An adaptive clustering method for radar echoes

Through the adaptive clustering method, frequency domain CFAR detection and feature similarity judgment are used to solve the clustering problem under the offshore target echo energy disconnection area in the radar system, and the target tracking performance and processing accuracy are improved.

CN118568525BActive Publication Date: 2024-11-12NAVAL AVIATION UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411016830.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-29
Publication Date
2024-11-12
Estimated Expiration
2044-07-29

AI Technical Summary

Technical Problem

When existing radar systems deal with sea target echoes, energy disconnection areas appear in strong scattering points due to improved distance resolution, and conventional point trace condensation methods are difficult to accurately cluster, resulting in the impact of target tracking performance.

Method used

Adaptive clustering method is adopted, including preliminary detection of frequency domain CFAR, point trace screening and point trace clustering, and area screening and feature similarity judgment, and time-frequency features assist in determining whether point traces belong to the same target , clustering is carried out based on the overlap of feature distribution intervals.

Benefits of technology

It improves the stability and reliability of point trace condensation, reduces the radar processing error rate, and effectively solves the target clustering problem in the energy disconnection area.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118568525B_ABST
    Figure CN118568525B_ABST
Patent Text Reader

Abstract

The present invention discloses an adaptive clustering method for radar echoes, which relates to the field of radar signal processing, and includes the following steps: step one: frequency domain CFAR preliminary detection; step two: point trace screening; step three: point trace clustering; the step one uses the existing frequency domain CFAR detection technology to analyze the radar echo to be detected, and obtains a binary detection result containing a series of point traces; the step two removes some point traces with a small area in the detection result. The present invention uses target characteristic mining information to assist in determining the category of radar point traces, so that the point trace aggregation result is more stable and reliable, and the overlap degree of the characteristic value distribution interval of two point traces is used as an indicator to measure the similarity of the point traces. On the basis of target detection, the similarity of target features is reasonably used to judge whether different strong scattering points belong to the same target, thereby effectively solving the problem of target clustering in the presence of energy disconnection areas.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of radar signal processing, and in particular to an adaptive clustering method for radar echoes. Background Art

[0002] As radar system parameters are gradually optimized, the target echo information that radar can receive is becoming more and more abundant, which is beneficial for correct detection at sea, but some new problems have also emerged. For example, when the radar range resolution is improved, the radar beam illuminates the target at sea (especially large targets, with the bow or stern facing the radar), and multiple strong scattering points will appear in the echo. These strong scattering points are at different distance units of the echo, while the target echo amplitude between the strong scattering points is not strong, that is, the echo belonging to the same target has an energy disconnection area. The following problems exist in the existing technology:

[0003] When using existing constant false alarm detection technology (such as frequency domain CFAR method) to process such radar echoes, the echoes belonging to the same target may appear multiple points in the detection results, which will bring additional calculations to the subsequent data processing and affect the target tracking performance. Therefore, it is necessary to perform condensation processing on the target points after constant false alarm processing, that is, clustering multiple detection points belonging to the same target to merge them into the same valid target. The schematic diagram of the conventional point condensation method is shown in the figure. Figure 5 As shown in the figure, in the azimuth-range dimension or distance-Doppler dimension, a detection window is formed with a certain unit as the center, and the energy of several units in the window is compared with the energy of the unit to be detected. If the unit to be detected is the maximum value, it is retained, otherwise it is discarded. If the target energy is continuous and there is no energy disconnection area, the conventional point trace condensation method with energy as the main information source can effectively solve the problem. However, for ship targets, in most cases, the energy of the bow or stern is much higher than that of the middle area of ​​the ship. At this time, when the conventional point trace condensation method is used for processing, since the strong scattering points at both ends of the target are separated by several or even dozens of weak scattering point distance units with lower energy, after processing, different strong scattering points of the same target are easily judged as two different targets, which in turn causes misleading to the subsequent processing of the radar (such as target tracking). From the perspective of scattering mechanism, the mutual coupling between the marine target and the environment is strong, and the target characteristics show strong fluctuation and time-varying. As one of the many characteristics, energy fluctuates violently with the change of attitude and observation angle, and the different geometric structures and materials of the hull have a great influence on the backscattered energy, so it is difficult to solve the above problems only by using energy information. Summary of the invention

[0004] The present application provides an adaptive clustering method for radar echoes to solve the problems raised in the above background technology.

