Sea surface target detection method based on CFAR and five-feature combination

Abstract

The invention discloses a sea surface target detection method based on CFAR and five-feature combination, and the method comprises the steps: carrying out the non-coherent accumulation of original radar echo data, improving the signal-to-clutter ratio, extracting a distance unit passing through a detection threshold through OS-CFAR detection, and shortening the time of subsequent feature detection; the distance units passing through the OS-CFAR detection threshold are sequentially subjected to feature extraction of fluctuation stability, maximum singular value proportion, time-frequency energy ridge confusion degree, relative average amplitude and maximum singular value, and the five features are subjected to serial filtering, so that accurate and rapid detection of the floating small target under the sea clutter background is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar target detection, and in particular relates to a sea surface target detection method based on CFAR and five-feature combination. Background Art

[0002] Detecting faint targets in sea clutter is a major challenge in radar remote sensing. Because floating targets on the sea surface have a small radar cross section (RCS), low velocity, and weak radar returns, they are easily submerged in sea clutter. At low grazing angles, the radar returns of sea spikes are similar to those of the target, resulting in a sharp increase in the false alarm rate (FAR). Therefore, research on sea surface target detection methods based on multi-domain features has important theoretical and practical value.

[0003] Traditional sea surface target detection methods are mostly based on multi-domain features combined with machine learning algorithms to distinguish targets. Existing research methods for sea surface target detection mainly include: Shui team (IEEE Transaction on Aerospace and Electronic Systems, 19(4), 1983: 608-621) uses relative average amplitude, relative Doppler peak height and Doppler entropy to construct a three-dimensional feature space, and applies a convex hull learning algorithm for classification; however, this method extracts features from a single domain to distinguish information, which makes it difficult to effectively distinguish targets from sea clutter, and the convex hull learning algorithm is more complex and difficult to meet real-time requirements. Fan Wanqing (Fan Wanqing. Sea surface target detection technology based on feature transformation driven deep network [D]. University of Electronic Science and Technology of China, 2022.DOI:10.27005 / d.cnki.gdzku.2022.001477.) uses three features combined with machine learning algorithm, using information entropy, maximum singular value, and low-frequency intrinsic mode function energy ratio as three features, and performs classification based on support vector machine. However, this technology takes a long time and the subsequent support vector machine algorithm model is highly complex, which is not conducive to real-time detection. Tian Yufang et al. (Journal of Ocean University of China (Natural Science Edition), 2013, 43(07):104-109) extract the decorrelation time and Hurst exponent to construct a two-dimensional feature space and combine it with the support vector machine algorithm for classification. However, since the characteristic differences between sea clutter and target echo in these domains are not significant, the detection accuracy is low. Summary of the Invention

[0004] To overcome the shortcomings of the aforementioned prior art, the present invention proposes a sea surface target detection method based on CFAR and five combined features. This method improves the signal-to-clutter ratio by performing non-coherent accumulation on the raw radar echo data. Combined with OS-CFAR detection, these methods extract distance cells that pass the detection threshold, reducing the time required for subsequent feature detection. These distance cells that pass the detection threshold are then sequentially subjected to five feature extractions: temporal fluctuation stability, maximum singular value ratio in the singular value domain, time-frequency energy ridge disorder, time-domain relative average amplitude, and maximum singular value in the singular value domain. Extracting diverse features from multiple domains facilitates feature decoupling of target information. Through five-feature serial filtering and CFAR detection, the present invention enables accurate and rapid detection of small floating targets in sea clutter.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A sea surface target detection method based on CFAR and five-feature combination includes the following steps:

[0007] Step 1: Perform non-coherent accumulation on the original radar echo data;

[0008] Step 2: Perform OS-CFAR detection on the radar echo data after non-coherent accumulation and extract the range cells that pass the detection threshold;

[0009] Step 3: Calculate the fluctuation stability characteristic values of the distance cells that pass the OS-CFAR detection threshold, sort them from large to small, and filter out the last half of the distance cells;

[0010] Step 4: Calculate the maximum singular value ratio eigenvalue of the remaining distance units, arrange them in descending order, and filter out the last half of the distance units;

[0011] Step 5: Calculate the time-frequency energy ridge chaos eigenvalues of the remaining distance units, arrange them in ascending order, and filter out the last half of the distance units;

[0012] Step 6: Calculate the relative average amplitude eigenvalues of the remaining distance units, sort them from large to small, and filter out the last half of the distance units;

[0013] Step 7: Calculate the maximum singular value feature of the remaining distance units, extract the distance unit where the maximum singular value is located, and complete the detection of sea surface targets.

