Three-feature sea surface target detection method based on smoothing filtering

Abstract

The application provides a three-feature sea surface target detection method based on smoothing filtering, and the implementation steps are as follows: initializing parameters; dividing each group of sea clutter data; performing smoothing filtering on each group of sea clutter data; extracting three-dimensional features of each data block in which the clutter is suppressed; and obtaining a target detection result under the sea clutter background. The application reconstructs each data block by using part of data of the adjacent data block of each distance unit, the reconstructed data block contains part of historical data and future data of the adjacent data block, the smoothing filter with the coefficient a is used to perform smoothing filtering on each data block by calculating the coefficient of the smoothing filter with the coefficient a from all the reconstructed data blocks, the ability to suppress the clutter in the data block is improved, the signal-to-clutter ratio of the target echo is improved, the defect that the existing technology is difficult to separate the small radar cross section (RCS) maneuvering target from the clutter is avoided, and the detection probability is effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of radar technology and relates to a method for detecting sea surface targets, specifically a three-feature sea surface target detection method based on smoothing filtering. Background Technology

[0002] Sea clutter is related to natural factors such as sea conditions, climate, wind speed, and wind direction, as well as many factors such as radar frequency, polarization, resolution, and ground-touching angle. It exhibits complex characteristics such as temporal non-stationarity, spatial non-uniformity, and non-Gaussian statistical properties. Effective detection of weak moving targets against the background of sea clutter is one of the important tasks of shore-based and shipborne maritime surveillance radars.

[0003] With the widespread application of coherent radar systems, the Doppler spectrum of radar echoes carries a wealth of information that incoherent radar systems cannot provide, such as Doppler velocity, Doppler bandwidth, and frequency domain energy distribution. Since traditional sea surface target detection methods do not fully utilize the information in the sea clutter data returned by coherent radar, single-feature-based sea surface target detection methods have been proposed. These methods use feature information extracted from the sea clutter data returned by coherent radar to distinguish between targets and clutter. However, single-feature extraction does not fully utilize the information in the sea clutter data returned by coherent radar, and single features are difficult to effectively detect small sea surface targets across all scenarios.

[0004] To address this issue, many multi-feature-based methods for sea surface target detection have been proposed. For example, Z.-X. Guo et al. published their paper, "Fast Dual Trifeature-Based Detection of Small Targets in Sea Clutter by Using Median Normalized Doppler Amplitude Spectra," in the IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. In *Sensing*, vol. 16, pp. 4050-4063, 2023, doi:10.1109 / JSTARS.2023.3268181., a three-feature-based method for sea surface target detection was proposed. This method divides the reference range cell and the target range cell of the sea clutter data returned from coherent radar into blocks. For each data block, three features are extracted: relative average amplitude, relative Doppler peak height, and relative Doppler spectral entropy, to obtain the feature sample sets of the reference range cell and the target range cell, respectively. Finally, the detection decision region is calculated using the reference range cell sample set of the sea clutter data for target detection. This method makes full use of the amplitude and Doppler information of the sea clutter data, improving the target detection probability. However, for maneuvering targets with small radar cross-section (RCS), the signal-to-clutter ratio of the target echo data received by the radar is low, making it difficult for this method to separate such targets from clutter, resulting in a still low target detection probability. Summary of the Invention

[0005] (1) Initialize parameters:

[0006] Initialize the Q groups of sea clutter data to be detected, where each group of sea clutter data includes K range cells consisting of P reference range cells and T range cells to be detected. Initialize the smoothing filter to the order of M, where Q≥2, P≥2, T≥1, K≥2, M≥2.

[0007] (2) Divide each group of sea clutter data:

[0008] Each reference range cell and each range cell to be detected in the q-th sea clutter data group are divided into J data blocks, each containing N pulses, resulting in P×J data blocks of P reference range cells. T×J data blocks of T distance units to be detected Then the K×J data block with K distance units is in, These represent the p-th reference distance cell, the t-th distance cell to be detected, and the j-th data block of the k-th distance cell, respectively.