[0005] To solve the above technical problems, the technical solution adopted in this application is as follows:

[0006] An adaptive clustering method for radar echoes, comprising the following steps:

[0007] Step 1: Preliminary detection in the frequency domain CFAR;

[0008] Step 2: Track screening;

[0009] Step 3: Track clustering;

[0010] In the above Step 1, the existing frequency domain CFAR detection technology is used to analyze the radar echoes to be detected, and a binary detection result containing a series of tracks is obtained;

[0011] In the above Step 2, in the detection result, the practical significance of small-area tracks is not prominent. Therefore, in this step, some tracks occupying a small area are screened out and eliminated;

[0012] In the above Step 3, it is assumed that there are remaining tracks after track screening. They are sorted according to the area size, and the set formed is denoted as , where the area of track 1 is the largest and the area of track is the smallest. Next, starting from the track with the largest area, first judge whether it belongs to the same class as the track with the second largest area. After that, judge with the track with the third largest area, and so on, to realize pairwise judgment between tracks. Auxiliary feature information is also needed for judgment.

[0013] A further improvement of the technical solution of this application is that in the radar echo map to be detected obtained in the above Step 1, the vertical axis represents the range cell dimension, the horizontal axis represents the number of pulses. In the obtained frequency domain CFAR detection result map, the vertical axis represents the range cell dimension, the horizontal axis represents the frame time dimension, and the black part represents the detected output tracks. Among them, one frame in the frame time dimension contains N pulses, and N represents the number of coherent pulses.

[0014] A further improvement of the technical solution of this application is that the screening method in the above Step 2 includes the following steps:

[0015] A1: Count the areas of all tracks, and here the area is replaced by the number of range-frame time units occupied by the track, to form a set ;

[0016] A2: Set a percentage p, 0 < p < 100%, and calculate the corresponding percentile in, and find the tracks with an area greater than this value;

[0017] A3: Eliminate the remaining tracks.

[0018] A further improvement of the technical solution of the present application is that: the method of using feature similarity in step 3 to determine whether the points K and M belong to the same type of target is as follows:

[0019] B1: Calculate the eigenvalues ​​of all frames in the trace K and obtain the distribution interval of the eigenvalues ​​of the trace K, recorded as ; Calculate the eigenvalues ​​of all frames in the trace M and obtain the distribution interval of the eigenvalues ​​of the trace M, recorded as ;

[0020] B2: According to , The overlap situation determines whether the points K and M belong to the same category. The criteria are as follows: , If there is no overlap, it is considered that points K and M do not belong to the same category; , There is overlap, and the overlapping interval is recorded as , calculate the overlap coefficient: , where represents the length of the interval, Represents the minimum value in the array. If Greater than similarity threshold , it is considered that points K and M belong to the same category, otherwise they are not the same category.

[0021] A further improvement of the technical solution of the present application is that the features include three different transformation domains:

[0022] C1: Relative average amplitude in the time domain, a signal of length N The RAA feature is calculated as follows: ;

[0023] C2: Get The method is as follows: First, you need to calculate The normalized smoothed pseudo-Wigner-Wiley distribution of is as follows:

[0024] 【1】;

[0025] In the formula, represent time delay and Doppler unit respectively, for The surrounding reference cell signals, represents the number of reference units, Respectively represent and The smoothed pseudo-Wigner-Wiley distribution of , Respectively The mean and variance sequence of It can be calculated by formula [2]:

[0026] 【2】;

[0027] In the formula, Represents the time-frequency ridge sequence in the time-frequency plane;

[0028] C3: To get , you need to first The bispectrum is defined as its third-order cumulant The 2D discrete Fourier transform of

[0029] ;

[0030] In the formula, represents the bispectral function, represents frequency, represents the conjugate transpose symbol, represents the expectation, and the diagonally integrated bispectrum can be calculated by the following formula:

[0031] ;

[0032] In the formula, the coordinate system From the original bispectral coordinate system Rotate 45º counterclockwise to get,

[0033] ;

[0034] In actual use, the above three features can be selected any one, or two or three can be used in combination.