[0014] In step 3, the fluctuation stability characteristic values of the distance cells that pass the OS-CFAR detection threshold are calculated, and the latter half of the distance cells are filtered out in descending order, including:

[0015] 3.1) Calculate the fluctuation stability characteristic value W of each distance unit that passes the OS-CFAR detection threshold. The calculation formula is:

[0016] W=std(x)

[0017] Wherein, x represents the echo sequence of the distance unit to be measured;

[0018] 3.2) Arrange all the fluctuation stability eigenvalues W from large to small and filter out the last half of the distance units.

[0019] In step 4, the maximum singular value ratio eigenvalues of the remaining distance units are calculated, and they are arranged in descending order, and the second half of the distance units are filtered out, including:

[0020] 4.1) Rearrange the sample data of each remaining distance unit to construct a Hankel matrix. Perform singular value decomposition on the Hankel matrix to obtain a singular value diagonal matrix. Then, extract the first five singular values in the singular value diagonal matrix and arrange them in descending order.

[0021] 4.2) Calculate the ratio of the first singular value to the sum of the second to fifth singular values to obtain the eigenvalue with the largest singular value ratio;

[0022] 4.3) Arrange all the eigenvalues with the largest singular value proportion from large to small, and filter out the last half of the distance units.

[0023] In step 5, the time-frequency energy ridge chaos eigenvalues of the remaining distance units are calculated, arranged in ascending order, and the last half of the distance units are filtered out, including:

[0024] 5.1) Perform short-time Fourier transform on the sample data of each remaining distance unit;

[0025] 5.2) Modulo the frequency dimension of the sample data after the short-time Fourier transform and perform maximum indexing to obtain an index vector;

[0026] 5.3) Perform standard deviation calculation on the index vector to obtain the time-frequency energy ridge disorder eigenvalue;

[0027] 5.4) Arrange all the time-frequency energy ridge chaos eigenvalues from small to large and filter out the last half of the distance units.

[0028] In step 6, the relative average amplitude eigenvalues of the remaining distance units are calculated, sorted from largest to smallest, and the second half of the distance units are filtered out, including:

[0029] 6.1) Calculate the relative average amplitude eigenvalue X of each remaining distance unit using the following formula:

[0030]

[0031] Among them, x jRepresents the echo sequence of the distance unit to be measured, x ij represents a discrete radar echo sequence;

[0032] 6.2) Arrange all the relative average amplitude eigenvalues X from large to small and filter out the last half of the distance units.

[0033] In step 7, the maximum singular value feature of the remaining distance units is calculated, and the distance unit where the maximum singular value is located is extracted, including:

[0034] 7.1) Rearrange the sample data of each remaining distance unit to construct a Hankel matrix, perform singular value decomposition on the Hankel matrix, and obtain a singular value diagonal matrix;

[0035] 7.2) Arrange the first five singular values in the singular value diagonal matrix in descending order, extract the first singular value in the diagonal matrix, that is, the distance unit where the largest singular value is located, and complete the detection of sea surface targets.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] (1) The features proposed in the present invention have the characteristics of high discrimination: the features of the fluctuation stability and relative average amplitude in the time domain, the time-frequency energy ridge disorder in the time-frequency domain, and the maximum singular value ratio and the maximum singular value in the singular value domain are extracted. Extracting different features from multiple domains is conducive to the feature decoupling of target information and can well distinguish target signals from sea clutter signals.

[0038] (2) The present invention has the characteristic of fast detection speed: compared with the traditional three-feature combined machine learning algorithm, since the feature only needs to extract data of a part of the distance unit, it saves feature extraction time and combines with CFAR detection to speed up the detection time, solving the technical problem of high time complexity.

