A method for target detection based on sea clutter spatial distance dimension features
By training a sparse representation dictionary of the spatial distance dimension features of sea clutter, the problem of separating sea clutter from targets under high sea states was solved, and the target detection performance under complex sea states was improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUILIN UNIV OF ELECTRONIC TECH
- Filing Date
- 2023-03-22
- Publication Date
- 2026-07-21
AI Technical Summary
Under high sea state conditions, existing technologies struggle to effectively separate sea clutter and slow, weak targets, leading to the failure of target detection algorithms and an increased probability of false alarms for radar, especially under strong sea state conditions.
By training a sparse representation dictionary of the spatial distance dimension features of sea clutter, and utilizing the spatial domain feature differences of sea clutter, clutter components are reconstructed and suppressed, thereby improving the signal-to-clutter ratio. Target signals are then extracted using conventional target detection methods.
It can effectively separate sea clutter and target signals under high sea states, reduce false alarm probability, improve target detection performance, and adapt to different sea state environments.
Smart Images
Figure CN116068522B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for processing radar echo data and detecting signals of slow-moving and weak targets on the sea surface in a marine environment. Specifically, it is a method for extracting the spatial range dimension features of sea clutter in the detection of small and slow targets at sea, and achieving target detection by separating clutter from target signals. Background Technology
[0002] Detecting slow, weak targets in ocean clutter environments is an extremely challenging task, especially under high sea states where complex and strong clutter makes target-clutter separation very difficult. Currently, the most mature methods for marine target detection are adaptive class detection methods, mainly including generalized likelihood ratio detection, adaptive matched filters, and normalized adaptive matched filters. These methods require establishing complex amplitude statistical models of sea clutter to form detection statistics, often resulting in difficulties in parameter estimation and poor generalization ability. Furthermore, as radar resolution or sea state increases, sea clutter exhibits significant non-Gaussian, nonlinear, and non-stationary statistical characteristics. At this point, target-like spikes frequently appear, and the statistical model of sea clutter suffers from severe tailing, increasing the probability of false alarms for the radar.
[0003] Discrete Fourier transform bases and various time-frequency transform bases are often used to extract and separate the time-frequency features of signals; however, they are not ideal for distinguishing between sea clutter and slow-moving, weak targets on the sea surface. Sparse representation signal processing is a data-driven feature extraction approach. Its core idea is to express complex signal features by rationally designing a dictionary (sparse feature space). In marine target detection, this means achieving the separation of sea clutter and target signals in the sparse feature domain. Methods for constructing sparse representation dictionaries are divided into fixed dictionary construction methods and adaptive training methods. Fixed dictionary construction methods require prior knowledge of the signal features and select a basis function of a linear transform to construct the signal feature space, such as wavelet bases or short-time Fourier bases. Adaptive dictionary training methods, on the other hand, are data-driven. They utilize the sparsity of signal features to obtain a sparse representation dictionary, or sparse feature space, through training, which can better represent complex signal features.
[0004] In actual strong sea conditions, when the radial velocity of the target and the wave motion are similar, the target signal and clutter characteristics in the slow time dimension of the echo data on which the existing target detection methods are based are almost identical, causing the target detection algorithm to fail. Summary of the Invention
[0005] To address the issue that target signals and clutter share similar characteristics in the slow time dimension during actual sea conditions, leading to reduced effectiveness of clutter filtering algorithms based on slow time dimension features, this invention provides a target detection method based on the spatial range dimension features of sea clutter. By rearranging and training the dictionary of captured marine radar echo data, a sparse representation dictionary capable of expressing the spatial domain features of marine clutter is obtained. This allows for the reconstruction and elimination of clutter components in the radar echo based on the clutter feature dictionary. By utilizing the differences in the correlation between targets and clutter in the spatial domain, the signal-to-clutter ratio of small, slow-moving targets at sea is improved.
