Radar target tracking and identification method based on multi-dimensional information fusion

Abstract

The invention relates to the field of signal processing, in particular to a radar target tracking and recognition method based on multi-dimensional information fusion, which comprises the following steps: performing pulse compression and coherent processing on a sea surface echo signal received by a radar, constructing a distance-Doppler graph, and detecting and positioning a suspected target through a constant false alarm rate. And the radial speed characteristic and the distance dimension contour fitting coefficient based on the weighted least square method are synchronously calculated. Afterwards, the motion and structure features are fused with a traditional energy state, an extended target state with a unique identification tag is constructed, and a probability distribution model of an initial multi-target state is established by adopting a tag multi-Bernoulli random finite set theory; and performing multi-target state estimation under a Bayesian recursion framework by using a generalized label multi-Bernoulli filter. Time domain statistical features are extracted based on a tracking trajectory, a vector machine is adopted to discriminate a target type, and through multi-dimensional information fusion and joint optimization processing, tracking stability and recognition reliability of a marine weak target are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of signal processing, in particular to a radar target tracking and identification method based on multi-dimensional information fusion. BACKGROUND

[0002] Maritime target detection and identification is of great significance in the fields of military defense, maritime safety, and marine rescue. Radar, as the main long-range detection sensor, is widely used in sea detection due to its characteristics of not being affected by day and night and bad weather. Radar discovers and identifies targets by receiving the echo signals reflected by targets, and the Range-Doppler graph formed thereby is the basis for target detection and feature analysis.

[0003] Radar is currently the core means of sea monitoring, but its detection and identification efficiency of weak small targets (such as small boats and floating objects) in complex sea conditions is severely restricted by sea clutter. Sea clutter appears as strong background noise in the radar image, and its amplitude and characteristics are highly similar to weak small targets, resulting in a very high false alarm rate for traditional threshold-based (such as CFAR) detection methods, and easy to cause missed detection.

[0004] Existing multi-target tracking algorithms mainly rely on the motion trajectory of the target for association and filtering, and fail to fully exploit and utilize the fine structure features of the target in the range dimension. Therefore, in complex scenes where targets and clutter, or different categories of targets are densely intertwined, it is difficult to achieve stable tracking and effective differentiation. SUMMARY

[0005] The present application aims to provide a radar target tracking and identification method based on multi-dimensional information fusion, which solves the problem of stable tracking and effective differentiation in the complex scene where targets and clutter, or different categories of targets are densely intertwined in the prior art.

[0006] The present application is implemented by the following technical solutions:

[0007] In a first aspect, the present application embodiment provides a radar target tracking and identification method based on multi-dimensional information fusion, comprising:

[0008] constructing a range-Doppler graph according to the echo signal obtained at the initial moment, the echo signal being the sea echo received by the radar after transmitting a pulse to the sea;

[0009] performing suspected target detection and feature calculation according to the range-Doppler graph, to obtain the index coordinates, Doppler velocity, and local polynomial fitting contour curve of each suspected target in the range-Doppler graph;

[0010] Multi-target state modeling is performed based on the index coordinates, Doppler velocity, and local polynomial fitted contour curve of each suspected target to obtain the initial multi-target state set and initial joint probability distribution of all suspected targets at the initial time, wherein each suspected target in the initial multi-target state set includes an identification label;

[0011] Based on the GLMB filter and the multi-target state set and initial joint probability distribution at the initial time, multi-target tracking is performed at multiple time points after the initial time to obtain continuous trajectory data of each suspected target at the multiple time points.

[0012] Based on the continuous trajectory data of each suspected target at the multiple time points, the type of each suspected target is identified, and the identification result is obtained.

[0013] Preferably, constructing the range Doppler map based on the echo signal acquired at the initial moment includes:

[0014] The echo signal acquired at the initial moment is demodulated to obtain the baseband echo signal. :

[0015] ,

[0016] in, Let m be the baseband echo signal of the m-th pulse. The sampling time within a single pulse corresponds to the range information, and A is the complex amplitude including propagation loss and target scattering cross section. It is a natural exponential function. The imaginary unit satisfies , For radar carrier frequency, For time-varying delay, A single pulse;

[0017] The baseband echo signal is discretized to obtain a two-dimensional data matrix, wherein the row dimension of the two-dimensional data matrix corresponds to the sampling points within a single pulse, and the column dimension corresponds to the index of the pulse sequence;

[0018] Perform a Fourier transform on each column of the two-dimensional data matrix to obtain the distance image matrix. :

[0019] ,

[0020] in, For distance frequency unit index, For pulse index, This represents the total number of sampling points within a single pulse. for a discrete version with equal interval sampling in time dimension t, is the index of sampling point within a single pulse, ranging from 0 to , each sampling point corresponds to a range cell;

[0021] performing Fourier transform on each row of the range image matrix and adding a window function to obtain a range Doppler map :

[0022] ,

[0023] wherein, is the index of Doppler frequency cell, is the total number of pulses, is a window function used to suppress the sidelobes of Fourier transform.

[0024] Preferably, the suspected target detection and feature calculation according to the range Doppler map to obtain the position, velocity and profile feature vector of each suspected target comprises:

[0025] calculating the amplitude map of the range Doppler map;

[0026] performing detection and screening processing on the amplitude map using a constant false alarm rate detection algorithm to obtain a plurality of suspected targets and the index coordinates of each suspected target , wherein, is the index of range frequency cell where the suspected target is located, is the index of Doppler frequency cell where the suspected target is located;

[0027] for each suspected target, the Doppler velocity of the suspected target is calculated according to the index of Doppler frequency cell of the suspected target;

[0028] performing inverse Fourier transform on the index of range frequency cell in the index coordinates of each suspected target to restore it to the index of range cell ;

[0029] according to the index of range cell of the suspected target , the physical distance corresponding to the range cell is calculated, and the signal reflection intensity in the range near the range is collected to construct the sampling sample of the profile curve ;

[0030] ,

[0031] wherein, is the signal reflection intensity value corresponding to the range cell corresponding to the suspected target; is a signal reflection intensity function; the actual physical distance of the distance unit;

[0032] According to the preset polynomial order p and the sampling samples of the profile curve , a local polynomial fitting profile curve is obtained.