[0039] (3) The present invention has the characteristic of a simple model: since the model does not use a machine learning algorithm, but uses a method combining CFAR and non-coherent accumulation, it is a very simple model.

[0040] In summary, compared with the existing technology, the present invention achieves accurate and rapid detection of small floating targets in sea clutter background by serially filtering five features: fluctuation stability and relative average amplitude in the time domain, time-frequency energy ridge disorder in the time-frequency domain, maximum singular value ratio and maximum singular value in the singular value domain, and CFAR detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a flow chart of the sea surface target detection method based on CFAR and five-feature combination of the present invention.

[0042] Figure 2It is a one-dimensional range image of the original radar echo data of the present invention.

[0043] Figure 3 It is the image obtained by non-coherent accumulation of the original radar echo data in step 1 of the present invention.

[0044] Figure 4 It is the image detected by OS-CFAR in step 2 of the present invention.

[0045] Figure 5 It is a characteristic diagram of the range unit fluctuation stability detected by OS-CFAR in step 3 of the present invention.

[0046] Figure 6 It is the maximum singular value ratio feature map of the remaining distance units in step 4 of the present invention.

[0047] Figure 7 It is the time-frequency energy ridge chaos characteristic diagram of the remaining distance unit in step 5 of the present invention.

[0048] Figure 8 It is the relative average amplitude characteristic diagram of the remaining distance units in step 6 of the present invention.

[0049] Figure 9 It is the distance unit map where the maximum singular value selected in step 7 of the present invention is located.

[0050] Figure 10 It is a one-dimensional distance image of the distance unit where the maximum singular value selected in step 7 of the present invention is located.

[0051] Figure 11 This is a comparison chart of the detection performance of the sea surface target detection method of the present invention and the traditional three-feature combined with machine learning algorithm.

[0052] Figure 12 This is a comparison chart of the detection time of the sea surface target detection method of the present invention and the traditional three-feature combined with machine learning algorithm. DETAILED DESCRIPTION

[0053] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is described in detail below with reference to the accompanying drawings.

[0054] A sea surface target detection method based on CFAR and five features is proposed. First, the original radar echo data is subjected to coarse detection, i.e., non-coherent accumulation is performed to improve the signal-to-noise ratio. Then, the distance units passing the detection threshold are extracted by combining OS-CFAR detection, thereby reducing the time for subsequent feature detection. Feature detection is then performed on the radar echo data after coarse detection, i.e., by analyzing the fluctuation degree of the target signal, the fluctuation stability feature is established, thereby realizing feature extraction in the time domain. By analyzing the singular values of the target signal, the maximum singular value ratio feature is established, thereby realizing feature extraction in the singular value domain. The time domain signal is transformed into the time-frequency domain by short-time Fourier transform. According to the different degrees of fluctuation of the time-frequency energy ridges of the target and sea clutter, the time-frequency energy ridge chaos degree feature is established to realize the extraction of features in the time-frequency domain; by analyzing the energy of the target signal, the relative average amplitude feature is established to realize the extraction of features in the time domain; by constructing the Hankel matrix of the radar echo data and performing singular value decomposition on the Hankel matrix, the maximum singular value feature is established according to the difference in the maximum singular value of the target and sea clutter signals, realizing the extraction of features in the singular value domain, and the serial filtering of these five features and CFAR detection realize the accurate and rapid detection of small floating targets under the background of sea clutter; the main steps include:

[0055] Step 1: Perform non-coherent accumulation on the original radar echo data. Due to the high energy of the target, the accumulated value is higher than that of the sea clutter unit.

[0056] The calculation formula is:

[0057]

[0058] Among them, x ij represents the discrete radar echo sequence, i.e. the original radar echo data, and N represents the number of range units;

[0059] The one-dimensional range image of the original radar echo data is as follows Figure 2 As shown in the figure, the radar echo data after non-coherent accumulation is as follows: Figure 3 As shown;

[0060] Step 2: Perform OS-CFAR detection on the radar echo data after non-coherent accumulation and extract the distance units that pass the detection threshold, such as Figure 4 As shown in the figure, OS-CFAR detection can distinguish some targets from sea clutter units, reducing the time for subsequent feature detection.

