A method for dynamic adjustment of false alarm rate for maritime target feature detection

Abstract

The present invention relates to a method for dynamically adjusting the false alarm rate for marine target feature detection, belonging to the technical field of radar target detection. The method comprises the following steps: 1) extracting features from the time and frequency domains of radar echo signals; 2) training a convex hull detector and a support vector machine (SVM) detector based on the extracted clutter features and target echo features to perform preliminary control of the false alarm rate; 3) performing detection based on the input echo data, calculating the total alarm rate over each period as an estimate of the actual false alarm rate, and determining an adjustment range for the decision region based on the actual false alarm rate and a preset false alarm rate to readjust the false alarm rate; 4) returning to step 3) to continue detection; and 5) validating the proposed method using measured data. The present invention is applied to a convex hull detector and a support vector machine (SVM) detector to form a target detection method, thereby improving the detector's false alarm control capability.

Description

Technical Field

[0001] The invention relates to a false alarm rate dynamic adjustment method for marine target feature detection, and belongs to the technical field of radar target detection. Background Art

[0002] False alarm rate control has always been an important issue in the field of radar detection. Feature detection has better detection performance than traditional detection methods, but features are different from amplitudes, and it is impossible to provide the optimal detection solution from the perspective of probability density. The false alarm rate control method for feature detectors also remains at the control of the initial false alarm rate. In feature detection, due to the quantitative differences between clutter data and target data, the detection problem is often classified as anomaly detection. The false alarm rate control link of the detector can also be regarded as a part of the understanding of sea clutter characteristics. It is well known that the spatial and temporal non-Gaussian and non-stationary characteristics of sea clutter are the fundamental reason for the difficulty in understanding sea clutter characteristics. This means that to achieve true constant false alarm detection, it is necessary to perceive the sea clutter characteristics that change with space and time in large scenes in real time.

[0003] Current feature detection false alarm rate control methods focus on maintaining the detector's false alarm rate at a preset level within the range of training clutter data. The spatial probability distribution of sea clutter features must be identical to the spatial probability distribution of sea clutter features used when training the detector, a fundamental requirement for these false alarm rate control methods to remain effective when receiving clutter data other than the training clutter data. However, due to the spatiotemporal nonstationarity and inhomogeneity of sea clutter, and the degree of instability of nonstationarity and inhomogeneity in the spatial dimension is often greater than in the temporal dimension, even with sufficient clutter data collected from a sufficient number of range-resolution units, it is impossible to accurately estimate the spatial probability density of the current clutter with time-varying nonstationarity. Instead, the only solution is to ensure that the clutter training sample set contains as many variations in the feature distribution as possible due to clutter nonstationarity. Therefore, increasing the number of training samples for the feature detector not only increases the time required for a single detection but also hinders false alarm rate control during the detection process.

[0004] Therefore, there is an urgent need for an online adjustment method for false alarm rate with low time cost, which can replan the judgment area according to the real-time sea clutter characteristics to cope with the sea clutter characteristics that change with time and space, which is very important for controlling the false alarm rate. Summary of the Invention

[0005] In order to address the deficiencies of the above-mentioned prior art, a method for dynamically adjusting the false alarm rate of marine target feature detection is provided. The method is applied to a convex hull detector and a SVM detector to form a target detection method, thereby improving the false alarm control capability of the detector.

[0006] The present invention provides a method for dynamically adjusting the false alarm rate of marine target feature detection, which is special in that it includes the following steps:

[0007] Step 1: Extract features from the time and frequency domains of radar echo signals;

[0008] Step 2: Train the convex hull detector and SVM detector based on the extracted clutter features and target echo features, and use the initial false alarm rate control method to preliminarily control the false alarm rate according to the preset false alarm rate value;

[0009] Step 3: Perform detection based on the input echo data. After each period of time, the unified alarm rate of that period is calculated as an estimate of the actual false alarm rate. The adjustment range of the judgment area is determined based on the actual false alarm rate and the preset false alarm rate. The false alarm rate is then readjusted to achieve the purpose of narrowing the difference with the preset false alarm rate.

[0010] Step 4: Return to step 3 and continue the inspection until the inspection is completed;

[0011] Step 5: Use measured data to verify the proposed method. Using a channel buoy as the small sea surface target to be detected, the performance of the proposed method is verified under two polarization modes: HH polarization refers to the horizontal polarization of the electromagnetic waves transmitted and received by the radar, and VV polarization refers to the vertical polarization of the electromagnetic waves transmitted and received by the radar.

[0012] Preferably, the features extracted in step 1 are from the time domain and the frequency domain. The following are specific extraction methods for three typical features:

[0013] 1) Relative average amplitude

[0014] The relative average amplitude is defined as the ratio of the time-domain average amplitude of the target unit within the coherent pulse number to the time-domain average amplitude of the reference unit. It is used to reflect the difference between the sea clutter echo and the target echo. The relative average amplitude of the sea clutter unit is 1. Since the amplitude of the target echo is larger than that of the sea clutter echo, the relative average amplitude of the target unit is often greater than that of the sea clutter unit.

