using dynamic time series weights, the specific process is as follows: Figure 2 As shown, the following steps are included: Step 1, extract features from the original data to obtain feature observation values: select radar echo data, calculate the relative average amplitude RAA, relative Doppler peak height RDPH, and spectrum peak to mean ratio FPAR to obtain RAA observation sequence, RDPH observation sequence, and FRAR observation sequence composed of M features respectively; Step 2: predict the features obtained in step 1 to obtain the predicted values ​​of the features: adopt the centralized AR(p) model and use q historical feature values ​​to predict each feature in the RAA observation sequence, RDPH observation sequence, and FRAR observation sequence extracted in step 1 to obtain the RAA prediction sequence, RDPH prediction sequence, and FRAR prediction sequence respectively; Step 3: Dynamically weight the observed value in step 1 and the predicted value in step 2 to obtain the time series fusion value of the feature: Select sliding window N, intercept the observation and prediction data of the same segment, calculate the variance of the two segments of data, and find the weighting coefficients of the observation and prediction values ​​according to the minimum variance of the fused data. Use the weighting coefficients as the weighting coefficients of the next step of sliding window to get the fused value. Calculate the fused value at each moment to get the fused value sequence. Step 4: Use a convex hull detector to detect the time series fusion value obtained in step 3 to obtain the detection probability.

[0021] Furthermore, the relative average amplitude RAA in step 1 is calculated as follows: , Where: Represents radar echo data; The RAA observation sequence is .

[0022] Furthermore, the relative Doppler peak height RDPH in step 1 is calculated as follows: , in, , , c = [ − d 1 , d 2 ] ∪ [ d 2 , d 1 ] , Where: is to make The Doppler frequency with the largest amplitude, Determined by the average Doppler bandwidth of sea clutter, Determined by the guard interval of the Doppler peak, Represents in the interval The number of Doppler cells within; The RDPH observation sequence is .

[0023] Furthermore, the spectrum peak-to-average ratio FPAR in step 1 is calculated as follows: , Where: express through The amplitude spectrum obtained by point discrete Fourier transform; The FRAR observation sequence is .

[0024] Furthermore, in step 2, the centralized AR(p) model is used to analyze the RAA observation sequence obtained in step 1. The fitting results are as follows: , Where: Model order are known system parameters, is Gaussian white noise, model parameters is the parameter to be estimated, and its moment estimate The calculation is as follows: [ r ̂ 1 r ̂ 2 ⋮ r ̂ p ] = [ 1 r ̂ 1 ⋯ r ̂ p − 1 r ̂ 1 1 ⋯ r ̂ p − 2 ⋮ ⋮ ⋮ r ̂ p − 1 r ̂ p − 2 ⋯ 1 ] [ c ̂ 1 c ̂ 2 ⋮ c ̂ p ] , Where: is the estimated value of the autocorrelation coefficient of the characteristic sequence; Take p=5, predict the features at each moment and get the RAA prediction sequence composed of M features .

[0025] Furthermore, in step 2, the centralized AR(p) model is used to analyze the RDPH observation sequence obtained in step 1. The fitting results are as follows: , Where: Model order are known system parameters, is Gaussian white noise, model parameters is the parameter to be estimated, and its moment estimate The calculation is as follows: [ r ̂ 1 r ̂ 2 ⋮ r ̂ p ] = [ 1 r ̂ 1 ⋯ r ̂ p − 1 r ̂ 1 1 ⋯ r ̂ p − 2 ⋮ ⋮ ⋮ r ̂ p − 1 r ̂ p − 2 ⋯ 1 ] [ c ̂ 1 c ̂ 2 ⋮ c ̂ p ] , Where: is the estimated value of the autocorrelation coefficient of the characteristic sequence; Take p=5, predict the features at each moment and get the RDPH prediction sequence composed of M features .

[0026] Furthermore, in step 2, the centralized AR(p) model is used to analyze the FRAR observation sequence obtained in step 1. The fitting results are as follows: , Where: Model order are known system parameters, is Gaussian white noise, model parameters is the parameter to be estimated, and its moment estimate The calculation is as follows: [ r ̂ 1 r ̂ 2 ⋮ r ̂ p ] = [ 1 r ̂ 1 ⋯ r ̂ p − 1 r ̂ 1 1 ⋯ r ̂ p − 2 ⋮ ⋮ ⋮ r ̂ p − 1 r ̂ p − 2 ⋯ 1 ] [ c ̂ 1 c ̂ 2 ⋮ c ̂ p ] , Where: is the estimated value of the autocorrelation coefficient of the characteristic sequence; Take p=5, predict the features at each moment and get a FRAR prediction sequence consisting of M features .