[0009] (3) Perform smoothing filtering on each group of sea clutter data:

[0010] Through each data block of each reference distance unit Partial data pairs of adjacent data blocks Reconstruct the data and use the reconstructed P×J data blocks. Calculate the coefficients 'a' of the smoothing filter, and then apply the smoothing filter to each data block for each distance cell. Smoothing filtering is performed to obtain K×J data blocks after clutter suppression. in, The j-th reconstructed data block of the p-th reference distance cell is The j-th clutter-suppressed data block in the k-th range cell is

[0011] (4) Extract the three-dimensional features of each data block with suppressed clutter:

[0012] Calculate the clutter suppression for each data block relative average amplitude Relative Doppler peak height and relative Doppler spectral entropy And and As For the three-feature sample, then The corresponding set of three features, including P×J suppressed clutter samples, is as follows: The corresponding set of three characteristic samples, including T×J clutter samples, is as follows: They represent The corresponding three-feature samples;

[0013] (5) Obtain target detection results against a sea clutter background:

[0014] Three-feature sample set through reference distance unit Calculate the target detection decision region Ω q Simultaneously calculate the three-feature sample set of the distance cell to be detected. Each three-feature sample Test statistic And judge Is it true? If so, then Falling into the target detection decision region Ω q Inside, representing a data block If no target exists, otherwise it means There is a target within it.

[0015] Compared with the prior art, the present invention has the following advantages:

[0016] This invention reconstructs a data block by using partial data from data blocks adjacent to each data block in each range cell. Since the reconstructed data block contains some historical and future data from adjacent data blocks, each data block is smoothed by a smoothing filter with coefficient 'a' calculated from all the reconstructed data blocks. This improves the ability to suppress clutter in the data block, thereby increasing the signal-to-clutter ratio of the target echo. It avoids the shortcomings of existing technologies that make it difficult to separate maneuvering targets with small radar cross-sections (RCS) from clutter, effectively improving the detection probability. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0018] Figure 2 This is a line graph comparing the target detection probability results of the present invention with those of existing technologies. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0020] Reference Figure 1 The present invention includes the following steps:

[0021] Step 1) Initialize parameters:

[0022] Initialize the Q groups of sea clutter data to be detected, where each group of sea clutter data includes K range cells consisting of P reference range cells and T range cells to be detected. Initialize the smoothing filter to the order of M, where Q≥2, P≥2, T≥1, K≥2, M≥2.

[0023] In this embodiment, Q=10 sets of sea clutter measured data were obtained from the South African Council for Scientific and Industrial Research (CSIR) database: TFA10_004, TFA10_005, TFA10_006, TFA10_007, TFA10_008, TFA17_001, TFA17_002, TFA17_004, TFA17_005, and TFA17_006. The first five sets of sea clutter measured data consist of K=64 range cells, composed of P=61 reference range cells and T=3 range cells to be detected. The latter five sets of sea clutter measured data consist of K=96 range cells, composed of P=93 reference range cells and T=3 range cells to be detected. Among them, TFA10_004, TF... The detection distance units for the four sets of data A10_005, TFA10_006, and TFA10_008 are 15, 16, and 17; the detection distance units for TFA10_007 are 16, 17, and 18; the detection units for TFA17_001_all are 25, 26, and 27; the detection distance units for TFA17_002 are 28, 29, and 30; the detection distance units for TFA17_004 are 27, 28, and 29; the detection distance units for TFA17_005 are 29, 30, and 31; and the detection units for TFA10_006 are 28, 29, and 30. Apart from the detection units, all other units are reference distance units.

[0024] The smoothing filter has an order of M = 8.

[0025] Step 2) Divide each group of sea clutter data:

[0026] Each reference range cell and each range cell to be detected in the q-th sea clutter data group are divided into J data blocks, each containing N pulses, resulting in P×J data blocks of P reference range cells. T×J data blocks of T distance units to be detected Then the K×J data block with K distance units is in, These represent the p-th reference distance cell, the t-th distance cell to be detected, and the j-th data block of the k-th distance cell, respectively.