[0033] Preferably, the Doppler velocity of the suspected target is:

[0034] ,

[0035] ,

[0036] ,

[0037] wherein, is the resolution of the Doppler frequency, PRF is the pulse repetition frequency, is the period of a single pulse, is the total observation time, is the Doppler frequency, is the radar operating wavelength, is the speed of light.

[0038] Preferably, the local polynomial fitting profile curve obtained according to the preset polynomial order p and the sampling samples of the profile curve , includes:

[0039] According to the preset polynomial order p and the sampling samples of the profile curve , the profile curve is described by a p-order truncated Taylor series to obtain a polynomial :

[0040] ,

[0041] wherein, D is a distance variable, representing the distance value of any distance unit; is a differential operator, is the q-order derivative of the signal reflection intensity function with respect to the distance D; q is the derivative order index, ranging from 0 to p; is the distance unit at the maximum intensity;

[0042] The polynomial is rewritten to obtain the undetermined parameter polynomial corresponding to the profile curve:

[0043] ,

[0044] wherein, is the undetermined coefficient of the polynomial,​ It is the set of real numbers;

[0045] The weighted least squares method is used to analyze the polynomial with undetermined parameters. The solution is performed to obtain the contour feature vector. ;

[0046] contour feature vector Substitute the polynomial with undetermined parameters Obtain the local polynomial fitting profile curve .

[0047] Preferably, the step of performing multi-target state modeling based on the index coordinates, Doppler velocity, and local polynomial fitted contour curve of each suspected target to obtain the initial multi-target state set and initial joint probability distribution of all suspected targets at the initial time includes:

[0048] For each suspected target, the energy state of the suspected target is obtained based on its index coordinates and range Doppler map. The energy state includes distance range, center azimuth angle, and signal strength;

[0049] Based on the energy state of the suspected target Doppler velocity and local polynomial fitting profile curve Build extended state ;

[0050] Based on the extended state of each suspected target, construct an initial multi-target state set for all suspected targets at the initial time.

[0051] Assign an identification label to each suspected target in the initial multi-target state set;

[0052] Substituting the initial multi-target state set and the identification label of each suspected target into the multi-target joint probability distribution formula, initial modeling is performed to obtain the initial joint probability distribution of all suspected targets at the initial time. :

[0053] ,

[0054] , ,

[0055] in, The number of suspected targets is t, where t is the target index. for The extended state of each target; Let be the joint probability distribution function of multiple objectives. It is the probability of an empty set. The corresponding probability distribution is , is a different Bernoulli random set index, is the total number of Bernoulli random sets, is the survival probability of the th Bernoulli random set; is the probability distribution corresponding to the th Bernoulli random set.

[0056] Preferably, the GLMB filter and the initial time point of the multi-target state set and the initial joint probability distribution are used for multi-target tracking at multiple time points after the initial time point to obtain continuous trajectory data of each suspected target at the multiple time points, including:

[0057] The multi-target joint probability distribution is used as an initial posterior distribution at the initial time point, and a multi-target tracking Bayesian recursion operation is performed, including:

[0058] According to a preset multi-target state transition equation and the initial posterior distribution, a prior distribution at a next time point of the initial time point is predicted to obtain a next prior distribution corresponding to the next time point;

[0059] According to a preset multi-target likelihood function and a next measurement set obtained at the next time point, a generation probability of the next prior distribution generating the next measurement set is calculated, the next measurement set including a plurality of observation points at the next time point and physical observation data of each observation point, the observation points including suspected targets and missed detection information;

[0060] According to the generation probability, the next predicted distribution is corrected to obtain a next posterior distribution at the next time point;

[0061] The multi-target tracking Bayesian recursion operation is repeatedly performed to obtain posterior distributions at multiple time points;

[0062] According to the posterior distributions at the multiple time points, continuous trajectory data of each suspected target at the multiple time points is obtained.

[0063] Preferably, the next measurement set is obtained according to the following method:

[0064] According to an amplitude map of a range-Doppler map at the next time point, a constant false alarm rate detection algorithm is used for detection to obtain a plurality of observation points;

[0065] According to an index coordinate of each observation point, data extraction is performed to obtain physical observation data of each observation point;

[0066] ​According to the physical observation data of each observation point, a measurement set of all observation points at the next moment is constructed.

[0067] Preferably, according to the continuous trajectory data of each suspected target at the plurality of moments, the type of each suspected target is identified to obtain an identification result, including:

[0068] According to the continuous trajectory data of each suspected target at the plurality of moments, a trajectory feature vector of each suspected target is constructed.

[0069] The trajectory feature vector of each suspected target is input into a preset vector machine for classification to obtain an identification result output by the vector machine, and the identification result is used to represent whether each suspected target is a ship target.

[0070] Preferably, the vector machine is obtained by training in the following manner:

[0071] A group of training samples are obtained, each group of training samples including a trajectory feature vector and a label corresponding to the state feature vector, and the label is used to represent whether the target corresponding to the state feature vector is a ship target.

[0072] According to the group of training samples and a preset kernel function, a dual optimization problem is constructed.

[0073] A sequence minimum optimization algorithm is used to solve the dual optimization problem to obtain an optimal Lagrange multiplier.

[0074] According to the optimal Lagrange multiplier, the group of training samples and the kernel function, a support vector set and a bias term of a decision function are calculated.

[0075] According to the support vector set and the bias term of the decision function, the optimal Lagrange multiplier and the kernel function, a trained vector machine is constructed.

[0076] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0077] 1. By jointly extracting the index coordinates, Doppler velocity and contour curve based on local polynomial fitting of the target from the range-Doppler graph, the motion features and structural features of the target are combined, the dependence on single energy features is reduced, the detection and feature description of weak targets in a complex sea clutter background are more stable and reliable, and the robustness of target detection and feature extraction is improved.

