A floating target recognition method using average amplitude feature time sequence information

Abstract

The present application relates to radar signal processing and pattern recognition cross technical field, mainly related to a kind of identification method for sea surface floating target.The present application utilizes the time domain variation law information of target radar echo to carry out floating target identification, improves the utilization rate of echo information in target identification process;The present application solves the problem that low-resolution radar cannot complete target identification, solves the problem that one-dimensional range image method cannot accurately identify when target physical size and RCS are similar;The identification accuracy of the method proposed in the present application for ship target and floating target and other mass difference larger target can reach more than 90% on the observation time of second level.

Description

TECHNICAL FIELD

[0001] The present application relates to the radar signal processing and pattern recognition cross technical field, and mainly relates to a kind of identification method for sea surface floating target. BACKGROUND

[0002] Radar target recognition technology is one of important function technologies of radar, and the accurate and rapid identification of sea surface target helps to determine the function purpose and threat level of target, so that the corresponding measures can be taken, which has important significance in civil and military fields.Common sea surface floating targets mainly include small fishing boats, buoys, ice floes, etc., which are generally different from large cruise ships, cargo ships, cargo roll ships, etc., in structure and physical size, so the floating targets and large ship targets usually have different instantaneous amplitude scattering characteristics, and therefore in high-resolution system radar, one-dimensional range image features are currently used to distinguish between the two.One-dimensional range image feature reflects the distribution of strong scattering points along the distance, and is an important target characteristic, which is the main basis for identifying floating targets and large ships.However, one-dimensional range image feature is closely related to the observation angle, especially when the radar beam is illuminated along the ship's side, the ship target and the floating target have similar one-dimensional range image characteristics, which seriously affects the identification performance.In addition, for low-resolution system radar, it is difficult to identify floating targets without relying on auxiliary information due to the lack of one-dimensional range image features.

[0003] Starting from the physical mechanism of floating targets and ship targets, the difference between the two is deeply explored, which is the fundamental way to solve the problems of the above-mentioned one-dimensional range image method and to realize accurate and stable identification of floating targets and ship targets.Although the ship target and the floating target may have similar one-dimensional range images, the mass of the two is quite different.In general, the density of the ship target is larger, and the density of the floating target is smaller, so under normal sea conditions, long-time observation can find that the ship target is usually in a relatively stable state and rarely fluctuates with the sea waves, while the floating target often fluctuates with the sea waves.

[0004] Therefore, the radar echoes of the two should have obviously different time sequence variation characteristics, so the present application intends to use the time domain variation rule information of target radar echo to distinguish between the two. SUMMARY

[0005] The purpose of the present application is to provide a floating target recognition method using average amplitude feature time sequence information, especially for low-resolution radar that cannot obtain one-dimensional range image, or sea surface floating targets and ship targets with similar physical size, which cannot be distinguished by one-dimensional range image.

[0006] The floating target recognition method using average amplitude feature time sequence information of the present application is characterized in that it comprises the following steps:

[0007] Step 1, extraction of average amplitude feature sequence

[0008] The sea detection radar detects the sea surface target. A target distance unit receives a coherent pulse train x(n), n = 1, 2, …, N, which is cut into I non-overlapping short vectors u i , as shown below:

[0009] [x(1), x(1), …, x(N)] T = [u1, u2, …, u NI ] T (1)

[0010] The average amplitude feature of each short vector u i is calculated respectively. The average amplitude feature used in the present application is the average of complex echoes, and the calculation formula is as follows:

[0011]

[0012] In the formula, z(n) is an echo pulse (complex number) with a length of N, and the above average is essentially the average of complex numbers. The average amplitude feature sequence of a certain sea surface target is denoted as X1, X2, …, X N / I .

[0013] Step 2, AR modeling of average amplitude feature sequence and secondary feature extraction

[0014] A centralized p-order auto-regressive (AR) model is used to fit the average amplitude feature sequence of all targets. The fitting result of the centralized AR(p) model of the average amplitude feature sequence X1, X2, …, X N / I is as follows:

[0015] X t = c1X t-1 + … + c p X t-p + ε t (3)

[0016] In the formula, the model order p is a known quantity, the model parameters c = (c1, c2, …, c p ) T are to be estimated parameters, and the matrix estimation value of the parameters can be calculated by formula (4).

[0017]

[0018] In the formula, is the estimation value of the autocorrelation coefficient, and formula (4) is called the Yule-Walker equation estimation method of parameters. The white noise variance the mean amplitude feature sequence As follows:

[0019]

[0020] In the formula, is the variance of the mean amplitude feature sequence, is the autocovariance function of interval j.

[0021] The formula (3) is the AR model fitting result of the mean amplitude feature sequence of a sea surface target, and the order p of the AR model is generally set to 3, so that the estimated AR model coefficient is denoted as AR(1), AR(2), and AR(3), since the mean amplitude feature is a complex number, the corresponding AR(1), AR(2), and AR(3) are also complex numbers, and the present application takes |AR(1)|, |AR(2)|, and |AR(3)| as the secondary features corresponding to the mean amplitude feature sequence.