[0015]

[0016] Where x represents the echo vector of the unit to be detected with a length of N, x(n) represents the echo sequence of the unit to be detected with a length of N, and n refers to the subscript of the echo sequence. Refers to the average amplitude of the echo of the unit to be detected, x p represents an echo sequence with a reference unit length of N, Refers to the average amplitude of the echo of the reference unit, P is the number of reference units, and ξ1(x) represents the relative average amplitude characteristics of the unit to be detected;

[0017] (2) Relative Doppler peak height

[0018] The Doppler spectrum is calculated as follows:

[0019]

[0020] Where, f d Refers to the Doppler frequency, T r refers to the pulse repetition period, n refers to the echo sequence subscript, N refers to the echo sequence length, x(n) represents the echo sequence with a unit length of N to be detected, X(f d ) refers to the Doppler spectrum of an echo sequence, from which the Doppler peak height and peak height location are extracted;

[0021] The reference Doppler bins are defined using δ1 and δ2, where 2δ2 It refers to the width of the Doppler main peak, and 2(δ1-δ2) refers to the Doppler frequency band unit for reference;

[0022] Δ=[-δ1,-δ2]∪[δ2,δ1]

[0023]

[0024] f d Refers to the Doppler frequency, Refers to X(f d ) takes its maximum value, Δ refers to the reference Doppler frequency range, #Δ refers to the number of Doppler bins falling within the Δ frequency interval, δ1 and δ2 refer to the frequency ranges for estimating relative Doppler peak heights, x represents the echo vector of the target bin with length N, and RDPH(x) refers to the Doppler peak height of the echo vector x of the target bin. After introducing the reference bin, the relative Doppler peak height eigenvalue ξ2 of the target bin is calculated as follows:

[0025]

[0026] ξ2(x) represents the relative Doppler peak height characteristic of the unit to be detected, x represents the echo vector of the unit to be detected with a length of N, x p represents the echo vector of the unit to be detected with a length of N, P refers to the number of reference units, RDPH(x) refers to the Doppler peak height of the echo vector x of the unit to be detected, and RDPH(x p ) refers to the reference unit echo vector x p Doppler peak height.

[0027] (3) Relative Doppler entropy

[0028] The difference in randomness is measured by entropy. The greater the randomness, the greater the amount of information and the greater the entropy.

[0029]

[0030] f d Refers to the Doppler frequency, X(f d ) refers to the Doppler spectrum of the echo train, It refers to the normalized Doppler spectrum of the echo sequence, x represents the echo vector of the unit length N to be detected, and VE(x) refers to the entropy of the echo vector of the unit length N to be detected.

[0031] After the reference unit is introduced, the relative Doppler entropy characteristic value ξ3 of the unit to be detected is calculated as follows:

[0032]

[0033] P refers to the number of reference units, x p represents the echo vector of the unit to be detected with a length of N, x represents the echo vector of the unit to be detected with a length of N, VE(x) refers to the entropy of the echo vector of the unit to be detected with a length of N, VE(x) p ) refers to the entropy of the echo vector of the unit to be detected with a length of N, and ξ3(x) represents the relative Doppler entropy characteristic of the unit to be detected

[0034] Preferably, the specific steps of step 2 are as follows:

[0035] (1) Training of convex hull classifier and false alarm control

[0036] Based on the collected three-feature clutter samples, the clutter feature sample set is converted into a convex area surrounded by triangular sections as the original decision area. Once the feature sample point to be detected falls into the decision area, it is judged as clutter. After the original decision area is formed, the sample points in the decision area are eliminated to reduce the false alarm rate. Based on the regional distribution of the three clutter features and the three target features, sample points whose spatial distribution is more similar to the spatial distribution of the three target features are selected from the three clutter feature sample points used in training, and the following conditions are met:

[0037]

[0038] In the formula The set of training sample points to be deleted, represents the mean point in the original training feature point set, represents the original training feature point set, ξ i =(ξ1,ξ2,ξ3) T Refers to the feature vector sample in the feature point set, H0 means the sample set is a clutter sample set;

[0039] From the training sample point set to be deleted Filter out the sample points as vertices as the sample point set to be eliminated, find the point with the farthest Euclidean distance from the center point of the original sample point set in the sample point set to be eliminated, and use it as the sample point to be eliminated in the current round. After eliminating it, re-form the decision area until the number of remaining sample points and the total number of original sample points meet the following relationship, which means that the initial control of false alarm rate is completed;

[0040]

[0041] P F Refers to the preset false alarm rate, Ω refers to the decision area formed by the clutter feature samples after false alarm rate control, and I refers to the total number of clutter feature samples before false alarm rate control. i ∈Ω} refers to the number of feature sample points in the Ω decision area.