[0078] 2. The initial multi-target state set containing identification labels and its joint probability distribution are constructed by using the fused multi-dimensional features, which can more completely describe the existence probability, state distribution and identity assumption of each potential target at the initial moment, provide more information and more rigorous representation for the initial conditions of subsequent tracking, and realize the unified modeling of multi-target state and its uncertainty.

[0079] 3. Based on the generalized labeled multi-Bernoulli (GLMB) filter, the characteristics of explicitly handling data association uncertainty and maintaining target identity uniqueness are used to estimate the state of the target group at multiple moments after the initial moment, and output continuous trajectory data with identity identification, thereby effectively associating the observations of the target at different moments.

[0080] 4. The continuous trajectory data obtained after tracking filtering is used instead of single-frame detection results to perform target type recognition. These time series data contain statistical rules of target motion and structural features, reduce the influence of random fluctuations, provide more discriminative input features for the classifier, thereby improving the classification accuracy of specific types of targets such as ships, and improving the accuracy of target type recognition. BRIEF DESCRIPTION OF DRAWINGS

[0081] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present application, the following will briefly introduce the drawings needed in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be considered as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor. In the drawings:

[0082] Figure 1 The flowchart of the radar target tracking and recognition method based on multi-dimensional information fusion provided by the present application. DETAILED DESCRIPTION

[0083] In order to make the purpose, technical solutions and advantages of the present application more clear and apparent, the following will further describe the present application in combination with embodiments and drawings. The exemplary embodiments of the present application and their descriptions are only used to explain the present application, and are not intended to limit the present application.

[0084] It is to be noted that the terms such as first and second, etc., are used herein merely to differentiate one entity or action from another, and do not necessarily require or imply any such actual relationship or order between such entities or actions. Also, the terms "including", "containing" or any other similar words are intended to encompass non-exclusive inclusion, so that a process, method, article, or apparatus including a list of elements does not necessarily include only those elements, but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without more limitations, the elements defined by the phrase "comprising" do not exclude the presence of additional identical elements in the process, method, article, or apparatus including the elements.

[0085] It is to be noted that all the actions of obtaining signals, information or data in the present application are carried out in compliance with the corresponding data protection regulations and policies of the place, and with the authorization given by the owner of the corresponding device.

[0086] Embodiment 1

[0087] See Figure 1 The embodiment of the present application provides a radar target tracking and identification method based on multi-dimensional information fusion, which comprises the following steps:

[0088] S1, constructing a range-Doppler map according to an echo signal obtained at an initial moment, wherein the echo signal is a sea echo received by a radar after transmitting a pulse to the sea;

[0089] This step aims to convert the original radar observation data into a time-frequency domain representation more suitable for target detection. The echo signal is a baseband complex signal returned after the electromagnetic pulse transmitted by the radar is scattered by the sea surface scene (which may contain targets and sea clutter), which contains the range and radial velocity information of the target. To solve the problem of low signal-to-noise ratio and information coupling in direct detection from the original signal, this step realizes information decoupling and energy aggregation through a signal processing flow: first, the echo of each pulse is subjected to a range dimension fast Fourier transform to convert the time domain signal into a range image, realizing pulse compression in the range dimension, improving the range resolution capability and signal-to-noise ratio; then, the echo sequence of multiple pulses across the same range unit is subjected to a Doppler dimension fast Fourier transform, and a window function is introduced to suppress spectral leakage, so as to focus the signal energy in the range-velocity (i.e. range-Doppler) two-dimensional plane. Finally, through the two orthogonal transformations, a range-Doppler map is obtained, the horizontal axis of which represents the range, and the vertical axis represents the radial velocity, and the amplitude value of each unit in the map corresponds to the signal energy in the range-velocity unit, providing a basis for subsequent detection and feature extraction.

[0090] In some embodiments, S1 constructs a range-Doppler map according to the echo signal obtained at the initial moment, including:

[0091] Demodulating the echo signal obtained at the initial moment to obtain a baseband echo signal :

[0092] ,

[0093] wherein, is the baseband echo signal of the mth pulse, is the sampling time within a single pulse, corresponding to range information, A is a complex amplitude containing propagation loss and target scattering cross section, is a natural exponential function, is an imaginary unit, satisfying , is the radar carrier frequency, is a time-varying delay, is a single pulse;

[0094] Discretizing the baseband echo signal to obtain a two-dimensional data matrix, wherein the row dimension of the two-dimensional data matrix corresponds to the sampling points within a single pulse, and the column dimension corresponds to the index of the pulse sequence;

[0095] Fourier transforming each column of the two-dimensional data matrix to obtain a range image matrix :

[0096] ,

[0097] wherein, is the distance frequency unit index, is the pulse index, is the total number of sampling points within a single pulse, is a discrete version of equal interval sampling in the time dimension t, is the sampling point index within a single pulse, ranging from 0 to , i.e. each sampling point corresponds to a range unit;

[0098] Fourier transforming each row of the range image matrix and adding a window function to obtain a range-Doppler map :

[0099] ,

[0100] wherein, is the Doppler frequency unit index, is the total number of pulses, is a window function used to suppress sidelobes of the Fourier transform.