[0061] OS-CFAR detection is specifically as follows: after sorting the radar echo data after non-coherent accumulation, the kth value is taken as the background noise power value. The background noise power value is multiplied by the scale factor to obtain the detection threshold, which is then compared with the radar echo data to extract the distance unit greater than the detection threshold. The calculation formula is:

[0062] x1≤x2≤…≤x 2N

[0063]

[0064] T OS =x k α,k∈{1,2,…,2N}

[0065] Where N is the sliding window length, P fa represents the false alarm rate, α represents the scale factor, x k Indicates the clutter amplitude of the reference range unit, T OS Indicates the detection threshold;

[0066] Step 3: Calculate the fluctuation stability characteristic values of the distance units that pass the OS-CFAR detection threshold, sort them from large to small, and filter out the last half of the distance units, such as Figure 5 As shown; specifically:

[0067] 3.1) Calculate the fluctuation stability characteristic value W of each distance unit that passes the OS-CFAR detection threshold. The calculation formula is:

[0068] W=std(x)

[0069] Wherein, x represents the echo sequence of the distance unit to be measured;

[0070] 3.2) Arrange all the fluctuation stability eigenvalues W from largest to smallest and filter out the last half of the distance units;

[0071] Step 4: Calculate the maximum singular value ratio eigenvalue of the remaining distance units, arrange them in descending order, and filter out the last half of the distance units, such as Figure 6 Shown: Specifically:

[0072] 4.1) Rearrange the sample data of each remaining distance unit to construct a Hankel matrix H. Perform singular value decomposition on the Hankel matrix H to obtain a singular value diagonal matrix D. Then, extract the first five singular values in the singular value diagonal matrix D and arrange them in descending order.

[0073] The expression of the Hankel matrix H is:

[0074]

[0075] Perform singular value decomposition on the Hankel matrix H, and the calculation formula is:

[0076]

[0077] Extract the first five singular values from the singular value angle matrix D and arrange them in descending order. The calculation formula is:

[0078] D=diag(σ1,σ2,σ3,σ4,σ5),σ1≥σ2≥σ3≥σ4≥σ5

[0079] Where S[k] represents the kth sample value of the sequence (k = 1, 2, 3, ..., C ..., N), which is the permuted input data sequence; H represents the Hankel matrix, which is formed by rearranging the input signal, where N is the length of the input sequence, C is the number of columns (window width), and the matrix has a constant diagonal structure; U and V are the left singular vector matrix and the right singular vector matrix in the singular value decomposition (SVD), respectively; D represents the singular value diagonal matrix, and σ1, σ2, σ3, σ4, and σ5 represent the five singular values, respectively.

[0080] 4.2) Calculate the ratio of the first singular value to the sum of the second to fifth singular values to obtain the maximum singular value ratio eigenvalue Q. The calculation formula is:

[0081] Q=σ1 / sum(σ2,σ3,σ4,σ5)

[0082] 4.3) Arrange all the largest singular value proportion eigenvalues Q from large to small, and filter out the last half of the distance units;

[0083] Step 5: Calculate the time-frequency energy ridge disorder eigenvalues of the remaining distance units, arrange them in ascending order, and filter out the last half of the distance units, such as Figure 7 As shown; specifically:

[0084] 5.1) Perform short-time Fourier transform on the sample data of each remaining distance unit, and the calculation formula is:

[0085]

[0086] Among them, STFT(t,f) represents the time-frequency matrix, t represents the time axis sampling point, f represents the frequency axis sampling point, x(m) represents the input signal, h represents the window function, and m represents the sampling point in the time window. represents a complex exponential basis;

[0087] 5.2) After modulo-calculating the frequency dimension of the STFT(t,f) obtained by the short-time Fourier transform, perform maximum indexing to obtain the index vector v. The calculation formula is:

[0088] v = argmax{|STFT(t,f)|}

[0089] 5.3) Perform standard deviation calculation on the index vector to obtain the time-frequency energy ridge disorder eigenvalue O, which is calculated as follows:

[0090] O=std(v)

[0091] 5.4) Arrange all the time-frequency energy ridge chaos eigenvalues from small to large and filter out the last half of the distance units;