[0027] In this embodiment, each distance unit is divided into J = 129 data blocks, each containing N = 512 pulses. The first five sets of data yield P × J = 7869 data blocks for P = 63 reference distance units and T × J = 387 data blocks for T = 3 units to be detected. K = 64 distance units are divided into K × J = 8256 data blocks. The last five sets of data yield P × J = 11997 data blocks for P = 93 reference distance units and T × J = 387 data blocks for T = 3 units to be detected. K = 96 distance units are divided into K × J = 12384 data blocks.

[0028] Step 3) Perform smoothing filtering on each group of sea clutter data:

[0029] Through each data block of each reference distance unit Partial data pairs of adjacent data blocks Reconstruct the data and use the reconstructed P×J data blocks. Calculate the coefficients 'a' of the smoothing filter, and then apply the smoothing filter to each data block for each distance cell. Smoothing filtering is performed to obtain K×J data blocks after clutter suppression. in, The j-th reconstructed data block of the p-th reference distance cell is The j-th clutter-suppressed data block in the k-th range cell is

[0030] Through each data block of each reference distance unit Partial data pairs of adjacent data blocks Refactoring is performed, and the refactoring formula is:

[0031]

[0032] in, Represents data block For the nth pulse in the current data block, where n∈1,2,...,N, when j=1, assign 0 to the (1-M)th to (0)th pulses of the current data block, and assign the (N+1)th to (N+M)th pulses of the current data block to the values ​​of the next data block; when j=J, assign 0 to the (N-M+1)th to (N)th pulses of the previous data block to the (1-M)th to (0)th pulses of the current data block, and assign 0 to the (N+1)th to (N+M)th pulses of the current data block; when j=2,...,J-1, assign 0 to the (N-M+1)th to (N)th pulses of the previous data block to the (1-M)th to (0)th pulses of the current data block, and assign 0 to the (N+1)th to (N+M)th pulses of the current data block. Data blocks after reconstruction It contains some historical and future data from adjacent data blocks. At the same time, since sea clutter information is acquired in batches, it is reasonable to use both historical and future information simultaneously.

[0033] The coefficients 'a' of the smoothing filter are calculated using the least squares fitting method, and the formula is as follows:

[0034]

[0035] a m =[-a - (M)-a - (M-1)…-a - (1)1-a + (1)…-a + (M+1)-a + (M)]

[0036] in: This represents the parameter vector a when the minimum value is reached. m The value of ∑ represents the summation operation, C 2M+1 Represents a 2M+1 dimensional vector space, -a - (M), -a + (M) represent a respectively m The first element value and the (2M+1)th element value are due to the reconstructed data block. It contains partial historical and future data information from adjacent data blocks, utilizing data blocks. The calculated smoothing filter with coefficient 'a' can effectively match clutter characteristics;

[0037] For each data block of each distance unit Smoothing filtering is performed; the formula for smoothing filtering is:

[0038]

[0039] Since the calculated smoothing filter can effectively match clutter characteristics, it is applicable to each data block of each distance cell. Smoothing filtering effectively suppressed noise.

[0040] Step 4) Extract the three-dimensional features of each data block where clutter is suppressed:

[0041] Calculate the clutter suppression for each data block relative average amplitude Relative Doppler peak height and relative Doppler spectral entropy And and As For the three-feature sample, then The corresponding set of three features, including P×J suppressed clutter samples, is as follows: The corresponding set of three characteristic samples, including T×J clutter samples, is as follows: They represent The corresponding three-feature samples;

[0042] In each data block where clutter is suppressed relative average amplitude Relative Doppler peak height and relative Doppler spectral entropy The calculation formulas are as follows:

[0043]

[0044]

[0045]

[0046]

[0047]

[0048]

[0049]

[0050] in This indicates the average amplitude operation, Peak(·) is the Doppler peak height, #Δ represents the number of Doppler information elements contained in each Doppler-transformed data block, and f d It's Doppler shift. It is the Doppler shift, X(f) d ) represents the Doppler amplitude spectrum, max{·} denotes the maximum value, argmax{·} denotes the value of the independent variable when the maximum value is reached, and T r This represents the pulse repetition period of the radar, |·| denotes the absolute value operation, exp denotes the exponential operation with the natural constant e as the base, and VE(·) represents the Doppler spectral vector entropy. For the normalized Doppler amplitude spectrum, log represents the logarithmic operation.