[0022] Step 3, construction and identification of the classifier

[0023] In the classifier training stage, the secondary features |AR(1)|, |AR(2)|, and |AR(3)| of a large number of mean amplitude features are extracted from the labeled ship target echo, to constitute the required training set feature samples, and the standard deviation normalization (Standard Deviation Normalization) processing is performed; a preset error decision probability of the ship target is used, and the fast convex hull learning algorithm is trained to obtain a decision region, so that the classifier training is completed, and the target echo sample data [|AR(1)|, |AR(2)|, |AR(3)|] collected by the radar is input into the convex hull single classifier which has been trained, to identify the target category.

[0024] Compared with the prior art, the sea surface floating target recognition method using the mean amplitude feature time sequence information has the beneficial effects that:

[0025] (1) The method uses the time domain variation law information of the target radar echo to recognize the floating target, and improves the utilization rate of the echo information in the target recognition process.

[0026] (2) The method solves the problem that the low-resolution radar cannot complete target recognition, and solves the problem that the one-dimensional range image method cannot accurately recognize the target when the physical size and RCS of the target are similar.

[0027] (3) On the observation time of seconds, the recognition accuracy of the method for the targets with large mass difference such as ship targets and floating targets can reach more than 90%. BRIEF DESCRIPTION OF DRAWINGS

[0028] Appendix Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0029] Appendix Figure 2 The distribution of ship targets and floating targets in the AR model coefficient domain of the average amplitude feature sequence proposed in this invention;

[0030] Appendix Figure 3 This is a schematic diagram of the training and decision-making process for a convex hull classifier. Detailed Implementation

[0031] To better understand and implement this invention, specific embodiments are provided below with reference to the accompanying drawings to illustrate a floating target identification method utilizing average amplitude feature time-series information.

[0032] The specific implementation process of this target recognition method is attached. Figure 1 As shown in the accompanying drawings, the processing flow of the present invention will be described in detail below:

[0033] 1) Extraction of average amplitude feature sequence

[0034] Suppose a maritime radar detects a target on the sea surface. A coherent pulse train x(n), n=1,2,…,N is received at a target range cell. x(n), n=1,2,…,N is truncated into a short, non-overlapping vector u of length I according to equation (1). i .

[0035] Calculate each short vector u according to equation (2). i The average amplitude characteristics are used to obtain the average amplitude characteristic sequence of a certain sea surface target, denoted as X1, X2, ..., X. N / I From a physical mechanism perspective, it can be seen that since ship targets and floating targets have different temporal fluctuation characteristics, their average amplitude characteristic sequences should have significantly different change patterns. Therefore, the temporal information of the average amplitude characteristic sequence can be used to distinguish between the two.

[0036] 2) AR modeling and secondary feature extraction of average amplitude feature sequences

[0037] The average amplitude feature sequence X1, X2, ..., X of a certain sea surface target is analyzed using a centered AR(p) model. N / I The fitting results are shown in Equation (3), and the model parameters are calculated using the Yul-Walker equation estimation method introduced in Equation (4).

[0038] The order p of the AR model used is generally set to 3, then the estimated AR model coefficients are... They can be denoted as AR(1), AR(2), and AR(3). Since the average amplitude feature of this invention is a complex number, its corresponding AR(1), AR(2), and AR(3) are also complex numbers. This invention uses |AR(1)|, |AR(2)|, and |AR(3)| as secondary features corresponding to the average amplitude feature sequence.

[0039] The typical distributions of ship targets and floating targets (buoys) obtained from measured data in the AR model coefficient domain of the average amplitude feature sequence are shown in the appendix. Figure 2 As shown. Obviously, ship targets and floating targets are well separable in the three-dimensional feature space composed of |AR(1)|, |AR(2)|, and |AR(3)|. Therefore, the AR modeling coefficients of the average amplitude feature sequence |AR(1)|, |AR(2)|, and |AR(3)| are used as secondary features to distinguish ship targets and floating targets.

[0040] 3) Classifier construction and recognition

[0041] After extracting the secondary features |AR(1)|, |AR(2)|, and |AR(3)| corresponding to the average amplitude feature sequences of different target echoes, the identification problem becomes a classification problem in a 3D feature space. This invention places this problem within an anomaly detection framework and designs a single-class classifier using the convex hull learning algorithm to achieve target recognition.