[0042] The decision criterion of the convex hull detector is summarized as follows:

[0043]

[0044] In the formula x Represents the current feature vector to be detected, Represents the coordinates of the three vertices of the triangle that makes up the convex hull surface, arranged clockwise. The subscript q represents the sequence number of the triangle that makes up the convex hull surface, and its maximum value is Q The convex hull surface consists of Q Different triangle vertices are connected to each other. Is the judgment result. If the value is less than zero, it means that the feature vector x to be detected is within the convex hull judgment area, and the feature vector to be detected is judged as a clutter vector. Otherwise, the feature vector to be detected is judged as a target vector.

[0045] (2) SVM classifier training and false alarm control

[0046] According to the collected clutter three-feature samples and target three-feature samples, each feature sample point is marked with the corresponding label y i , where the target feature sample label is +1, indicating the target class, which is represented as a positive sample in the SVM classifier, and the clutter feature sample point is marked as -1, indicating the clutter class, which is represented as a negative sample in the SVM classifier. The corresponding hyperplane is then found using the soft margin SVM constraint criterion;

[0047]

[0048] In the formula, ω and b are the hyperplane weight coefficients, and F i is the i-th feature sample, k(·) is the kernel transformation function, c i is the penalty factor, β0 is the penalty parameter for clutter samples, and β1 is the penalty parameter for target samples;

[0049] To initially control the false alarm rate, the penalty parameter of the clutter sample β0 Will be bounded on the penalty parameter βh and the penalty parameter lower bound β l After repeated adjustments are made, when the hyperplane is formed, the clutter data used in training the SVM is input into the classifier, and the number of clutter feature points misclassified as targets is counted to obtain the false alarm rate within the training sample range.

[0050] When the false alarm rate is high, use the following formula to update the penalty parameter and the upper and lower bounds of the penalty parameter;

[0051]

[0052] β0 is the penalty parameter for clutter samples, β h is the upper bound of the penalty parameter for clutter samples, β l is the lower bound of the penalty parameter for clutter samples.

[0053] When the false alarm rate is low, use the following formula to update the penalty parameter and the upper and lower bounds of the long hair parameter;

[0054]

[0055] β0 is the penalty parameter for clutter samples, β h is the upper bound of the penalty parameter for clutter samples, β l is the lower bound of the penalty parameter for clutter samples.

[0056] This updating step is repeated until the false alarm rate P′ obtained from the training clutter feature samples is F and the preset false alarm rate P f The difference between them is within a preset range, and the control of false alarm rate is completed.

[0057] |P′ F -P f |<η (13)

[0058] P′ F Refers to the false alarm rate obtained from the training clutter feature samples, P f Refers to the preset false alarm rate, η refers to the difference between the two, usually P f / 10.

[0059] The decision criterion of the SVM detector can be summarized as follows:

[0060]

[0061] Where F j is the feature vector to be detected, y j is the corresponding detection result, ω Τ, b is the hyperplane weight coefficient

[0062] Preferably, the specific steps of step 3 are:

[0063] After the initial false alarm control of the detector is completed, the actual detection work begins; the false alarm rate obtained by detecting unknown data is used as a means of perceiving the characteristics of the sea surface. After each pre-set time period, the false alarm rate obtained by the detector detecting the clutter unit during the time period is counted as an estimate of the actual false alarm rate. In order to ensure that the detector can make a correct response to the characteristics of the sea surface, the preset false alarm rate P is set based on a large number of experimental observations. f and the actual false alarm rate P′ F The relationship is established as a linear function with a slope of 1 and a varying intercept:

[0064] P F '=P f +d (15)

[0065] In the actual detection process, since the intercept d is unknown, it is necessary to estimate this value. The intercept d estimate is a reflection of the detector's response to the clutter environment. It is obtained by calculating the difference between the actual false alarm rate and the preset false alarm rate. The intercept d is updated by statistically analyzing the difference between the actual false alarm rate and the preset false alarm rate. The updated relationship is used to guide the update of the detector's decision area. In terms of behavior, the value of the false alarm rate that needs to be maintained is changed when input to the detector. By continuously changing this value, the size of the decision area is controlled, and the adjusted detector is applied to the detection work in the next time period.

[0066] d=P F '-P f (16)

[0067] P F ' refers to the false alarm rate obtained in the actual detection process, P f Refers to the preset false alarm rate

[0068] After estimating the intercept d, use the intercept d to calculate the false alarm rate value P that the detector should be adjusted to at this time F ;

[0069] P F =P Fc -d (17)

[0070] P Fc Refers to the false alarm rate of the current detector setting, P F refers to the false alarm rate that the current detector should be adjusted to, and d refers to the previously estimated intercept.