[0101] S2, performing suspected target detection and feature calculation according to the range-Doppler map, obtaining index coordinates, Doppler velocity and local polynomial fitting contour curve of each suspected target in the range-Doppler map;

[0102] This step aims to extract potential target tracks from the background containing noise and clutter and calculate their discriminative features. Since sea clutter and weak targets may overlap in amplitude, a single amplitude threshold alone is prone to cause high false alarm or missed detection, so a more robust detection and feature extraction method is needed. In implementation, firstly, a constant false alarm rate detection algorithm is applied in the amplitude domain of the range-Doppler map, which can adaptively estimate the local background noise and clutter level around each detection cell and set a dynamic threshold, thereby screening out peak points with amplitude significantly higher than the local background. These peak points are suspected targets, and their positions are uniquely determined by the coordinates composed of the range cell index and the Doppler cell index. Then, based on the coordinates, feature calculation is performed: the Doppler cell index and system parameters (such as pulse repetition frequency) are used to calculate the accurate Doppler frequency, and then the radial velocity of the target is calculated according to the radar carrier frequency and the speed of light. This velocity information helps to distinguish between stationary clutter and moving targets. At the same time, a segment of signal reflection intensity sequence centered on the target's range cell is intercepted in the range dimension to construct its contour curve, and a weighted least squares method is used for local polynomial fitting. This fitting method gives higher weight to the sampling points close to the target center, effectively suppressing the fitting error caused by the fluctuation of the contour edge, and finally obtains a set of low-dimensional polynomial coefficient vectors representing the scattering structure characteristics of the target, i.e. the local polynomial fitting contour curve. This step enhances the ability of the subsequent processing link to distinguish between targets and clutter by fusing position, motion and contour multi-dimensional features.

[0103] In some embodiments, S2, the range-Doppler map performs suspected target detection and feature calculation to obtain the position, velocity and contour feature vector of each suspected target, including:

[0104] calculating an amplitude map of the range-Doppler map;

[0105] detecting and screening the amplitude map using a constant false alarm rate detection algorithm to obtain a plurality of suspected targets and index coordinates of each suspected target wherein is the range frequency cell index of the suspected target, is the Doppler frequency cell index of the suspected target;

[0106] For each suspected target, the Doppler velocity of the suspected target is calculated according to the Doppler frequency cell index of the suspected target.

[0107] inverse Fourier transform of the distance frequency unit index in the index coordinate of each suspected target, restoring to the distance unit index ;

[0108] According to the distance unit index of the suspected target , the physical distance corresponding to the distance unit is calculated, and the signal reflection intensity in the range near the distance is collected to construct the sampling sample of the profile curve ;

[0109] ,

[0110] Among them, is the signal reflection intensity value corresponding to the i-th distance unit corresponding to the suspected target; is the signal reflection intensity function; is the actual physical distance of the i-th distance unit; According to the preset polynomial order p and the sampling sample of the profile curve , the local polynomial fitting profile curve is obtained.

[0111] Among them, in some embodiments, the Doppler velocity of the suspected target is expressed as:

[0112]

[0113] ,

[0114] ,

[0115] ,

[0116] Among them, is the resolution of the Doppler frequency, PRF is the pulse repetition frequency, is the total number of pulses, is the period of a single pulse, is the total observation time, is the Doppler frequency, is the radar operating wavelength, is the speed of light.

[0117] Among them, in some embodiments, the local polynomial fitting profile curve obtained according to the preset polynomial order p and the sampling sample of the profile curve , includes:

[0118] According to the preset polynomial order p and the sampling sample of the profile curve ​​The contour curve is described by a Taylor series of order p truncation, resulting in a polynomial. :

[0119] ,

[0120] Where D is a distance variable, representing the distance value of any distance unit; For differential operators, Let q be the q-th derivative of the signal reflection intensity function with respect to distance D; q is the derivative order index, ranging from 0 to p; At the distance element where the intensity is maximum;

[0121] For the polynomial By rewriting, the polynomial with undetermined parameters corresponding to the profile curve is obtained. :

[0122] ,

[0123] in, The coefficients of the polynomial are undetermined. It is the set of real numbers;

[0124] Because the polynomial fitting has a large marginal error, a weighted least squares method is used to estimate the polynomial with the undetermined parameters using the Nadaraya-Watson method. The solution is performed to obtain the contour feature vector. :

[0125] ,

[0126] in, The coefficients are undetermined and include , This represents the total number of sampling points involved in the fitting process. , For kernel functions with tight support; For the first The weight of each sampling point;

[0127] contour feature vector Rewritten in matrix form:

[0128] ,

[0129] Where W is the weight matrix, and the diagonal elements are... Off-diagonal elements are 0. To design a matrix, the elements are , Here, q is the index of the sampling point, and q is the index of the derivative order. L is the vector of undetermined coefficients; L is the vector of signal reflection intensity.

[0130] The cost function is written as:

[0131] ,

[0132] The profile feature vector of the suspected target is described by the estimated parameter vector as:

[0133] ,

[0134] Where the superscript T denotes matrix transpose;

[0135] The profile feature vector is substituted into the undetermined parameter polynomial to obtain the local polynomial fitting profile curve .

[0136] S3, multi-target state modeling is performed according to the index coordinates, Doppler velocity and local polynomial fitting profile curve of each suspected target to obtain an initial multi-target state set and an initial joint probability distribution of all suspected targets at the initial time, wherein each suspected target in the initial multi-target state set includes an identification label;

[0137] This step aims to establish a complete and identity-included initial multi-target state probability model for subsequent Bayesian tracking. In complex sea conditions, multiple targets coexist with clutter, and their states (existence or not, specific parameters) are uncertain. Traditional list-based initialization cannot express this uncertainty and is difficult to maintain target identity in tracking. Therefore, this step uses the random finite set theory to model the multi-target state uniformly. In specific implementation, first, an extended state vector is constructed for each suspected target, which integrates its traditional energy state (distance, bearing and amplitude information converted from index coordinates), Doppler velocity and profile feature vector, thereby forming a richer target description. Then, all the extended state vectors are combined into a multi-target state set. To handle the uncertainty of target existence and maintain the uniqueness of identity, a labeled Bernoulli random finite set model is introduced: a unique identification label (which can be a binary tuple containing the starting time and serial number) is assigned to each target in the state set, and an existence probability and a probability distribution about its extended state are specified for each labeled target, wherein the existence probability is initialized based on the signal-to-noise ratio and other information of the CFAR detection in S2, and the state distribution is usually modeled as a Gaussian distribution with the measurement value as the mean and the sensor error as the covariance. Finally, the Bernoulli components of all labeled targets jointly constitute the joint probability distribution at the initial time, which describes which targets may exist at the initial time, their states and their identity labels, providing a recursive prior for Bayesian filtering.