[0027] Furthermore, the relative average amplitude RAA in step 3 is specifically as follows: select sliding window N, intercept the i-th segment data of the observation value , the i-th segment of the predicted value , calculate the variance of two segments of data , ; Assume that the radar echo is a stable signal in a short time, that is, its mean and variance are considered to be approximately unchanged. Take a small sliding window before the feature frame to calculate the weight, in which the mean and variance of the data are approximately unchanged, and the mathematical expectations of the predicted and observed data are the same, which conforms to the normal distribution, that is, , ,in is the mathematical expectation of the observed value and the predicted value, so that the variance of the fused data Minimum, that is, satisfying the following conditions: , Solving the above formula yields: , The solution to , , as the weighted coefficient of the sliding window step, the fusion value is obtained ; Calculate the fusion value at each moment to get the RAA fusion value sequence .

[0028] Furthermore, the relative Doppler peak height RDPH in step 3 is specifically as follows: select a sliding window N, intercept the i-th segment of the observation value , the i-th segment of the predicted value , calculate the variance of two segments of data , ; Assume that the radar echo is a stable signal in a short time, that is, its mean and variance are considered to be approximately unchanged. Take a small sliding window before the feature frame to calculate the weight, in which the mean and variance of the data are approximately unchanged, and the mathematical expectations of the predicted and observed data are the same, which conforms to the normal distribution, that is, , ,in is the mathematical expectation of the observed value and the predicted value, so that the variance of the fused data Minimum, that is, satisfying the following conditions: , Solving the above formula yields: , The solution to , , as the weighted coefficient of the sliding window step, the fusion value is obtained ; Calculate the fusion value at each moment to get the FRAR fusion sequence .

[0029] Furthermore, the spectrum peak-to-average ratio FPAR in step 3 is specifically as follows: select a sliding window N, intercept the i-th segment of the observation value , the i-th segment of the predicted value , calculate the variance of two segments of data , ; Assume that the radar echo is a stable signal in a short time, that is, its mean and variance are considered to be approximately unchanged. Take a small sliding window before the feature frame to calculate the weight, in which the mean and variance of the data are approximately unchanged, and the mathematical expectations of the predicted and observed data are the same, which conforms to the normal distribution, that is, , ,in is the mathematical expectation of the observed value and the predicted value, so that the variance of the fused data Minimum, that is, satisfying the following conditions: , Solving the above formula yields: , The solution to , , as the weighted coefficient of the sliding window step, the fusion value is obtained ; Calculate the fusion value at each moment to get the RDPH fusion sequence .

[0030] The variance comparison of different polarization data after non-time series fusion, static weighted fusion, and dynamic weighted fusion algorithms is shown in Figure 2. Figure 3-Figure 4 As shown in the figure, the Bhattacharyya distance between the target characteristic coordinates and the sea clutter characteristic coordinates is as follows: Figure 5-Figure 6As shown. After calculation, under HH polarization, the data variance of the dynamic weighted fusion algorithm is reduced by 8.3441 on average compared with the non-time series fusion, and by 1.1967 on average compared with the static weighted fusion algorithm; under VV polarization, the data variance of the dynamic weighted fusion algorithm is reduced by 7.0802 on average compared with the non-time series fusion, and by 0.9471 on average compared with the static weighted fusion algorithm; under HH polarization, the feature Bhattacharyya distance of the dynamic weighted fusion algorithm is increased by 1.1959 on average compared with the non-time series fusion, and by 0.0441 on average compared with the static weighted fusion algorithm; under VV polarization, the Bhattacharyya distance of the dynamic weighted fusion algorithm is increased by 0.2505 on average compared with the non-time series fusion, and by 0.0545 on average compared with the static weighted fusion algorithm.

[0031] Further, the step 4 is as follows Figure 7 As shown, the specific steps include: Step 41, initializing the convex hull detector: converting the calculated three eigenvalues ​​into eigenvectors in a three-dimensional space, and the three features of each data frame constitute three-dimensional coordinate points in the space; Step 42: Calculate the number of false alarms and remove sample points: Assume that the feature set of small targets and sea clutter is The number of eigenvectors of is , set the false alarm rate to , calculate the number of false alarms as ,turn up The point with the longest Euclidean distance from the mean center of the sea clutter sample point Sample points are removed; Step 43: Generate a convex hull decision area using the sea clutter feature points after removing the sample points from the total sea clutter sample points ; Step 44: Calculate the target feature point falling into the convex hull judgment area The number of outliers divided by the number of target feature points gives the detection probability P d .

[0032] Example 2. This example is a verification test of the method of Example 1.