[0042] A single-class classifier can be trained using only samples from one class of data, that is, based solely on the set of labeled ship target echo feature vectors S, it can be trained in a 3D feature space according to a preset ship target error probability P. f By progressively training, the convex hull of the ship target echo is obtained, effectively solving the problem of imbalance between the two types of samples. First, according to the Minimal-Volume Criterion, the standard deviation of the set S is normalized to avoid the excessive variation range of one feature value affecting the recognition of other features. Assume the training set is represented as:

[0043] S0=[|AR(1)|0,|AR(2)|0,|AR(3)|0] (6)

[0044] In the formula, |AR(1)|0, |AR(2)|0, and |AR(3)|0 are column vectors composed of a large number of secondary features |AR(1)|, |AR(2)|, and |AR(3)| values ​​of the average amplitude feature sequence of the ship target unit, respectively, and the length of the vector is denoted as Q. The standard deviation of each feature component can be estimated by the following formula:

[0045]

[0046] In the formula, |AR(1)|q ,|AR(2)| q ,|AR(3)| q Let |AR(1)|0, |AR(2)|0, and |AR(3)|0 represent the q-th eigenvalues ​​of the eigenvectors |AR(1)|0, |AR(2)|0, and |AR(3)|0, respectively, and mean(·) represent the mean operator. Then the normalized sample set can be represented as:

[0047]

[0048] In the formula, |AR(1)|, |AR(2)|, and |AR(3)| represent column vectors composed of the corresponding normalized features.

[0049] Clearly, ship target samples and floating target samples should cluster in different regions within the 3D feature space. Therefore, during training, feature points within the convex hull formed by ship target samples that are closer to the clustering region of floating target samples are more likely to become outliers and be removed. Thus, the process of forming the classifier's decision region is as follows:

[0050] 1) Initialization: Let W be the number of ship target echo feature vectors, and calculate the number of ship target anomalies F. num =W·P f , where P f Let l = 0, representing the preset probability of incorrect ship target identification.

[0051] 2) Find the maximum values ​​of the eigenvalues ​​|AR(1)|, |AR(2)|, and |AR(3)| in set S to form a new spatial vertex v0 = [max(|AR(1)|), max(|AR(2)|), max(|AR(3)|)].

[0052] 3) Generate the convex hull CH(S) of the current data point, with vertices {v1, v2, ..., v...} r}. Count the number of feature points falling into the convex hull CH(S), let it be n. all .

[0053] 4) Calculate the Euclidean distance from all feature points to vertex v0, and find the feature point v that is farthest from v0. i And remove;

[0054] 5) Generate a new convex hull CH(S-{v i Then calculate the number of feature points in the new convex hull, let it be n. q .

[0055] 6) Let S-{v i}=S,l+n all -n q =l.

[0056] 7) If l <Fnum If the process ends, return to step 2) and continue removing the next vertex. Otherwise, terminate the removal process and output the final decision region Ω = CH(S).

[0057] After the convex hull single classifier is trained, the target echo samples [|AR(1)|,|AR(2)|,|AR(3)|] are input to complete the recognition. The recognition rules are as follows:

[0058]

[0059] The diagrams illustrating the training and decision processes of the convex hull classifier obtained from the measured data are attached. Figure 3 As shown, the average amplitude feature of the ship target echo, a secondary feature, is used as a sample to train a convex hull single classifier, with a preset ship target misclassification probability P. f The accuracy rate was 6%, and the correct identification probability for floating targets (buoys) was 93.72%.

Claims

1. A method for identifying floating targets using time-series information of average amplitude characteristics, characterized in that... Includes the following steps: Step 1, Extraction of the average amplitude feature sequence The maritime radar detects targets on the sea surface. A target range cell receives a signal with a length of... N coherent pulse train ,Will Cut to length of I Non-overlapping short vectors As shown below: （1）； Calculate each short vector separately The average amplitude characteristic, which is the average of the complex echoes, is calculated using the following formula: （2）； In the formula, It is a length of The average of the echo pulses, which is essentially a complex number average, yields the average amplitude characteristic sequence of a certain sea surface target, denoted as... ; Step 2, AR modeling and secondary feature extraction of the average amplitude feature sequence: Use centralized p The autoregressive model fits the average amplitude feature sequence of all targets, and the average amplitude feature sequence... The fitting results of the centered AR(p) model are as follows: （3）； In the formula, the model order is... For known quantities, model parameters For the parameter to be estimated, its moment estimate is... It can be calculated using equation (4): （4）； In the formula, The autocorrelation coefficient is estimated using equation (4), which is called the Yul-Walker equation estimation method for the parameter. The white noise variance... The moment estimate as follows: （5）； In the formula, The variance of the average amplitude feature sequence, For interval The autocovariance function; Equation (3) is the AR model fitting result of the average amplitude feature sequence of a certain sea surface target, and the AR model order is... Generally, it is set to 3, then the estimated AR model coefficients are... Let them be denoted as AR(1), AR(2), and AR(3). Since the average amplitude characteristic is complex, their corresponding AR(1), AR(2), and AR(3) are also complex. , , As a secondary feature corresponding to the average amplitude feature sequence; Step 3, Classifier Construction and Recognition: During the classifier training phase, a large number of secondary features of average amplitude are extracted from the labeled ship target echoes. , , The required training set feature samples are constructed and their standard deviation is normalized. A pre-defined error probability for a ship target is established, and a fast convex hull learning algorithm is used for training to obtain the decision region, thus completing the classifier training. The target echo sample data acquired by the radar are then used for this purpose. Input a pre-trained convex hull single classifier to identify the target category.

Images