[0138] In some implementations, S3 involves performing multi-target state modeling based on the index coordinates, Doppler velocity, and local polynomial fitted contour curve of each suspected target, to obtain the initial multi-target state set and initial joint probability distribution of all suspected targets at the initial time, including:

[0139] For each suspected target, the energy state of the suspected target is obtained based on its index coordinates and range Doppler map. The energy state includes distance range, center azimuth angle, and signal strength;

[0140] Based on the energy state of the suspected target Doppler velocity and local polynomial fitting profile curve Build extended state ;

[0141] Based on the extended state of each suspected target, construct an initial multi-target state set for all suspected targets at the initial time.

[0142] Assign an identification label to each suspected target in the initial multi-target state set;

[0143] Substituting the initial multi-target state set and the identification label of each suspected target into the multi-target joint probability distribution formula, initial modeling is performed to obtain the initial joint probability distribution of all suspected targets at the initial time. :

[0144] ,

[0145] , ,

[0146] in, The number of suspected targets is t, where t is the target index. for The extended state of each target; Let be the joint probability distribution function of multiple objectives. It is the probability of an empty set. The corresponding probability distribution is , Indexing different Bernoulli random sets, Let be the total number of Bernoulli random sets. For the first The survival probability of a Bernoulli random set; For the first Each Bernoulli random set corresponds to The probability distribution.

[0147] wherein the initial multi-target state set can be represented as:

[0148] ,

[0149] wherein, is a multi-target state set at the initial time, is a multi-target state set at the initial time, is a target state space .

[0150] S4, based on the GLMB (generalized labeled multi-Bernoulli) filter and the initial multi-target state set and the initial joint probability distribution at the initial time, performing multi-target tracking at a plurality of times after the initial time to obtain continuous track data of each suspected target at the plurality of times;

[0151] This step aims to recursively estimate the evolution state of the target over time using Bayesian filtering and maintain the identity continuity thereof in the presence of missed detection and clutter. Due to the complexity of the motion of the target at sea and the false alarm and missed detection of the radar measurement, it is difficult for a standard filter to stably track. The GLMB filter solves this problem by explicitly modeling a plurality of hypotheses of data association. The implementation process is a recursive prediction-update cycle: in the prediction step, the filter pushes forward each possible target state in the posterior distribution at the previous time according to the target motion model (such as the uniform speed or coordinated turn model) and the survival probability, while incorporating a new target probability distribution describing how a new target appears, to form a prediction distribution at the current time; in the update step, when a new measurement set (i.e., the result of S2 after processing a new frame of radar data) arrives, the filter considers all possible associations of the measurement with the existing target tracks (including missed detection), for each hypothesis, the weight of the hypothesis is calculated using a measurement likelihood function (the core of which is to calculate the probability of producing a specific measurement under a given target state), and the higher the weight, the more consistent the association mode is with the observation data. Finally, the posterior distribution is a mixture of all these association hypotheses with weights, and by estimating the posterior distribution (such as selecting the maximum posterior hypothesis), the most likely multi-target state estimation set at the current time can be output, wherein the state of each target is bound to a unique and unchanged identification label. Through recursive processing at a plurality of times, the state sequence corresponding to each label forms continuous track data, achieving stable tracking and track maintenance of multiple targets.

[0152] wherein, in some embodiments, the based on the GLMB filter and the initial multi-target state set and the initial joint probability distribution at the initial time, performing multi-target tracking at a plurality of times after the initial time to obtain continuous track data of each suspected target at the plurality of times includes:

[0153] perform a multi-target tracking Bayesian recursion operation, the multi-target tracking Bayesian recursion operation comprising:

[0154] predict a prior distribution of a next time point corresponding to the initial time point according to a preset multi-target state transition equation and the initial posterior distribution, to obtain a next prior distribution corresponding to the next time point;

[0155] calculate a generation probability of the next prior distribution generating the next measurement set according to a preset multi-target likelihood function and the obtained next measurement set, the next measurement set comprising a plurality of observation points and physical observation data of each observation point at the next time point, the observation points comprising suspected targets and missed detection information;

[0156] correct the next predicted distribution according to the generation probability, to obtain a next posterior distribution of the next time point;

[0157] perform the multi-target tracking Bayesian recursion operation cyclically, to obtain posterior distributions of a plurality of time points;

[0158] obtain continuous trajectory data of each suspected target at the plurality of time points according to the posterior distributions of the plurality of time points.

[0159] Further, the multi-target likelihood function is obtained by the following way:

[0160] constructing a probability distribution of appearing new suspected targets

[0161]

[0162] wherein Y is a random finite set of new-born targets; is a label distinguishability indicator function, 1 when Y is a finite set, otherwise 0, and is a multi-target new-born distribution given parameters; is a label set of Y corresponding weight function; is a probability parameter of new-born targets; is a probability contribution item of Y, is a probability distribution of the new-born target set Y;

[0163] In multi-target tracking, the targets do not interfere with each other when transitioning to the next state, and the new-born targets are independent of the surviving targets, so that the target likelihood function is:

[0164]

[0165] ​​​ ,

[0166] ,

[0167] where, is the Cartesian product of the target states and labels, i.e. the value space of the target states, and the function is a mapping from the label RFS to the label value space. is the next time step multi-target state set; X is the current time step multi-target state set; L is the label space (the set of all possible labels); is the probability distribution of the current state set X transitioning to the next state set ; is the state transition probability distribution of the surviving targets; is the state distribution of the new-born targets; W is the transition state set of the surviving targets; is an indicator function, which is 1 when the set is a finite set, otherwise 0; is an indicator function, which is 1 when the label set of X is valid, otherwise 0; is the weight function of the surviving target transition (for the target set X and the transition state set W); is the weight function of a single target (identification label , state ) transition; is the survival probability of a target with identification label , state ; is the death probability of a target with identification label , state ; is the probability distribution of a target with identification label , state transitioning to state ; is the pair of the current state and the next state of a target; is the label set of the transition state set W.