[0033] This test uses 128 pulses to extract features, performs FFT (Fast Fourier Transform) with 128 points, and uses 100 eigenvalues ​​for one-step prediction. The convex hull algorithm is used to test the target detection probabilities of three methods: no time series fusion, time series fusion using the averaging method, and time series fusion with dynamic weighting.

[0034] The IPIX dataset is used for data verification under different sea conditions under different polarization modes. The parameters corresponding to the file numbers of the IPIX dataset are shown in Table 1. The detection probability comparison of the methods without time series fusion, static weighted fusion, and dynamic weighted fusion under different polarizations and different datasets is shown in Table 1. Figure 8-Figure 9 shown.

[0035] Table 1 Parameters corresponding to the file numbers of the IPIX dataset , After calculation, under HH polarization, the detection probability of the dynamic weighted fusion algorithm is increased by 15.44% compared with the non-time series fusion, and by 2.48% compared with the static weighted fusion algorithm; under VV polarization, the detection probability of the dynamic weighted fusion algorithm is increased by 13.80% compared with the non-time series fusion, and by 2.84% compared with the average weighted fusion algorithm.

[0036] Since radar echo data is only approximated in a very short period of time, the stable sliding window should not be too large. By selecting data under different sea conditions and polarization, it is found that the detection probability changes with the sliding window in a trend of first increasing and then decreasing, and the extreme value is between 10 and 20. Figure 10-Figure 11 As shown, the sliding window N is set to 10 in this application.

[0037] The contribution of the predicted value is larger in high sea conditions, and the contribution of the observed value is larger in low sea conditions. Essentially, the higher the sea conditions, the greater the reflection of sea clutter, and because the sea clutter is time-varying, the echo data is unstable, resulting in a large variance of the eigenvalues ​​and poor data separability. This application dynamically calculates the weighting coefficients from the perspective of variance in order to reduce the variance of the eigenvalues, thereby improving the separability of the features and increasing the probability of detection. Therefore, in different sea conditions or with random disturbances, the performance is improved compared to the original time series detector.

[0038] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the defined scope.

Claims

1. A method for detecting marine target features using dynamic temporal weights, characterized in that: The following steps are involved: Step 1: Extract features from the original data to obtain the observed values ​​of the features: Select radar echo data, calculate relative average amplitude RAA, relative Doppler peak height RDPH, spectrum peak to mean ratio FPAR, and obtain RAA observation sequence, RDPH observation sequence, and FRAR observation sequence composed of M features respectively; Step 2: Predict the features obtained in step 1 to get the predicted values ​​of the features: The centralized AR(p) model is used to predict each feature of the RAA observation sequence, RDPH observation sequence, and FRAR observation sequence extracted in step 1 using q historical feature values, and the RAA prediction sequence, RDPH prediction sequence, and FRAR prediction sequence are obtained respectively; Step 3: Dynamically weight the observed value in step 1 and the predicted value in step 2 to obtain the time series fusion value of the feature: Select sliding window N, intercept the observation and prediction data of the same segment, calculate the variance of the two segments of data, and find the weighting coefficients of the observation and prediction values ​​according to the minimum variance of the fused data. Use the weighting coefficients as the weighting coefficients of the next step of sliding window to get the fused value. Calculate the fused value at each moment to get the fused value sequence. Step 4: Use a convex hull detector to detect the time series fusion value obtained in step 3 to obtain the detection probability.

2. The method for detecting marine target features using dynamic temporal weights according to claim 1, characterized in that: The relative average amplitude RAA in step 1 is calculated as follows: , Where: Represents radar echo data; The RAA observation sequence is .

3. A method for detecting marine target features using dynamic temporal weights according to claim 1 or 2, characterized in that: The relative Doppler peak height RDPH in step 1 is calculated as follows: ,in, , , , Where: is to make The Doppler frequency with the largest amplitude, Determined by the average Doppler bandwidth of sea clutter, Determined by the guard interval of the Doppler peak, Represents in the interval The number of Doppler cells within; The RDPH observation sequence is .

4. The method for detecting marine target features using dynamic temporal weights according to claim 3, characterized in that: The calculation of the spectrum peak-to-average ratio FPAR in step 1 is as follows: , Where: express through The amplitude spectrum obtained by point discrete Fourier transform; The FRAR observation sequence is .

5. The method for detecting marine target features using dynamic temporal weights according to claim 4, characterized in that: In step 2, the centralized AR(p) model is used to analyze the RAA observation sequence obtained in step 1 The fitting results are as follows: , Where: Model order are known system parameters, is Gaussian white noise, model parameters is the parameter to be estimated, and its moment estimate The calculation is as follows: ， Where: is the estimated value of the autocorrelation coefficient of the characteristic sequence; Take p=5, predict the features at each moment and get the RAA prediction sequence composed of M features .