[0051] Step 5) Obtain target detection results against the background of sea clutter:

[0052] Three-feature sample set through reference distance unit Calculate the target detection decision region Ω qSimultaneously calculate the three-feature sample set of the distance cell to be detected. Each three-feature sample Test statistic And judge Is it true? If so, then Falling into the target detection decision region Ω q Inside, representing a data block If no target exists, otherwise it means There is a target within;

[0053] Among them, the target detection decision region Ω is calculated. q The fast convex hull learning algorithm is adopted, and the implementation steps are as follows:

[0054] (5a) Initialize the false alarm probability as P F The number of iterations is l, and the maximum number of iterations is L = I × P. F I = P × J, and let l = 1;

[0055] (5b) Calculate the sample set after the l-th iteration. convex hull

[0056]

[0057] in, This represents the sample set after the l-th iteration. The w-th element that forms the convex hull surface after the l-th iteration represents the convex hull surface. l The triangle is defined by its three vertices, and triangle(·,·,·) represents a triangle composed of these three vertices. l SP represents the total number of triangles that make up the convex hull, and SP represents the total number of triangles that make up W. l The operation of forming a convex hull using triangular faces, with the vertices of the convex hull being... represent The r-th vertex, R represents The total number of vertices;

[0058] (5c) Calculate the removal of each vertex separately. back volume And delete the vertex corresponding to the volume with the smallest value among the R calculated volumes. Achieve The update will be the updated sample set. The convex hull is used as the detection decision region for the l-th iteration. in:

[0059]

[0060]

[0061] (5d) Determine whether l = L is true. If so, obtain the detection decision region. Otherwise, let l = l + 1 and proceed to step (5b);

[0062] In this embodiment, the false alarm probability is P. F =0.001;

[0063] The fast convex hull learning algorithm used to calculate the detection decision region has certain advantages in terms of efficiency, versatility, ease of implementation, and space efficiency. It is suitable for processing point sets of various sizes and shapes, and also has good processing capabilities for the randomly distributed point sets in this invention. The algorithm is relatively simple to implement, easy to understand and debug. Furthermore, the fast convex hull algorithm usually does not need to generate all intermediate data structures during the process of solving the convex hull, and can solve it with relatively small space complexity, thus enabling efficient calculation of the detection decision region.

[0064] Each three-feature sample Test statistic The calculation formula is:

[0065]

[0066] in, Indicates the detection decision region Ω q The wth surface L The three vertices of the triangle, W L Indicates the composition of Ω q The total number of triangular faces, det(·) is the determinant calculation operation.

[0067] The technical effects of this invention will be further explained below with reference to simulation experiments:

[0068] 1. Experimental conditions and contents:

[0069] The simulation experiment of this invention was run on an Intel(R) Core i7-6700 CPU@3.40GHz, a 64-bit Windows operating system, and the simulation software used was Matlab R2021b.

[0070] The detection probabilities of this invention and existing three-feature-based sea surface target detection methods are compared through simulation, and the results are as follows: Figure 2 As shown.

[0071] 2. Analysis of experimental results:

[0072] Reference Figure 2The horizontal axis represents 10 sets of sea clutter data, and the vertical axis represents the target detection probability. Line segments with cross-shaped dots represent the target detection probability of this invention, while line segments with hollow dots represent the target detection probability of the prior art. Figure 2 As can be seen, the detection results of the present invention on 10 sets of sea clutter data are mostly higher than those of the prior art. After calculation, the average detection probabilities of the present invention and the prior art are 82% and 70% respectively, which is an improvement of 12%.