[0006] This invention discloses a method for target detection based on the spatial range dimension features of sea clutter, referring to... Figure 1 It includes the following steps:
[0007] (1) Training the sea clutter range dimension feature representation dictionary: In a target-free sea area, the radar-captured ocean echo data, i.e., pure clutter data, is rearranged to obtain the sea clutter range dimension training sample set Y. Train ;
[0008] (2) Using a dictionary learning algorithm, a training sample set Y for the sea clutter distance dimension was prepared. Train The eigenvectors of the data matrix are used for dictionary training to obtain the sparse feature space of the sea clutter distance dimension.
[0009] (3) The sea clutter range dimension sparse feature space is used to deconstruct the echo data to be measured. According to formula (5), the sparse representation coefficient vector is estimated according to the minimum Euclidean distance principle. That is, the echo sample y is represented by the dictionary D. i coefficient make Closest echo sample data y i The sea clutter data components obtained after reconstruction The reconstruction results of multiple samples are then used to construct C = [c1, ..., c2]. N ];
[0010] (4) For the measured echo data Y test , through Y test -C cancellation processing yields the sea clutter-suppressed signal E;
[0011] (5) Target extraction processing is performed on signal E. Conventional target detection methods are used, such as MTI (Moving Target Indication) and MTD (Moving Target Detection) based on fast and slow time-dimensional radar echo data, to detect the target.
[0012] Effectively extracting sea clutter features and fully reflecting the spatial correlation of sea clutter are the key to separating the target from the clutter signal in this invention.
[0013] To train the sea clutter range dimension feature representation dictionary, the echo data of the sea area under test without targets (i.e., pure clutter data) must first be rearranged to obtain the sea clutter range dimension training sample set. The specific method is as follows:
[0014] Reference Figure 2 For the range-pulse radar pure sea clutter echo complex data matrix under the sea state of the sea area to be measured, P represents the number of accumulated pulses, and R represents the number of range cells. The real vector y of the i-th sample is constructed from the in-phase or quadrature components of the pulse echo, grouped into a set of n adjacent pulse periods. i A total of N = P / n samples can be obtained (where P is an integer multiple of n, i = 1, ..., N). After extracting the real and imaginary parts of each sample, the length is L = 2nR, thus obtaining the sea clutter training sample set. It is a real sample matrix.
[0015] There are various methods for the specific process of dictionary training and learning, with the main difference being in the dictionary update stage. For example, the existing K-SVD algorithm (Dong Ziwei, Sun Jun, Sun Jingming, et al. Sparse dictionary learning for detecting weak moving targets on the sea surface [J]. Systems Engineering and Electronics Technology, 2020, 42(1): 30-36.) decomposes the error set of dictionary reconstruction clutter into SVD and updates the dictionary atoms by using the eigenvectors corresponding to the maximum singular values. The specific update steps of dictionary learning are as follows:
[0016] (1) For a given training sample set Y Train Randomly select m samples to use as the initial dictionary D0;
[0017] (2) In each iteration, the K-SVD algorithm uses the Orthogonal Matching Pursuit (OMP) algorithm in the sparse coding stage to implement the Y... Train Sparse solution;
[0018] (3) During the dictionary update phase, the dictionary atoms are updated one by one using a sequential update method:
[0019] In the T-th iteration, for the i-th dictionary atom that is to be updated, let p i The i-th atom d from the dictionary is used to solve the sparse representation. i The training sample index set, For index p i The corresponding sample set, excluding the i-th dictionary atom, the remaining dictionary atom pairs in the sample set. The reconstruction error is calculated as follows:
[0020]
[0021] In the formula, It is dictionary D deleting dictionary atom d i The dictionary matrix following (the i-th column of D) It is the original sparse representation coefficient matrix Delete the i-th row The result;
[0022] right Perform singular value decomposition:
[0023]
[0024] To make dictionary D correct To minimize the reconstruction error, update the i-th atom in the dictionary and its corresponding sparse representation coefficients as follows:
[0025]
[0026]
[0027] In the formula, U1 is the first column of the left singular matrix, and Δ1V1 is the product of the first singular value and the first column of the right singular matrix. and Used to update the dictionary matrix D and the coefficient matrix. The dictionary is updated atomically until convergence or the specified number of iterations is reached.