[0071] The present invention provides a method for dynamically adjusting the false alarm rate of marine target feature detection, which has the following beneficial effects:

[0072] 1. It can reuse training data without re-collecting data within a certain time range, and maintain the false alarm rate near a given false alarm rate under the condition that the fluctuation range of the false alarm rate remains roughly unchanged;

[0073] 2. It can control the false alarm rate at a sufficiently low time cost, thus ensuring the real-time detection to a certain extent;

[0074] 3. This method can be applied to a variety of feature detection methods, which is beneficial for evaluating the performance of detectors. After verification with HH and VV measured data, it can effectively improve the false alarm control capability of detectors. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 It is a three-level sea state relationship diagram between the preset false alarm rate and the actual false alarm rate when the detector receives clutter data other than the training data;

[0076] Figure 2 It is a five-level sea state relationship diagram between the preset false alarm rate and the actual false alarm rate when the detector receives clutter data other than the training data;

[0077] Figure 3 This is a graph showing the change in false alarm rate before and after using the method for level 3 sea conditions;

[0078] Figure 4 This is a graph showing the changes in false alarm rate before and after using the method for level 5 sea conditions. DETAILED DESCRIPTION

[0079] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0080] A method for dynamically adjusting the false alarm rate of marine target feature detection according to this embodiment includes the following steps:

[0081] Step 1: Extract features from the time and frequency domains of radar echo signals;

[0082] Step 2: Train the convex hull detector and SVM detector based on the extracted clutter features and target echo features, and use the initial false alarm rate control method to preliminarily control the false alarm rate according to the preset false alarm rate value;

[0083] Step 3: Perform detection based on the input echo data. After each period of time, the unified alarm rate of that period is calculated as an estimate of the actual false alarm rate. The adjustment range of the judgment area is determined based on the actual false alarm rate and the preset false alarm rate. The false alarm rate is then readjusted to achieve the purpose of narrowing the difference with the preset false alarm rate.

[0084] Step 4: Return to step 3 and continue the inspection until the inspection is completed;

[0085] Step 5: Use measured data to verify the proposed method. Take the channel buoy as the small sea surface target to be detected, and verify the performance of the proposed method under HH and VV polarization modes.

[0086] The echo feature extraction involved in step 1 comes from the time domain and frequency domain. The following introduces three typical feature extraction methods.

[0087] (1) Relative Average Amplitude (RAA)

[0088] This feature is defined as the ratio of the time-domain average amplitude of the target unit within the coherent pulse number to the time-domain average amplitude of the reference unit. It can be used to reflect the difference between the sea clutter echo and the target echo. The relative average amplitude of the sea clutter unit will be close to 1. Because the amplitude of the target echo is often larger than that of the sea clutter echo, the relative average amplitude of the target unit is often greater than that of the sea clutter unit.

[0089]

[0090] Where x represents the echo vector of the unit to be detected with a length of N, x(n) represents the echo sequence of the unit to be detected with a length of N, and n refers to the subscript of the echo sequence. Refers to the average amplitude of the echo of the unit to be detected, x p represents an echo sequence with a reference unit length of N, Refers to the average amplitude of the echo of the reference unit, P is the number of reference units, and ξ1(x) represents the relative average amplitude characteristics of the unit to be detected;

[0091] (2) Relative Doppler Peak Height (RDPH)

[0092] In the Doppler domain, the energy of the target echo Doppler spectrum is more concentrated than that of the clutter spectrum, which means that there is a large difference between the peak value of the target echo spectrum and the spectrum value outside the main lobe. If the peak value is compared with the spectrum value outside the main lobe, a larger value will be obtained. Compared with the target echo Doppler spectrum, the energy distribution of the clutter Doppler spectrum is more even, and the difference between the peak value and the spectrum value outside the main lobe is not as large as that of the target echo Doppler spectrum. The Doppler spectrum can be calculated according to the following formula

[0093]

[0094] Where, fd Refers to the Doppler frequency, Tr refers to the pulse repetition period, n refers to the echo sequence subscript, N refers to the echo sequence length, x(n) represents the echo sequence with a unit length of N to be detected, X(fd ) refers to the Doppler spectrum of an echo sequence, from which the Doppler peak height and peak height location are extracted;

[0095]

[0096] f d Refers to the Doppler frequency, Refers to X(f d ) takes its maximum value, Δ refers to the reference Doppler frequency range, #Δ refers to the number of Doppler bins falling within the Δ frequency interval, δ1 and δ2 refer to the frequency ranges for estimating relative Doppler peak heights, x represents the echo vector of the target bin with length N, and RDPH(x) refers to the Doppler peak height of the echo vector x of the target bin. After introducing the reference bin, the relative Doppler peak height eigenvalue ξ2 of the target bin is calculated as follows:

[0097]

[0098] ξ2(x) represents the relative Doppler peak height characteristic of the unit to be detected, x represents the echo vector of the unit to be detected with a length of N, x p represents the echo vector of the unit to be detected with a length of N, P refers to the number of reference units, RDPH(x) refers to the Doppler peak height of the echo vector x of the unit to be detected, and RDPH(x p ) refers to the reference unit echo vector x p Doppler peak height.