[0092] Step 6: Calculate the relative average amplitude eigenvalue (RAA) of the remaining distance units, sort them from large to small, and filter out the last half of the distance units, such as Figure 8 As shown; specifically:

[0093] 6.1) Calculate the relative average amplitude eigenvalue X of each remaining distance unit using the following formula:

[0094]

[0095] Among them, x j Represents the echo sequence of the distance unit to be measured, x ij represents a discrete radar echo sequence;

[0096] 6.2) Arrange all relative average amplitude eigenvalues X from largest to smallest and filter out the last half of the distance units;

[0097] Step 7: Calculate the maximum singular value feature of the remaining distance units, such as Figure 9 As shown; extract the distance unit where the largest singular value is located to complete the detection of sea surface targets, as shown Figure 10 As shown; specifically:

[0098] 7.1) Rearrange the sample data of each remaining distance unit to construct a Hankel matrix H. Perform singular value decomposition on the Hankel matrix H to obtain a singular value diagonal matrix D. The expression of the Hankel matrix H is:

[0099]

[0100] Perform singular value decomposition on the Hankel matrix H, and the calculation formula is:

[0101]

[0102] Where S[k] represents the kth sample value of the sequence (k = 1, 2, 3, ..., C ..., N), the permuted input data sequence; H represents the Hankel matrix, formed by rearranging the input signal, where N is the length of the input sequence, C is the number of columns (window width), and the matrix has a constant diagonal structure; U and V are the left and right singular vector matrices in the singular value decomposition (SVD), respectively; D represents the singular value diagonal matrix;

[0103] 7.2) Extract the first five singular values from the singular value angle matrix D and arrange them in descending order. The calculation formula is:

[0104] D=diag(σ1,σ2,σ3,σ4,σ5),σ1≥σ2≥σ3≥σ4≥σ5

[0105] Among them, σ1, σ2, σ3, σ4, and σ5 represent five singular values respectively.

[0106] Then extract the first singular value T in the singular value diagonal matrix, that is, the distance unit where the largest singular value is located, to complete the detection of sea surface targets. The calculation formula is:

[0107] T=σ1.

[0108] Based on OS-CFAR detection, this method sequentially extracts five features: the eigenvalue of fluctuation stability, the eigenvalue of the maximum singular value ratio, the eigenvalue of the time-frequency energy ridge disorder, the eigenvalue of the relative average amplitude, and the maximum singular value. This feature is then gradually filtered out of sea clutter units to achieve sea surface target detection. Verification results using the IPIX radar measured data set demonstrate that, compared to traditional three-feature combined machine learning algorithms, the proposed algorithm significantly reduces detection time while maintaining detection probability, effectively completing detection, as shown in Tables 1 and 2.

[0109] Table 1 Detection probability under different data lengths

[0110] Data length 128 512 1024 The present invention 0.8815 0.9056 0.9278 Original three characteristics 0.8925 0.8845 0.9387

[0111] Table 2 Detection time under different data lengths

[0112] Data length 128 512 1024 The present invention 5.12s 7.22s 10.74s Original three characteristics 8.45s 12.74s 20.82s

[0113] Table 1 shows the IPIX measured sea clutter data with different data lengths and the false alarm rate of 1e -3 , under HH polarization conditions, the detection probability of the present invention is compared with that of the traditional three-feature combined machine learning algorithm, and the above table is plotted into a line graph, as shown Figure 11 shown.

[0114] Table 2 shows the IPIX measured sea clutter data with different data lengths, with a false alarm rate of 1e -3 , under HH polarization conditions, the detection time of the present invention is compared with that of the traditional three-feature combined machine learning algorithm, and the above table is plotted into a line graph, as shown Figure 12 shown.

[0115] This shows that compared with the traditional three-feature combined machine learning algorithm, the target detection algorithm of the present invention significantly reduces the time required for detection while maintaining a similar detection probability. After multiple verifications using the IPIX radar measured sea clutter dataset, the present invention has high detection performance and strong robustness.