[0028] The sea clutter dictionary training in step (2) of the method of the present invention can be regarded as using a dictionary learning algorithm to train the sea clutter training sample set data matrix Y. Train The process of adaptive learning of feature vectors, and the dictionary training method is:
[0029] Suppose the dictionary contains a set of atomic vectors. ||d j ||2=1, where d j Let m be the number of atoms in the dictionary, j = 1, ..., m, and L be the dimension of the dictionary atom. Dictionary atoms can sparsely and linearly represent the sample vector y. i ,like Figure 3 As shown;
[0030] Dictionary learning for sea clutter can be modeled as an iterative process involving two stages: sparse coding and dictionary updating.
[0031] In the sparse coding stage, the dictionary D remains unchanged, and the vector of sparse representation coefficients X is solved.
[0032]
[0033] in, Let Y be the sparse representation of the sea clutter training sample set Y on the dictionary D, and the sparsity of each column vector is restricted to a range of size ε.
[0034] During the dictionary update phase, the sparse representation coefficients X remain unchanged, and the dictionary D is updated:
[0035]
[0036] The optimization problem is solved iteratively in two stages until the convergence threshold or the maximum number of iterations is reached. The learned dictionary D is then output, which yields the distance-dimensional feature space of the sea clutter.
[0037] This invention proposes a method for separating sea clutter from target echoes by leveraging differences in spatial range-dimensional correlation features. By reconstructing radar echo samples, the spatial range-dimensional correlation features of sea clutter are learned and captured, and a sparse representation dictionary is used to reflect the spatial range-dimensional correlation features of the sea clutter. This method is crucial for reconstructing and canceling sea clutter, thereby improving the signal-to-clutter ratio. By training a sparse representation dictionary to reflect the spatial range-dimensional correlation features of clutter under specific sea conditions, and using this dictionary to reconstruct and cancel sea clutter components in received echo data, the signal-to-clutter ratio is improved, and the target signal is resolved.
[0038] This invention utilizes the differences in the correlation between targets and clutter in the spatial domain to improve the signal-to-clutter ratio of small, slow-moving targets. It has the advantage of adapting to the non-Gaussian, nonlinear, and non-stationary statistical characteristics of sea clutter in high sea states, and can adapt to different sea states through adaptive dictionary learning without requiring prior knowledge of sea clutter statistical characteristics. Under complex clutter characteristics in high sea states, this invention's method exhibits better clutter suppression and target signal extraction performance than methods based on slow-time-dimensional sea clutter feature extraction. This invention is applicable to clutter suppression in radar detection and processing of slow-moving, weak marine targets. Attached Figure Description
[0039] Figure 1 This is a flowchart of the method of the present invention;
[0040] Figure 2 This is a schematic diagram of sample data rearrangement in the spatial distance dimension in the method of the present invention;
[0041] Figure 3 This is a schematic diagram of the sparse dictionary representation of the distance dimension vector of the echo signal in the method of the present invention. Detailed Implementation
[0042] The present invention will be further described in detail below with reference to the embodiments, but this is not intended to limit the present invention.
[0043] Example:
[0044] Ocean clutter data measured by the Naval Aviation University were used, with the datasets 20210106155330_01_staring.mat and 20210106155432_01_staring.mat (hereinafter referred to as 330_01 and 432_01, respectively) showing good consistency as data support to verify the method of this invention. This data was generated by a shore-based X-band solid-state fully coherent radar in staring mode. The test site was the Yantai No. 1 Bathing Beach test point, with the radar erected at a height of 80m and sea state reaching level 3-4. Its main performance parameters are shown in Table 1. In addition to the parameters given in the table, the range sampling interval can be calculated from the radar parameters in the table to be 2.5m, and the range resolution is 6m.
[0045] Table 1 Key performance parameters of the Hainan Airlines measured dataset
[0046]
[0047] Both 330_01 and 432_01 are pure sea clutter data. All pulse units of 330_01, and range units 1701-1892 are selected as training data. The last 192 pulse units of 432_01 are selected. To ensure consistency with the training set, range units 1701-1892, identical to those in the training set, are used as test data. The target is injected into range unit 40 of the pure sea clutter test data. The target amplitude is adjusted according to a preset signal-to-clutter ratio. The sea clutter power is determined by the average power of the pure sea clutter in the range unit of the target and its two adjacent range units. The target frequency f... d The target's radial velocity and radar parameters are jointly determined. The example is selected under the condition of a given signal-to-clutter ratio (SCR) of -15 dB and target data parameters with a velocity of 2 m / s.