[0099] (3) Relative Doppler entropy (RVE)

[0100] The target echo Doppler spectrum and the clutter Doppler spectrum differ not only in the degree of energy concentration, but also in the degree of chaos. The clutter Doppler spectrum is more random than the target echo Doppler spectrum.

[0101] The difference in randomness can be measured by entropy. The greater the randomness, the greater the amount of information and the greater the entropy.

[0102]

[0103] f d Refers to the Doppler frequency, X(f d ) refers to the Doppler spectrum of the echo train, It refers to the normalized Doppler spectrum of the echo sequence, x represents the echo vector of the unit length N to be detected, and VE(x) refers to the entropy of the echo vector of the unit length N to be detected.

[0104] After the reference unit is introduced, the relative Doppler entropy characteristic value ξ3 of the unit to be detected is calculated as follows.

[0105]

[0106] P refers to the number of reference units, x p represents the echo vector of the unit to be detected with a length of N, x represents the echo vector of the unit to be detected with a length of N, VE(x) refers to the entropy of the echo vector of the unit to be detected with a length of N, VE(x) p ) refers to the entropy of the echo vector of the unit to be detected with a length of N, and ξ3(x) represents the relative Doppler entropy characteristic of the unit to be detected

[0107] The specific steps of step 2 are as follows:

[0108] (1) Training of convex hull classifier and false alarm control

[0109] Based on the collected three-feature clutter samples, the clutter feature sample set is converted into a convex area surrounded by triangular sections as the original decision area. Once the feature sample point to be detected falls into the decision area, it is judged as clutter. After the original decision area is formed, the sample points in the decision area are eliminated to reduce the false alarm rate. Based on the regional distribution of the three clutter features and the three target features, sample points whose spatial distribution is more similar to the spatial distribution of the three target features are selected from the three clutter feature sample points used in training, and the following conditions are met:

[0110]

[0111] In the formula The set of training sample points to be deleted, represents the mean point in the original training feature point set, represents the original training feature point set, ξ i =(ξ1,ξ2,ξ3) T Refers to the feature vector sample in the feature point set, H0 means the sample set is a clutter sample set;

[0112] From the training sample point set to be deleted Filter out the sample points as vertices as the sample point set to be eliminated, find the point with the farthest Euclidean distance from the center point of the original sample point set in the sample point set to be eliminated, and use it as the sample point to be eliminated in the current round. After eliminating it, re-form the decision area until the number of remaining sample points and the total number of original sample points meet the following relationship, which means that the initial control of false alarm rate is completed;

[0113]

[0114] P F Refers to the preset false alarm rate, Ω refers to the decision area formed by the clutter feature samples after false alarm rate control, and I refers to the total number of clutter feature samples before false alarm rate control.i ∈Ω} refers to the number of feature sample points in the Ω decision area.

[0115] The decision criterion of the convex hull detector is summarized as follows:

[0116]

[0117] In the formula x Represents the current feature vector to be detected, Represents the coordinates of the three vertices of the triangle that makes up the convex hull surface, arranged clockwise. The subscript q represents the sequence number of the triangle that makes up the convex hull surface, and its maximum value is Q The convex hull surface consists of Q Different triangle vertices are connected to each other. Is the judgment result. If the value is less than zero, it means that the feature vector x to be detected is within the convex hull judgment area, and the feature vector to be detected is judged as a clutter vector. Otherwise, the feature vector to be detected is judged as a target vector.

[0118] (2) SVM classifier training and false alarm control

[0119] According to the collected clutter three-feature samples and target three-feature samples, each feature sample point is marked with the corresponding label y i , where the target feature sample label is +1, indicating the target class, which is represented as a positive sample in the SVM classifier, and the clutter feature sample point is marked as -1, indicating the clutter class, which is represented as a negative sample in the SVM classifier. The corresponding hyperplane is then found using the soft margin SVM constraint criterion;

[0120]

[0121] In the formula, ω and b are the hyperplane weight coefficients, and F i is the i-th feature sample, k(·) is the kernel transformation function, c i is the penalty factor, β0 is the penalty parameter for clutter samples, and β1 is the penalty parameter for target samples;

[0122] To initially control the false alarm rate, the penalty parameter of the clutter sample β0 Will be bounded on the penalty parameter βh and the penalty parameter lower bound β l After repeated adjustments are made, when the hyperplane is formed, the clutter data used in training the SVM is input into the classifier, and the number of clutter feature points misclassified as targets is counted to obtain the false alarm rate within the training sample range.

[0123] To initially control the false alarm rate, the penalty parameter of the clutter sample β0 Will be bounded on the penalty parameter βh and the penalty parameter lower bound β lAfter repeated adjustments are made, when the hyperplane is formed, the clutter data used in training the SVM is input into the classifier, and the number of clutter feature points misclassified as targets is counted to obtain the false alarm rate within the training sample range.