6. The method for detecting marine target features using dynamic temporal weights according to claim 5, characterized in that: In step 2, the centralized AR(p) model is used to analyze the RDPH observation sequence obtained in step 1 The fitting results are as follows: ， Where: Model order are known system parameters, is Gaussian white noise, model parameters is the parameter to be estimated, and its moment estimate The calculation is as follows: ， Where: is the estimated value of the autocorrelation coefficient of the characteristic sequence; Take p=5, predict the features at each moment and get the RDPH prediction sequence composed of M features ; In step 2, the centralized AR(p) model is used to analyze the FRAR observation sequence obtained in step 1 The fitting results are as follows: , Where: Model order are known system parameters, is Gaussian white noise, model parameters is the parameter to be estimated, and its moment estimate The calculation is as follows: ， Where: is the estimated value of the autocorrelation coefficient of the characteristic sequence; Take p=5, predict the features at each moment and get a FRAR prediction sequence consisting of M features .

7. A method for detecting marine target features using dynamic temporal weights according to claim 6, characterized in that: The relative average amplitude RAA in step 3 is specifically as follows: Select sliding window N and intercept the i-th segment of observation data , the i-th segment of the predicted value , calculate the variance of two segments of data , ; In a short period of time, the radar echo is a stable signal, that is, its mean and variance are considered unchanged. A small sliding window is taken before the feature frame to calculate the weight, in which the mean and variance of the data are unchanged, and the mathematical expectations of the predicted value and the observed value are the same, which conforms to the normal distribution, that is, , ,in is the mathematical expectation of the observed value and the predicted value, so that the variance of the fused data Minimum, that is, satisfying the following conditions: , Solving the above formula yields: , The solution to , , as the weighted coefficient of the sliding window step, the fusion value is obtained ; Calculate the fusion value at each moment to get the RAA fusion value sequence .

8. The method for detecting marine target features using dynamic temporal weights according to claim 7, characterized in that: The relative Doppler peak height RDPH in step 3 is as follows: Select sliding window N and intercept the i-th segment of observation data , the i-th segment of the predicted value , calculate the variance of two segments of data , ; In a short period of time, the radar echo is a stable signal, that is, its mean and variance remain unchanged. A small sliding window is taken before the feature frame to calculate the weight, in which the mean and variance of the data remain unchanged, and the mathematical expectations of the predicted value and the observed value are the same, which conforms to the normal distribution, that is, , ,in is the mathematical expectation of the observed value and the predicted value, so that the variance of the fused data Minimum, that is, satisfying the following conditions: , Solving the above formula yields: , The solution to , , as the weighted coefficient of the sliding window step, the fusion value is obtained ; Calculate the fusion value at each moment to get the FRAR fusion sequence ; The spectrum peak-to-average ratio FPAR described in step 3 is specifically as follows: Select sliding window N and intercept the i-th segment of observation data , the i-th segment of the predicted value , calculate the variance of two segments of data , ; In a short period of time, the radar echo is a stable signal, that is, its mean and variance remain unchanged. A small sliding window is taken before the feature frame to calculate the weight, in which the mean and variance of the data remain unchanged, and the mathematical expectations of the predicted value and the observed value are the same, which conforms to the normal distribution, that is, , ,in is the mathematical expectation of the observed value and the predicted value, so that the variance of the fused data Minimum, that is, satisfying the following conditions: , Solving the above formula yields: , The solution to , , as the weighted coefficient of the sliding window step, the fusion value is obtained ; Calculate the fusion value at each moment to get the RDPH fusion sequence .

9. The method for detecting marine target features using dynamic temporal weights according to claim 8, characterized in that: The step 4 specifically comprises the following steps: Step 41, initializing the convex hull detector: converting the calculated three eigenvalues ​​into eigenvectors in a three-dimensional space, and the three features of each data frame constitute a three-dimensional coordinate point in the space; Step 42: Calculate the number of false alarms and remove sample points: Assume that the feature set of small targets and sea clutter is The number of eigenvectors of is , set the false alarm rate to , calculate the number of false alarms as ,turn up The point with the longest Euclidean distance from the mean center of the sea clutter sample point Sample points are removed; Step 43: Generate a convex hull decision area using the sea clutter feature points after removing the sample points from the total sea clutter sample points ; Step 44: Calculate the target feature point falling into the convex hull judgment area The detection probability Pd is obtained by dividing the number of the target feature points.

Images