[0048] Consider training and testing the dictionary on the same-direction and orthogonal component data of the training and test sets respectively, and generate the training sample set Y according to the sample construction method given in this invention. Train and test sample set Y test .
[0049] To prevent overfitting, Y is randomly selected. Train A number of samples were used for dictionary training; in this embodiment, 5000 training samples were selected. Regarding the parameter settings for the K-SVD algorithm, considering that the number of dictionary atoms m, the dimension of dictionary atoms L, and the sparse representation error affect the algorithm's complexity and underfitting / overfitting degree, to weigh their advantages and disadvantages, and based on experimental results, the optimal parameters for the dictionary learning algorithm were chosen as L = 192, m = 640 dictionary atoms, and error = 600, with 15 iterations.
[0050] After the dictionary training, sparse coding and reconstruction, clutter filtering and MTD target detection processes described above, the obtained range-Doppler spectrum can achieve excellent target detection performance under the condition of SCR = -15dB.
Claims
1. A method for target detection based on the spatial range dimension features of sea clutter, characterized in that, Includes the following steps: (1) Training the sea clutter range dimension feature representation dictionary: In a target-free sea area, the radar-captured ocean echo data, i.e., pure clutter data, is rearranged to obtain the sea clutter range dimension training sample set Y. Train ; (2) Using a dictionary learning algorithm, a training sample set Y for the sea clutter distance dimension was prepared. Train The eigenvectors of the data matrix are used for dictionary training to obtain the sparse feature space of the sea clutter distance dimension. (3) The sea clutter range dimension sparse feature space is used to deconstruct the echo data to be measured. According to formula (5), the sparse representation coefficient vector is estimated according to the minimum Euclidean distance principle. That is, the echo sample y is represented by the dictionary D. i coefficient make Closest echo sample data y i The sea clutter data components obtained after reconstruction The reconstruction results of multiple samples are then used to construct C = [c1, ..., c2]. N ]; (4) For the measured echo data Y test , through Y test -C cancellation processing yields the sea clutter-suppressed signal E; (5) Target extraction processing is performed on signal E, and the target is detected using conventional target detection methods; the pure clutter data described in step (1) is rearranged to obtain the sea clutter range dimension training sample set Y. Train The specific method is: For the range-pulse radar pure sea clutter echo complex data matrix under the sea state of the sea area to be measured P represents the number of accumulated pulses, and R represents the number of range cells. The real vector y of the i-th sample is constructed from the in-phase or quadrature components of the pulse echo, grouped into a set of n adjacent pulse periods. i A total of N = P / n samples can be obtained, where P is an integer multiple of n, i = 1, ..., N. After extracting the real and imaginary parts of each sample, the length is L = 2nR, thus obtaining the sea clutter training sample set. It is a real sample matrix; The sea clutter dictionary training described in step (2) involves using a dictionary learning algorithm to train the sea clutter training sample set Y. Train The process of adaptively learning the eigenvectors of a data matrix, using dictionary training methods, is as follows: Suppose the dictionary contains a set of atomic vectors. ||d j ||2=1, where d j Let m be the number of atoms in the dictionary, j = 1, ..., m, and L be the dimension of the dictionary atom. Dictionary atoms can sparsely and linearly represent the sample vector y. i ; Dictionary learning for sea clutter is modeled as an iterative process involving two stages: sparse coding and dictionary updating. In the sparse coding stage, the dictionary D remains unchanged, and the vector of sparse representation coefficients X is solved. in, Let Y be the sparse representation of the sea clutter training sample set Y on the dictionary D, and the sparsity of each column vector is restricted to a range of size ε. During the dictionary update phase, the sparse representation coefficients X remain unchanged, and the dictionary D is updated: The optimization problem is solved iteratively in two stages until the convergence threshold or the maximum number of iterations is reached. The learned dictionary D is then output, which yields the distance-dimensional feature space of the sea clutter.