[0035] Due to the adoption of the above technical solution, the present application has achieved the following technical progress compared with the prior art:

[0036] 1. The present application provides an adaptive clustering method for radar echoes, which assists in determining the category of radar traces by mining target characteristic information, making the trace aggregation result more stable and reliable, and taking the degree of overlap of the characteristic value distribution intervals of two traces as an indicator to measure the similarity of the traces. On the basis of target detection, the similarity of target features is reasonably utilized to judge whether different strong scattering points belong to the same target, thereby effectively solving the target clustering problem in the presence of energy disconnection areas.

[0037] 2. The present application provides an adaptive clustering method for radar echoes, in which the selection of features is open and flexible, and can be pre-stored in the software and flexibly called during operation to reduce the error rate of radar processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1It is a schematic diagram of the time domain display of the radar echo and the frequency domain CFAR detection result of this application;

[0039] Figure 2 This is a schematic diagram of the point trace screening results of this application;

[0040] Figure 3 It is a schematic diagram of the process of point trace condensation of this application;

[0041] Figure 4 Schematic diagram of initial clustering and adaptive clustering results of point traces in this application;

[0042] Figure 5 This is a schematic diagram of the conventional point trace aggregation method of this application. DETAILED DESCRIPTION

[0043] The present application is further described in detail below with reference to the embodiments:

[0044] Example 1

[0045] like Figure 1-Figure 5 As shown, the present application provides an adaptive clustering method for radar echoes, comprising the following steps:

[0046] Step 1: Preliminary detection of CFAR in frequency domain;

[0047] Step 2: Point screening;

[0048] Step 3: point clustering;

[0049] The step 1 uses the existing frequency domain CFAR detection technology to analyze the radar echo to be detected to obtain a binary detection result including a series of point traces;

[0050] In the detection results of step 2, the practical significance of analyzing small-area dots is not prominent, so this step will screen out some dots with a smaller area;

[0051] In step 3, it is assumed that after the point traces are screened, there are a total of points, sort them according to their area size, and the resulting set is denoted by , among which, point 1 has the largest area, and point The point with the smallest area, then start with the point with the largest area, first determine whether it belongs to the same category as the point with the second largest area, then determine whether it belongs to the same category as the point with the third largest area, and repeat this cycle to achieve pairwise determination between points, with the help of auxiliary feature information.

[0052] Example 2

[0053] like Figure 1-Figure 5As shown, on the basis of Embodiment 1, the present application provides a technical solution: Preferably, in the radar echo map to be detected obtained in Step 1, the vertical axis represents the range cell dimension, and the horizontal axis represents the number of pulses. In the obtained frequency domain CFAR detection result map, the vertical axis represents the range cell dimension, and the horizontal axis represents the frame time dimension. The black part represents the detected output traces. Among them, one frame in the frame time dimension contains N pulses, and N represents the number of coherent pulses. The screening method in Step 2 includes the following steps:

[0054] A1: Statistically calculate the area of all traces, and here the area is replaced by the number of range - frame time units occupied by the traces, to form a set ;

[0055] A2: Set a percentage p, 0 < p < 100%, and calculate the corresponding percentile in, and find the traces with an area greater than this value;

[0056] A3: Eliminate the remaining traces.

[0057] Embodiment 3

[0058] As Figure 1-Figure 5 shown, on the basis of Embodiment 1, the present application provides a technical solution: Preferably, the method for using feature similarity to determine whether traces K and M belong to the same type of target in Step 3 is as follows:

[0059] B1: Calculate the eigenvalues of all frames within trace K to obtain the distribution interval of the eigenvalues of trace K, denoted as ; Calculate the eigenvalues of all frames within trace M to obtain the distribution interval of the eigenvalues of trace M, denoted as ;

[0060] B2: Judge whether traces K and M belong to the same class according to the overlap situation of , . The criterion is as follows: If , have no overlap, it is considered that traces K and M do not belong to the same class; if , have overlap, the overlapping interval is denoted as , and calculate the overlap coefficient: , where in the formula, represents the length of the interval, represents taking the minimum value in the array. If is greater than the similarity threshold , it is considered that traces K and M belong to the same class, otherwise they do not belong to the same class. The features include three different transform domains:

[0061] C1: The relative average amplitude in the time domain, a signal of length N The RAA feature is calculated as follows: ;

[0062] C2: Get The method is as follows: First, you need to calculate The normalized smoothed pseudo-Wigner-Wiley distribution of is as follows:

[0063] 【1】;

[0064] In the formula, represent time delay and Doppler unit respectively, for The surrounding reference cell signals, represents the number of reference units, Respectively represent and The smoothed pseudo-Wigner-Wiley distribution of , Respectively The mean and variance sequence of It can be calculated by formula [2]:

[0065] 【2】;

[0066] In the formula, Represents the time-frequency ridge sequence in the time-frequency plane;

[0067] C3: To get , you need to first The bispectrum is defined as its third-order cumulant The 2D discrete Fourier transform of

[0068] ;

[0069] In the formula, represents the bispectral function, represents frequency, represents the conjugate transpose symbol, represents the expectation, and the diagonally integrated bispectrum can be calculated by the following formula:

[0070] ;

[0071] In the formula, the coordinate system From the original bispectral coordinate system Rotate 45º counterclockwise to get,

[0072] ;

[0073] In actual use, the above three features can be selected any one, or two or three can be used in combination.

[0074] Example 4

[0075] like Figure 4 As shown, based on Example 1, the present application provides a technical solution: preferably, Figure 4 middle , , , , when the feature is FC, the initial clustering results and adaptive clustering results of the points. The numbers 1~6 marked in the figure represent the target category numbers after aggregation. There is a ship target in the scene corresponding to the data, which occupies a distance unit range of 30~60. The radar beam is in staring observation mode. After CFAR detection, the left figure corresponds to the result obtained by the conventional clustering method. It can be seen that due to the disconnection area of ​​the target energy, the target cannot be aggregated into the same category after clustering. A total of 5 types of targets are judged within the distance unit range where the target appears; the right figure corresponds to the adaptive clustering result given by this patent. Due to the auxiliary judgment based on the similarity of the target feature information, the target clustering problem can still be effectively solved in the presence of energy disconnection areas.

[0076] The following is a detailed description of the working principle of the adaptive clustering method of radar echoes.

[0077] like Figure 1-Figure 5 As shown in the figure, first, the frequency domain CFAR detection method is used to process the radar echo to obtain a binary detection result containing a series of points. Then, some points with smaller areas are screened out. Subsequently, information that can be used to assist in determining the category of points is mined based on the target characteristics. The auxiliary information here includes two categories: one category is definitely not information of the same category of points. The method is: first set the maximum number of distance units that a sea surface target can occupy. If the number of units occupied by the centroid positions of two points in the distance dimension is greater than this value, it is considered that the two points are not of the same category; the other category is definitely information of the same category of points. If there are multiple points in a certain distance unit, and the sum of the number of frames occupied by these points in the frame time dimension corresponding to the distance unit is greater than the total number of frames multiplied by a preset percentage, it is recorded as , it is considered that these points belong to the same category. Then, with the help of auxiliary feature information, it is judged whether the radar points belong to the same category. For the points that cannot be judged, the degree of overlap of the characteristic value distribution intervals in the points is used for judgment. The higher the degree of overlap, the stronger the similarity of the two points is, and the more likely they are to be in the same category.

[0078] The above generally describes the present application in detail, but it is obvious to a person skilled in the art that some modifications or improvements can be made to the present application. Therefore, modifications or improvements that do not depart from the spirit of the present invention are within the scope of protection of the present application.