[0168] wherein, in some embodiments, the next measurement set is obtained according to the following method:

[0169] performing detection according to a constant false alarm rate detection algorithm of an amplitude map of the next time step range Doppler map to obtain a plurality of observation points;

[0170] performing data extraction according to the index coordinates of each observation point to obtain physical observation data of each observation point;

[0171] constructing a measurement set of all observation points at the next time step according to the physical observation data of each observation point.

[0172] The measurement set is represented as:

[0173] ,

[0174] in, for Multiple observation sets at different times; It is the observation space A set consisting of finite subsets of . It includes information on clutter, target observation, and missed detections.

[0175] In some implementations, the posterior distribution of the multi-objectives utilizes time 0 to... All measurement information at any given time, and satisfying the following relationship:

[0176] ,

[0177] in, A multi-objective state sequence from time 0 to k In the observation sequence The posterior distribution is given below; , It is a multi-objective The multi-objective likelihood function at time 1 / 2. It is a multi-objective The state transition function at time t; For the target set at time k Generate observation set Conditional events; For the target set at time k-1 Transition to the target set at time k The state transition function. In the multi-suspect target detection process under the Bayesian framework, the prediction and update of the suspected target state are as follows:

[0178] ,

[0179] ,

[0180] in, The multi-objective likelihood function at time k, specifically referring to... , Let k be the set of multi-objective states at time k. Let k be the observation set. For the integration over the target set X (the integral operator for random finite sets); For the target set at time k In the observation set The posterior probability distribution is as follows. For the target set at time k The prior probability distribution is predicted from the posterior distribution at time k-1.

[0181] The multi-objective likelihood function is obtained as follows:

[0182] The likelihood of generating the measurement set is Transition to the next state The probability distribution is z is the observed value (measured value) from the observation space Z. This is the state after the target has been transferred. The label after the target is transferred. For label indicator functions, when The value is 1 if it is true, otherwise it is 0 (the label value remains unchanged). In the current state ,Label Next, transition to the next state. The probability distribution, In the state ,Label The likelihood function for the observed value z is then generated.

[0183] The mapping between the target and the measurement is unique unless a detection is missed at that moment. The constructed multi-objective likelihood function is:

[0184] ,

[0185] ,

[0186] in, To establish a correlation mapping between suspected targets and observation points, the target is mapped to the observation ( Indicates label The target is related to the first One observation, (Indicates missed detection) For association mapping The corresponding weighting function, The set of labels for the target set X (extract the labels of all targets in X). For the set of all possible target-observation correlation mappings, To generate a multi-objective likelihood function for the observation set Z under the objective set X, Let be the clutter intensity function, describing the probability distribution of the clutter. This is the integral of the clutter intensity function (total clutter expectation). This represents the contribution of clutter to the observation set Z. For the tags The observations associated with the target, For tags ,state The probability of detecting the target. For tags ,state The probability of a missed detection of the target. For tags ,state Lower generation observation The likelihood function is given by X, where X is the target set and Z is the observation set.

[0187] S5. Based on the continuous trajectory data of each suspected target at the multiple time points, identify the type of each suspected target and obtain the identification result.

[0188] This step aims to use the stable trajectory data provided by tracking to perform semantic classification of targets in order to distinguish specific types of targets (such as ships). Since different types of targets (such as ships, buoys, and sea clutter false alarms) differ in motion characteristics and physical structure, but their features may overlap at a single moment and be affected by noise, it is necessary to use their stable patterns that evolve over time for identification. In implementation, firstly, discriminative features are extracted from each stable trajectory data. These features are the target extended state sequence output from S4, which may include statistically derived average radial velocity, radar cross-section statistics, and time-series mean or principal components of the contour feature vector, forming a high-dimensional feature vector. Subsequently, these feature vectors are used as training samples and input into the support vector machine classifier for model training. The SVM solves a convex optimization problem to find a hyperplane in the feature space that can optimally separate samples of different categories. For nonlinearly separable problems, kernel functions (such as radial basis functions) are used to map the samples to the high-dimensional feature space to achieve linear separation. In the recognition stage, the feature vector extracted from the target trajectory to be recognized is input into the trained SVM model. The model calculates the position of the feature vector relative to the classification hyperplane and outputs a classification decision function value. The target type (such as ship or non-ship) can be determined based on the sign of this value. This step extracts robust features by utilizing the temporal consistency information provided by the tracking process, and learns the nonlinear mapping between complex features and target categories through machine learning models, thereby achieving automated target recognition at the end of the overall process and improving the information output level of the system.

[0189] In some implementations, S5, based on the continuous trajectory data of each suspected target at the plurality of time points, identifies the type of each suspected target and obtains an identification result, including:

[0190] Based on the continuous trajectory data of each suspected target at the multiple time points, construct the trajectory feature vector of each suspected target;

[0191] The trajectory feature vector of each suspected target is input into a preset vector machine for classification to obtain an identification result output by the vector machine, and the identification result is used to represent whether each suspected target is a ship target.

[0192] In some embodiments, the vector machine is trained by the following method:

[0193] A set of training samples are obtained, each of which includes a trajectory feature vector and a corresponding label of the state feature vector, and the label is used to represent whether the target corresponding to the state feature vector is a ship target.

[0194] A dual optimization problem is constructed according to the set of training samples and a preset kernel function.

[0195] The sequence minimum optimization algorithm is used to solve the dual optimization problem to obtain an optimal Lagrange multiplier.

[0196] According to the optimal Lagrange multiplier, the set of training samples and the kernel function, a support vector set and a bias term of a decision function are calculated.

[0197] According to the support vector set and the bias term of the decision function, the optimal Lagrange multiplier and the kernel function, a trained vector machine is constructed.

[0198] Specifically, the purpose of vector machine classification is to find a hyperplane that can distinguish two classes and maximize the margin. In a dimensional space, the equation of the hyperplane is defined as:

[0199] ,

[0200] wherein, is a state vector of a target, is a component of the state vector of the target; is a coefficient vector of the hyperplane, which is used to define the direction of the hyperplane; b is an intercept term of the hyperplane, which determines the position of the hyperplane in the space. For dimensional linear SVM, the distance of the state vector of the target to the hyperplane is:

[0201] ,

[0202] wherein, is a constant term in the distance formula (obtained by transforming the hyperplane equation), and d is the distance of the state vector of the target to the hyperplane.