[0124] When the false alarm rate is high, use the following formula to update the penalty parameter and the upper and lower bounds of the penalty parameter

[0125]

[0126] β0 is the penalty parameter for clutter samples, β h is the upper bound of the penalty parameter for clutter samples, β l is the lower bound of the penalty parameter for clutter samples.

[0127] When the false alarm rate is low, use the following formula to update the penalty parameter and the upper and lower bounds of the long hair parameter.

[0128]

[0129] β0 is the penalty parameter for clutter samples, β h is the upper bound of the penalty parameter for clutter samples, β l is the lower bound of the penalty parameter for clutter samples.

[0130] This updating step is repeated until the false alarm rate P obtained from the training clutter feature samples is F ' and the preset false alarm rate P f The difference between them is within a preset range, and the control of the false alarm rate is completed.

[0131] |P F '-P f |<η (13)

[0132] P F ' refers to the false alarm rate obtained from the training clutter feature samples, P f Refers to the preset false alarm rate, η refers to the difference between the two, usually P f / 10.

[0133] The decision criterion of the SVM detector can be summarized as follows:

[0134]

[0135] Where F j is the feature vector to be detected, y j is the corresponding detection result, ω Τ , b is the hyperplane weight coefficient. The specific steps of step 3 are:

[0136] When the initial false alarm control of the detector is completed, the actual detection work can begin. The non-stationary and non-uniform characteristics of the sea surface mean that in order to maintain the false alarm rate during the detection process, the characteristics of the sea surface must be perceived in real time. The false alarm rate obtained by detecting unknown data can be used as a means of perceiving the characteristics of the sea surface. After each pre-set time period, the false alarm rate obtained by the detector detecting the clutter unit during the time period is counted as an estimate of the actual false alarm rate. In order to ensure that the detector can respond correctly to the characteristics of the sea surface, based on a large number of experimental observations, the preset false alarm rate P can be f The actual false alarm rate P F The relationship is established as a linear function with a slope of 1 and a change in intercept, as shown in the attached figure. Figure 1-2 shown.

[0137] P F '=P f +d (15)

[0138] In the actual detection process, since the intercept d is unknown, it is necessary to estimate the value. d The estimation of is a reflection of the detector to the clutter environment, which is obtained by the difference between the actual false alarm rate and the preset false alarm rate. The difference between the actual false alarm rate and the preset false alarm rate is obtained by statistically analyzing the intercept. d Update the judgment area of ​​the detector using the updated relationship. The behavior is to change the value of the false alarm rate input to the detector that needs to be maintained. By continuously changing this value, the size of the judgment area is controlled and the adjusted detector is applied to the detection work in the next time period.

[0139] d=P F '-P f (16)

[0140] P F ' refers to the false alarm rate obtained in the actual detection process, P f Refers to the preset false alarm rate

[0141] After estimating the intercept d, use the intercept d to calculate the false alarm rate value P that the detector should be adjusted to at this time F ;

[0142] P F =P Fc -d (17)

[0143] P Fc Refers to the false alarm rate of the current detector setting, P F refers to the false alarm rate that the current detector should be adjusted to, and d refers to the previously estimated intercept.

[0144] Specific implementation steps for step 5: Test on HH- and VV-polarized X-band sea detection datasets with channel buoys as the target to be detected. First, extract a portion of the clutter data as training data and train a classifier based on a preset false alarm rate. Then, at the beginning of detection, collect the false alarm rate in the clutter area, correct the preset false alarm rate according to equation (17), and re-form the decision region based on the corrected false alarm rate. Target detection continues using the updated decision region. The false alarm rate control is shown in Table 1.

[0145] Table 1: Comparison of false alarm rate control performance under different data sets

[0146]

[0147] Under the third level of sea conditions, with a preset false alarm rate of 1%, the method can reduce the average false alarm rate difference to below 0.1% or even lower. The curve of the actual false alarm rate change over time before and after the method is used is attached. Figure 3 , Attachment Figure 4 From the attached Figure 3 、 4 It can be seen that the mean of the actual false alarm rate after using the method is closer to the preset false alarm rate.

[0148] Under level 5 sea conditions, the actual false alarm rate fluctuated more violently over time before the method was used. Although the degree of fluctuation did not decrease after the method was used, the actual false alarm rate fluctuated around the preset false alarm rate, and the difference between its mean and the preset false alarm rate was still maintained within 0.1%.

[0149] Regardless of HH, VV polarization or low-altitude sea conditions, the proposed method can improve the false alarm rate control capability of the detector and control the false alarm rate within an ideal range, which is not only beneficial to the detection work but also facilitates more effective evaluation of the detector performance.