Claims

1. An adaptive clustering method for radar echoes, characterized in that: It includes the following steps: Step 1: Preliminary frequency-domain CFAR detection; Step 2: Plot screening; Step 3: Plot clustering; In Step 1, the existing frequency-domain CFAR detection technology is used to analyze the radar echo to be detected, and a binary detection result containing a series of plots is obtained; The screening method in Step 2 includes the following steps: A1: Statistically calculate the area of all plots, and here the area is replaced by the number of range-frame time units occupied by the plot to form a set Area; A2: Set a percentage p, 0 < p < 100%, calculate the corresponding percentile in Area, and find the plots with an area greater than p; A3: Eliminate the remaining plots; In Step 3, it is assumed that there are I plots remaining after plot screening. They are sorted according to the area size, and the formed set is denoted as {plot 1, plot 2,..., plot I}, where plot 1 has the largest area and plot I has the smallest area. Next, starting from the plot with the largest area, first judge whether it belongs to the same class as the plot with the second-largest area. After completion, judge with the plot with the third-largest area, and so on to achieve pairwise judgment between plots. During this process, auxiliary feature information is also needed for judgment. Information that can be used to assist in judging the plot category is mined according to the target characteristics. Here, the auxiliary information includes two categories: one is information that must not be plots of the same class. The method is: first set the maximum number of range units that a sea target can occupy. If the number of units occupied by the centroid positions of two plots in the range dimension is greater than this value, it is considered that these two plots are not of the same category; the other is information that must be plots of the same class. If there are multiple plots in a certain range unit, and the sum of the number of frames occupied by these plots in the frame time dimension corresponding to this range unit is greater than the total number of frames multiplied by a pre-set percentage, denoted as Q, then it is considered that these plots belong to the same class. Then, with the help of auxiliary feature information, pairwise judgment is made on whether the radar plots belong to the same class. For plots that still cannot be judged, judgment is made according to the overlapping degree of the eigenvalue distribution intervals within the plots. The higher the overlapping degree, the stronger the similarity between the two plots is considered; The method for using feature similarity to judge whether plots K and M belong to the same class target in Step 3 is as follows: B1: Calculate the eigenvalues ​​of all frames in the trace K and obtain the distribution interval of the eigenvalues ​​of the trace K, denoted as L K ; Calculate the eigenvalues ​​of all frames in the trace M and obtain the distribution interval of the eigenvalues ​​of the trace M, denoted as L M ; B2: According to L K , L M The overlap situation determines whether the points K and M belong to the same category. The criteria are as follows: K , L M There is no overlap, and it is considered that points K and M do not belong to the same category; if L K , L M There is overlap, and the overlapping interval is recorded as L C , calculate the overlap coefficient: cc = length (L c ) / min{length(L K ),length(L M )}, where length(·) represents the length of the interval, min(·) represents the minimum value in the array, and if cc is greater than the similarity threshold T, it is considered that the traces K and M belong to the same class, otherwise they are not of the same class.

2. The adaptive clustering method of radar echo according to claim 1, characterized in that: In the radar echo map to be detected obtained in Step 1, the vertical axis represents the range unit dimension, and the horizontal axis represents the number of pulses. In the frequency-domain CFAR detection result map obtained, the vertical axis represents the range unit dimension, and the horizontal axis represents the frame time dimension. The black part represents the detected plots. Among them, one frame in the frame time dimension contains N pulses, and N represents the number of coherent pulses.

3. The adaptive clustering method of radar echo according to claim 1, characterized in that: The features include three different transform domains: C1: Relative average amplitude in the time domain. The calculation method of the RAA feature of a signal x(n) with a length of N is as follows: C2: The method for obtaining RI(x) is as follows: First, it is necessary to calculate the normalized smoothed pseudo Wigner-Ville distribution of x(n), as shown in Equation [1] Where n and l represent the time delay and Doppler unit respectively, and x e is the reference unit signal around x(n), E represents the number of reference units, SPWVD(n,l|x), SPWVD(n,l|x e ) represent x and x respectively. e The smoothed pseudo-Wigner-Wiley distribution of They represent SPWVD(n,l|x e ), then RI(x) can be calculated by formula [2]: In the formula, Ridge(n|x) represents the time-frequency ridge sequence in the time-frequency plane; C3: To obtain FC(x), it is necessary to first calculate the bispectrum of x(n). The bispectrum is defined as the two-dimensional discrete Fourier transform of its third-order cumulant c3(τ1, τ2), that is c3(τ1,τ2)=E[x H (t)x(t+τ1)x(t+τ2)] Where B(ω1,ω2) represents the bispectral function, ω1 and ω2 both represent frequencies, H represents the conjugate transpose symbol, and E(·) represents the expectation. The diagonally integrated bispectrum can be calculated by the following formula: DIB(ω)=DIB(ω′1)=∫|B(ω′1,ω′2)|dω′2 In the formula, the coordinate system (ω1′, ω2′) is obtained by rotating the original bispectral coordinate system (ω1, ω2) counterclockwise by 45°. FC(x)=∫ω·DIB(ω)dω / ∫DIB(ω)dω; In actual use, the above three features can be selected any one, or two or three can be used in combination.

Citation Information

Cited By

  • Radar echo curve fitting correction method and system based on normal regression and weighted augmented matrix disturbance

    CN122654463A