[0203] The state vector and the coefficient are both vectorized, and the distance formula can be regarded as:

[0204] ,

[0205] where, x is the target state vector (contains ); is the normal vector of the hyperplane; is the normal vector ; is the norm of the normal vector; is the dimension of the target state vector. The hyperplane is found by first finding the hyperplane that is closest to each class, then finding the hyperplane that is farthest from the sum of the distances:

[0206] ,

[0207] where, is the product of the sample class label and the classification function, used to determine whether the sample is within the classification interval, is the state vector of the th sample, and n is the number of samples.

[0208] Linear SVM only supports binary classification models, only two values, when the state vector is determined to be a ship target, the value is 1, otherwise it is not a target, and the value is -1. The hyperplane is:

[0209] ,

[0210] The points on both sides of the hyperplane satisfy:

[0211] ,

[0212] where y is the sample class label.

[0213] According to optimization theory, for can be converted into a restricted convex optimization problem, using the Karush-Kukn-Tucker (KKT) condition and the Lagrange formula, it can be deduced that the "decision boundary" is:

[0214] ,

[0215] where, is the number of support vector points, is the th support vector point, is the class label of the support vector (i.e. the state vector of the suspected target), is the intercept term of the decision boundary, is the Lagrange multiplier of the th support vector.

[0216] For information with nonlinear characteristics such as target contour, the input space sample is transformed to high-dimensional feature space by nonlinear mapping using nonlinear kernel function, and then the optimal classification hyperplane is solved. The nonlinear mapping is: The training algorithm only involves inner product operation between vectors, i.e. The dot product in the linear support vector classifier is replaced by the function , i.e.:

[0217] ,

[0218] where, is the kernel function, which is used to represent the inner product of the i-th and j-th sample feature vectors in the high-dimensional feature space (to avoid direct calculation of high-dimensional inner product), is the original state vector of the i-th sample, is the original state vector of the j-th sample, is the mapping of the original state vector of the i-th sample to the high-dimensional feature space, is the mapping of the original state vector of the j-th sample to the high-dimensional feature space.

[0219] The kernel function is applied to the dual Lagrange function, and the following is obtained:

[0220] ,

[0221] where, is the Lagrange multiplier of the i-th sample, is the Lagrange multiplier of the j-th sample, is the hyperplane normal vector, b is the hyperplane intercept term, and a is the Lagrange multiplier vector (containing ), is the dual Lagrange function, which is used to solve the optimal hyperplane of SVM, and N is the total number of training samples, is the i-th component of the normal vector w, is the j-th component of the normal vector w.

[0222] The decision function generated is:

[0223] ,

[0224] The radial basis kernel function is selected, i.e.:

[0225] ,

[0226] where, is the width of the kernel function, which controls the smoothness of the kernel function, is the Euclidean distance between the i-th and j-th sample original state vectors, is the natural exponential function, so that,

[0227] ,

[0228] wherein, is a radial basis kernel based decision function.

Claims

1. A radar target tracking and identification method based on multi-dimensional information fusion, characterized in that, The method comprises the following steps: constructing a range-Doppler map according to echo signals obtained at an initial moment, the echo signals being sea echo received after a radar transmits pulses to the sea; performing suspected target detection and feature calculation according to the range-Doppler map, to obtain index coordinates, Doppler velocities and local polynomial fitting contour curves of each suspected target in the range-Doppler map; performing multi-target state modeling according to the index coordinates, Doppler velocities and local polynomial fitting contour curves of each suspected target, to obtain an initial multi-target state set and an initial joint probability distribution of all suspected targets at the initial moment, wherein each suspected target in the initial multi-target state set comprises an identification label; performing multi-target tracking at multiple moments after the initial moment based on a GLMB filter and the initial multi-target state set and the initial joint probability distribution at the initial moment, to obtain continuous trajectory data of each suspected target at the multiple moments; identifying the type of each suspected target according to the continuous trajectory data of each suspected target at the multiple moments, to obtain an identification result.

2. The method of claim 1, wherein, The method of constructing a range-Doppler map according to echo signals obtained at an initial moment comprises the following steps: demodulating the echo signal acquired at the initial moment to obtain a baseband echo signal : ， wherein, is the baseband echo signal for the mth pulse, is the sampling time within a single pulse, corresponding to range information, A is the complex amplitude including propagation loss and target scattering cross section, is the natural exponential function, is the imaginary unit, satisfying , is the radar carrier frequency, is the time-varying delay, is a single pulse; performing discretization processing on the baseband echo signals, to obtain a two-dimensional data matrix, wherein the row dimension of the two-dimensional data matrix corresponds to sampling points within a single pulse, and the column dimension corresponds to the index of a pulse sequence; performing a Fourier transform on each column of the two-dimensional data matrix to obtain a range image matrix : ， wherein, is a distance bin index, is a pulse index, is the total number of sampling points within a single pulse, is is a discrete version of the signal sampled at equal intervals in the time dimension t, is a sampling point index within a single pulse, ranging from 0 to i.e. each sampling point corresponds to a distance bin; performing a Fourier transform on each row of the range image matrix and applying a window function to obtain a range Doppler map : ， wherein is the Doppler frequency unit index, is the total number of pulses, is a window function to suppress sidelobes of the Fourier transform.