Claims

1. A method for dynamically adjusting the false alarm rate of marine target feature detection, characterized in that The following steps are involved: Step 1) Extract features from the time domain and frequency domain of the radar echo signal; Step 2) The convex hull detector and the support vector machine (SVM) detector are trained based on the extracted clutter features and target echo features, and the false alarm rate is preliminarily controlled based on the preset false alarm rate using the initial false alarm rate control method. Step 3) Detection is performed based on the input echo data. The false alarm rate during each period is calculated as an estimate of the actual false alarm rate. The adjustment range of the judgment area is determined based on the actual false alarm rate and the preset false alarm rate. The false alarm rate is then readjusted to reduce the difference from the preset false alarm rate. Step 4) Return to step 3) and continue the inspection until the inspection is completed; Step 5) The proposed method is verified using measured data. A channel buoy is used as the small sea surface target to be detected. The performance of the proposed method is verified under both HH and VV polarization modes (HH polarization refers to the radar transmitting and receiving electromagnetic waves with horizontal polarization, and VV polarization refers to the radar transmitting and receiving electromagnetic waves with vertical polarization).

2. The method for dynamically adjusting the false alarm rate of marine target feature detection according to claim 1 is characterized in that The features extracted in step 1) come from the time domain and frequency domain. The following are the specific extraction methods of three typical features: (1) Relative average amplitude The relative average amplitude is defined as the ratio of the time-domain average amplitude of the target unit within the coherent pulse number to the time-domain average amplitude of the reference unit. It is used to reflect the difference between the sea clutter echo and the target echo. The relative average amplitude of the sea clutter unit is 1. Since the amplitude of the target echo is larger than that of the sea clutter echo, the relative average amplitude of the target unit is often greater than that of the sea clutter unit. ； Where, Represents the echo vector of the unit to be detected with a length of N, Represents an echo sequence with a unit length of N to be detected, Refers to the echo sequence subscript, Refers to the average amplitude of the echo of the unit to be detected, represents an echo sequence with a reference unit length of N, Refers to the average echo amplitude of the reference unit, P is the number of reference units, Represents the relative average amplitude characteristics of the unit to be detected; (2) Relative Doppler peak height The Doppler spectrum is calculated as follows: ； Where, Refers to the Doppler frequency, Refers to the pulse repetition period, Refers to the echo sequence subscript, Refers to the echo sequence length, Represents an echo sequence with a unit length of N to be detected, Refers to the Doppler spectrum of the echo sequence, based on which the Doppler peak height and peak height location are extracted; use 、 The reference Doppler unit is defined as follows: Refers to the width of the Doppler main peak, Refers to the Doppler frequency band unit for reference; ； Refers to the Doppler frequency, refer to The frequency corresponding to the maximum value is, Refers to the Doppler frequency range available for reference, Finger falls into The number of Doppler bins in the frequency bin, 、 Refers to the frequency range for estimating the relative Doppler peak height, Represents the echo vector of the unit to be detected with a length of N, Refers to the echo vector of the unit to be detected The Doppler peak height of the unit to be detected is the relative Doppler peak height characteristic value of the unit to be detected after the reference unit is introduced. Calculated as follows: ； Indicates the relative Doppler peak height characteristics of the unit to be detected, Represents the echo vector of the unit to be detected with a length of N, Represents the echo vector of the unit to be detected with a length of N, P refers to the number of reference units, Refers to the echo vector of the unit to be detected The Doppler peak height, Refers to the reference unit echo vector Doppler peak height; (3) Relative Doppler entropy The difference in randomness is measured by entropy. The greater the randomness, the greater the amount of information and the greater the entropy. ； Refers to the Doppler frequency, Refers to the Doppler spectrum of the echo sequence, refers to the normalized Doppler spectrum of the echo sequence, Represents the echo vector of the unit to be detected with a length of N, Refers to the entropy of the echo vector of the unit to be detected with a length of N; After the reference unit is introduced, the relative Doppler entropy characteristic value of the unit to be detected is Calculated as follows: ； P refers to the number of reference cells, Represents the echo vector of the unit to be detected with a length of N, Represents the echo vector of the unit to be detected with a length of N, Refers to the entropy of the echo vector of the unit to be detected with a length of N, Refers to the entropy of the echo vector of the unit to be detected with a length of N, Represents the relative Doppler entropy characteristics of the unit to be detected.