Claims

1. A sea surface target detection method based on CFAR and five features, characterized in that: The steps include: Step 1: Perform non-coherent accumulation on the original radar echo data; Step 2: Perform OS-CFAR detection on the radar echo data after non-coherent accumulation and extract the range cells that pass the detection threshold; Step 3: Calculate the fluctuation stability characteristic values of the distance cells that pass the OS-CFAR detection threshold, sort them from large to small, and filter out the last half of the distance cells; Step 4: Calculate the maximum singular value ratio eigenvalue of the remaining distance units, arrange them in descending order, and filter out the last half of the distance units; Step 5: Calculate the time-frequency energy ridge chaos eigenvalues of the remaining distance units, arrange them in ascending order, and filter out the last half of the distance units; Step 6: Calculate the relative average amplitude eigenvalues of the remaining distance units, sort them from large to small, and filter out the last half of the distance units; Step 7: Calculate the maximum singular value feature of the remaining distance units, extract the distance unit where the maximum singular value is located, and complete the detection of sea surface targets.

2. The method for sea surface target detection based on CFAR and five-feature combination according to claim 1, characterized in that: In step 3, the fluctuation stability characteristic values of the distance cells that pass the OS-CFAR detection threshold are calculated, and the latter half of the distance cells are filtered out in descending order, including: 3.1) Calculate the fluctuation stability characteristic value W of each distance unit that passes the OS-CFAR detection threshold. The calculation formula is: W=std(x) Wherein, x represents the echo sequence of the distance unit to be measured; 3.2) Arrange all the fluctuation stability eigenvalues W from large to small and filter out the last half of the distance units.

3. The sea surface target detection method based on CFAR and five-feature combination according to claim 1, characterized in that: In step 4, the maximum singular value ratio eigenvalues of the remaining distance units are calculated, and they are arranged in descending order, and the second half of the distance units are filtered out, including: 4.1) Rearrange the sample data of each remaining distance unit to construct a Hankel matrix. Perform singular value decomposition on the Hankel matrix to obtain a singular value diagonal matrix. Then, extract the first five singular values in the singular value diagonal matrix and arrange them in descending order. 4.2) Calculate the ratio of the first singular value to the sum of the second to fifth singular values to obtain the eigenvalue with the largest singular value ratio; 4.3) Arrange all the eigenvalues with the largest singular value proportion from large to small, and filter out the last half of the distance units.

4. The method for sea surface target detection based on CFAR and five-feature combination according to claim 1, characterized in that: In step 5, the time-frequency energy ridge chaos eigenvalues of the remaining distance units are calculated, arranged in ascending order, and the second half of the distance units are filtered out, including: 5.1) Perform short-time Fourier transform on the sample data of each remaining distance unit; 5.2) Modulo the frequency dimension of the sample data after the short-time Fourier transform and perform maximum indexing to obtain an index vector; 5.3) Perform standard deviation calculation on the index vector to obtain the time-frequency energy ridge disorder eigenvalue; 5.4) Arrange all the time-frequency energy ridge chaos eigenvalues from small to large and filter out the last half of the distance units.

5. The method for sea surface target detection based on CFAR and five-feature combination according to claim 1, characterized in that: In step 6, the relative average amplitude eigenvalues of the remaining distance units are calculated, sorted from largest to smallest, and the second half of the distance units are filtered out, including: 6.1) Calculate the relative average amplitude eigenvalue X of each remaining distance unit using the following formula: Among them, x j Represents the echo sequence of the distance unit to be measured, x ij represents a discrete radar echo sequence; 6.2) Arrange all the relative average amplitude eigenvalues X from large to small and filter out the last half of the distance units.

6. The method for sea surface target detection based on CFAR and five-feature combination according to claim 1, characterized in that: In step 7, the maximum singular value feature of the remaining distance units is calculated, and the distance unit where the maximum singular value is located is extracted, including: 7.1) Rearrange the sample data of each remaining distance unit to construct a Hankel matrix, perform singular value decomposition on the Hankel matrix, and obtain a singular value diagonal matrix; 7.2) Arrange the first five singular values in the singular value diagonal matrix in descending order, extract the first singular value in the diagonal matrix, that is, the distance unit where the largest singular value is located, and complete the detection of sea surface targets.

Images