3. The method of claim 1, wherein, The method of performing suspected target detection and feature calculation according to the range-Doppler map, to obtain the position, velocity and contour feature vector of each suspected target, comprises the following steps: calculating an amplitude map of the range-Doppler map; The amplitude diagram is detected and screened by using a constant false alarm rate detection algorithm to obtain multiple suspected targets and index coordinates of each suspected target wherein is a distance frequency unit index where the suspected target is located, is a Doppler frequency unit index where the suspected target is located; for each suspected target, calculating the Doppler velocity of the suspected target according to the Doppler frequency unit index of the suspected target; inverse fourier transforming the distance frequency cell indices in the indexed coordinates of each suspect object, restoring to distance cell indices ; According to the distance unit index of the suspected target , a physical distance corresponding to the distance unit is calculated, and the signal reflection intensity in the range near the distance is collected to construct a sampling sample of the profile curve ; ， wherein, is a signal reflection intensity value corresponding to the i-th distance unit of the suspected target; is a signal reflection intensity value corresponding to the i-th distance unit of the suspected target; is a signal reflection intensity function; is a signal reflection intensity value corresponding to the i-th distance unit of the suspected target; is a signal reflection intensity value corresponding to the i-th distance unit of the suspected target; According to a preset polynomial order p and sampling samples of the profile curve , a local polynomial fitting profile curve is obtained.

4. The method of claim 3, wherein, the doppler velocity of the suspected target is: ， ， ， wherein, is the resolution of the Doppler frequency, PRF is the pulse repetition frequency, is the period of a single pulse, is the total observation time, is the Doppler frequency, is the radar operating wavelength, is the speed of light.

5. The method of claim 3, wherein, The polynomial order p and the sampling samples of the profile curve are preset The local polynomial fitting profile curve is obtained by According to a predetermined polynomial order p and a sampling sample of the profile curve The profile curve is described by a polynomial of order p using a truncated Taylor series : ， wherein D is a distance variable representing a distance value of any distance bin; is a differential operator, is a qth derivative of the signal reflection intensity function with respect to distance D; q is a derivative order index, ranging from 0 to p; is a distance bin at which the maximum intensity occurs. The polynomial is rewritten as follows to obtain the undetermined parameter polynomial corresponding to the profile curve : ， wherein the undetermined coefficients of the polynomial, is the set of real numbers; Solving the undetermined parameter polynomial by using the weighted least square method to obtain the contour feature vector ; The profile feature vector is calculated as The profile feature vector is calculated as The profile feature vector is calculated as The profile feature vector is calculated as 6. The method of claim 1, wherein, The method of performing multi-target state modeling according to the index coordinates, Doppler velocities and local polynomial fitting contour curves of each suspected target, to obtain an initial multi-target state set and an initial joint probability distribution of all suspected targets at the initial moment, comprises the following steps: For each suspected target, an energy state of the suspected target is obtained according to the index coordinates and the range-Doppler map of the suspected target , the energy state including a range range, a center azimuth and a signal strength; According to the energy state of the suspected object Doppler velocity And local polynomial fitting profile curve Construct an extended state ; constructing an initial multi-target state set of all suspected targets at the initial moment according to the extended state of each suspected target; allocating an identification label to each suspected target in the initial multi-target state set; substitute the initial multi-target state set and the identification label of each suspected target into a multi-target joint probability distribution formula, perform initial modeling, and obtain an initial joint probability distribution of all suspected targets at the initial moment : ， ， ， wherein, is the number of targets of interest, t is the target index; is the extended state of the target; is the joint probability distribution function of multiple targets, is the probability of an empty set, is the corresponding probability distribution , is the different Bernoulli random set index, is the total number of Bernoulli random sets, is the survival probability of the th Bernoulli random set; is the probability distribution of the th Bernoulli random set corresponding to .

7. The method of claim 1, wherein, The method of performing multi-target tracking at multiple moments after the initial moment based on a GLMB filter and the initial multi-target state set and the initial joint probability distribution at the initial moment, to obtain continuous trajectory data of each suspected target at the multiple moments, comprises the following steps: performing a multi-target tracking Bayesian recursion operation by taking the multi-target joint probability distribution as an initial posterior distribution at the initial moment, the multi-target tracking Bayesian recursion operation comprising: predicting a prior distribution of a next moment corresponding to the initial moment according to a preset multi-target state transition equation and the initial posterior distribution, to obtain a next prior distribution corresponding to the next moment; According to a preset multi-target likelihood function and an obtained next measurement set of the next time, a generation probability of the next prior distribution generating the next measurement set is calculated, the next measurement set comprising a plurality of observation points and physical observation data of each observation point at the next time, the observation points comprising suspected targets and missed detection information; According to the generation probability, the next predicted distribution is corrected to obtain a next posterior distribution at the next time; The multi-target tracking Bayesian recursion operation is repeatedly executed to obtain posterior distributions at a plurality of times; According to the posterior distributions at the plurality of times, continuous trajectory data of each suspected target at the plurality of times is obtained.

8. The method of claim 7, wherein, The next measurement set is obtained according to the following method: Constant false alarm rate detection algorithm is performed according to an amplitude graph of a range-Doppler graph at the next time to obtain a plurality of observation points; According to index coordinates of each observation point, data extraction is performed to obtain physical observation data of each observation point; According to the physical observation data of each observation point, a measurement set of all observation points at the next time is constructed.

9. The method of claim 1, wherein, According to the continuous trajectory data of each suspected target at the plurality of times, the type of each suspected target is identified to obtain an identification result, comprising: According to the continuous trajectory data of each suspected target at the plurality of times, a trajectory feature vector of each suspected target is constructed; The trajectory feature vector of each suspected target is input into a preset vector machine for classification to obtain an identification result output by the vector machine, the identification result being used to represent whether each suspected target is a ship target.

10. The method of claim 9, wherein, The vector machine is obtained by training in the following manner: A group of training samples are obtained, each training sample comprising a trajectory feature vector and a label corresponding to the state feature vector, the label being used to represent whether the target corresponding to the state feature vector is a ship target; According to the group of training samples and a preset kernel function, a dual optimization problem is constructed; A sequential minimal optimization algorithm is used to solve the dual optimization problem to obtain an optimal Lagrange multiplier; According to the optimal Lagrange multiplier, the group of training samples and the kernel function, a support vector set and a bias term of a decision function are calculated; According to the support vector set and the bias term of the decision function, the optimal Lagrange multiplier and the kernel function, a trained vector machine is constructed.

Images