3. The method for dynamically adjusting the false alarm rate of marine target feature detection according to claim 1 is characterized in that The specific steps of step 2) are as follows: (1) Training of convex hull classifier and false alarm control Based on the collected three-feature clutter samples, the clutter feature sample set is converted into a convex area surrounded by triangular sections as the original decision area. Once the feature sample point to be detected falls into the decision area, it is judged as clutter. After the original decision area is formed, the sample points in the decision area are eliminated to reduce the false alarm rate. Based on the regional distribution of the three clutter features and the three target features, sample points whose spatial distribution is more similar to the spatial distribution of the three target features are selected from the three clutter feature sample points used in training, and the following conditions are met: ； In the formula The set of training sample points to be deleted, represents the mean point in the original training feature point set, represents the original training feature point set, Refers to the feature vector sample in the feature point set, Refers to the sample set as a clutter sample set; From the training sample point set to be deleted Filter out the sample points as vertices as the sample point set to be eliminated, find the point with the farthest Euclidean distance from the center point of the original sample point set in the sample point set to be eliminated, and use it as the sample point to be eliminated in the current round. After eliminating it, re-form the decision area until the number of remaining sample points and the total number of original sample points meet the following relationship, which means that the initial control of false alarm rate is completed; （8）； Refers to the preset false alarm rate, Refers to the decision area formed by the clutter feature samples after false alarm rate control. Refers to the total number of clutter feature samples before false alarm rate control, refer to The number of characteristic sample points in the decision area; The decision criterion of the convex hull detector is summarized as follows: ； In the formula Represents the current feature vector to be detected, Represents the coordinates of the three vertices of the triangle that makes up the convex hull surface, arranged clockwise, with subscripts Represents the sequence number of the triangles that make up the convex hull surface, and its maximum value is The convex hull surface consists of Different triangle vertices are connected to each other. is the judgment result. If the value is less than zero, it means that the feature vector to be detected is In the convex hull judgment area, the feature vector to be detected is judged as a clutter vector, otherwise it is judged as a target vector; (2) SVM classifier training and false alarm control According to the collected clutter three-feature samples and target three-feature samples, each feature sample point is marked with a corresponding label , where the target feature sample label is +1, indicating the target class, which is represented as a positive sample in the SVM classifier, and the clutter feature sample point is marked as -1, indicating the clutter class, which is represented as a negative sample in the SVM classifier. The corresponding hyperplane is then found using the soft margin SVM constraint criterion; ； Where, 、 is the hyperplane weight coefficient, is the i-th feature sample, is the kernel transformation function, is the penalty factor, is the penalty parameter of the clutter sample, is the penalty parameter of the target sample; To initially control the false alarm rate, the penalty parameter of the clutter sample Will be bounded on the penalty parameter and the penalty parameter lower bound After repeated adjustments are made, when the hyperplane is formed, the clutter data used in training the SVM is input into the classifier, and the number of clutter feature points misclassified as targets is counted to obtain the false alarm rate within the training sample range. When the false alarm rate is high, use the following formula to update the penalty parameter and the upper and lower bounds of the penalty parameter; ； is the penalty parameter of the clutter sample, is the upper bound of the penalty parameter for clutter samples, is the lower bound of the penalty parameter for clutter samples; When the false alarm rate is low, use the following formula to update the penalty parameter and the upper and lower bounds of the long hair parameter; ； is the penalty parameter of the clutter sample, is the upper bound of the penalty parameter for clutter samples, is the lower bound of the penalty parameter for clutter samples; This updating step is repeated until the false alarm rate obtained from the training clutter feature samples is and the preset false alarm rate The difference between them is within a preset range, and the control of false alarm rate is completed. （13）； Refers to the false alarm rate obtained from the training clutter feature samples, Refers to the preset false alarm rate, Refers to the difference between the two, usually taken ; The decision criterion of the SVM detector can be summarized as follows: ； In the formula is the feature vector to be detected, is the corresponding test result, 、 is the hyperplane weight coefficient.

4. The method for dynamically adjusting the false alarm rate of marine target feature detection according to claim 1 is characterized in that The specific steps of step 3) are: When the detector has completed the initial false alarm control, the actual detection work begins; the false alarm rate obtained by detecting unknown data is used as a means of perceiving the characteristics of the sea surface. After each pre-set time period, the false alarm rate obtained by the detector detecting the clutter unit during the time period is counted as an estimate of the actual false alarm rate. In order to ensure that the detector can make a correct response to the characteristics of the sea surface, the preset false alarm rate is set based on a large number of experimental observations. The actual false alarm rate The relationship is established as a linear function with a slope of 1 and a varying intercept: （15）； In the actual detection process, due to the intercept is unknown and needs to be estimated. The estimation of is a reflection of the detector to the clutter environment, which is obtained by the difference between the actual false alarm rate and the preset false alarm rate. The difference between the actual false alarm rate and the preset false alarm rate is obtained by statistically analyzing the intercept. Update the judgment area of ​​the detector using the updated relationship. The behavior is to change the value of the false alarm rate input to the detector that needs to be maintained. By continuously changing this value, the size of the judgment area is controlled and the adjusted detector is applied to the detection work in the next time period. （16）； Refers to the false alarm rate obtained during the actual detection process. Refers to the preset false alarm rate; Estimate the intercept Then, use the intercept Calculate the false alarm rate value that the detector should be adjusted to at this time ; （17）； Refers to the false alarm rate of the current detector setting, Refers to the false alarm rate that the current detector should be adjusted to, refers to the previously estimated intercept; Throughout the entire detection process, the detector's decision area continuously changes according to changes in the clutter environment, coping with the sea clutter environment with spatial and temporal non-stationary characteristics and achieving real-time and precise control of the false alarm rate.

Images