Claims

1. A three-feature sea surface target detection method based on smoothing filtering, characterized in that, Includes the following steps: (1) Initialize parameters: Initialize the object to be detected Groups of sea clutter data, each group of sea clutter data including data from... Reference distance cells and Composed of individual distance units to be detected The distance unit is used to initialize the smoothing filter to the order of . ,in, , , , , ; (2) Divide each group of sea clutter data into subgroups: The first Each reference range cell and each detectable range cell in the sea clutter data set are respectively divided into: Each includes A data block of pulses is obtained Reference distance cells Data blocks and One distance unit to be detected Data blocks ,but distance units Data blocks are ,in, , , They represent the first The reference distance cell, the first The first distance unit to be detected, the first The distance unit of the first One data block; (3) Smooth the sea clutter data for each group: Through each data block of each reference distance unit Partial data pairs of adjacent data blocks Perform refactoring, and then use the refactored version. Data blocks Calculate the coefficients of the smoothing filter Then, a smoothing filter is applied to each data block of each distance unit. Smoothing filtering is performed to obtain clutter-suppressed clutter. Data blocks ,in, The first in The first reference distance unit The reconstructed data blocks are , The first in The distance unit of the first Each clutter-suppressed data block is ; coefficients of the smoothing filter The least squares fitting method is used for calculation: ； ； in: Represents the parameter vector when the minimum value is taken. The value, This represents the summation operation. express 3D vector space, , They represent The first element value, the second Each element value; (4) Extract the three-dimensional features of each data block where clutter is suppressed: Calculate the clutter suppression for each data block relative average amplitude Relative Doppler peak height and relative Doppler spectral entropy and will , and As For the three-feature sample, then Corresponding to include The three-feature sample set of suppressed clutter is as follows: , Corresponding to include The three-feature sample set of suppressed clutter is as follows: , , They represent , The corresponding three-feature samples; (5) Obtain target detection results against a sea clutter background: Three-feature sample set through reference distance unit Calculate the target detection decision region Simultaneously calculate the three-feature sample set of the distance cell to be detected. Each three-feature sample Test statistic and judge Is it true? If so, then Falling into the target detection judgment area Inside, representing a data block If no target exists, otherwise it means There is a target within it.

2. The method according to claim 1, characterized in that, The step (3) described above involves each data block with each reference distance unit. Partial data pairs of adjacent data blocks Refactoring is performed, and the refactoring formula is: ； in, Represents data block The first in One pulse, .

3. The method according to claim 1, characterized in that, Step (3) describes processing each data block for each distance unit. Smoothing filtering is performed; the formula for smoothing filtering is: 。 4. The method according to claim 1, characterized in that, Each data block in step (4) where clutter is suppressed relative average amplitude Relative Doppler peak height and relative Doppler spectral entropy The calculation formulas are as follows: ； ； ； ； ； ； ； in This indicates an average amplitude operation. It is higher than Doppler Peak. This indicates the number of Doppler information elements contained in each Doppler-transformed data block. It's Doppler shift. It's Doppler shift. It is the Doppler amplitude spectrum. This indicates taking the maximum value. This represents the value of the independent variable when it reaches its maximum value. It is the pulse repetition period of the radar. This indicates the absolute value operation. This represents an exponential operation with the natural constant e as the base. Represents the entropy of the Doppler spectral vector. For the normalized Doppler amplitude spectrum, This indicates the logarithmic operation.

5. The method according to claim 1, characterized in that, The calculation of the target detection decision region described in step (5) The fast convex hull learning algorithm is adopted, and the implementation steps are as follows: (5a) Initialize the false alarm probability as The number of iterations is The maximum number of iterations is , and order ; (5b) Calculate the first Sample set after the second iteration convex hull : ； in, Indicates the first The sample set after the next iteration Indicates the first After the nth iteration, the convex hull surface is formed. The three vertices of a triangle. This represents a triangle formed by three vertices. This represents the total number of triangular faces that make up the convex hull. Indicates will The operation of forming a convex hull using triangular faces, with the vertices of the convex hull being... , represent No. One vertex, express The total number of vertices; (5c) Calculate the removal of each vertex separately. back volume And delete the calculated The vertex corresponding to the volume with the smallest value among the volumes. To achieve The update will be the updated sample set. The convex hull as the first The detection decision region of the next iteration ,in: ； ； (5d) Judgment Whether it is valid or not, if so, obtain the detection judgment area. Otherwise, let Then proceed with step (5b).

6. The method according to claim 1, characterized in that, Each three-feature sample described in step (5) Test statistic The calculation formula is: ； in, Indicates the detection judgment area The surface of the first The three vertices of the triangle face, Indicates composition The total number of triangular faces, It is an operation for calculating determinants.

Images