A Maneuvering Target Cluster Tracking Method Based on Gaussian Process

Abstract

The present invention discloses a method for tracking a cluster of maneuvering targets based on a Gaussian process, comprising the following steps: S1, selecting the centroid of a cluster, describing the cluster morphological contour through a star convex model, and constructing a contour state space model; S2, constructing a centroid motion state space model, and constructing a direction evolution state space model; S3, constructing a joint state space equation; S4, obtaining target cluster measurements, and recursively estimating the target cluster state online; S5, visualizing the target cluster state, and obtaining the target cluster tracking result. The present invention takes into account the target cluster contour, centroid motion, and direction angle, and solves the problem of inaccurate tracking when unpredictable maneuvers occur within the cluster; constructing a target cluster state estimator through an interactive multi-model, and recursively estimating the target cluster state online, thereby achieving high-precision and stable tracking of a maneuvering target cluster.

Description

Technical Field

[0001] The invention belongs to the technical field of unmanned aerial vehicle tracking, and in particular relates to a mobile target cluster tracking method based on Gaussian process. Background Art

[0002] When drone swarms are in motion, traditional multi-target tracking algorithms are difficult to track them correctly due to the limitations of radar resolution and computer computing power. This type of highly dense and agile target swarms poses severe challenges to air defense radars.

[0003] A series of representative methods have been developed to solve the tracking problem of such maneuvering target clusters, especially the motion modeling problem, including the state expansion method based on point targets, the ellipse / multi-ellipse contour modeling method based on random matrices, and the star convex contour modeling method based on Gaussian processes. Among them, the state expansion method based on point targets can obtain the motion state of each target in the cluster, but the high dimension of the expanded state leads to a large amount of calculation, and when the number of targets is too large, this method often faces the risk of failure due to matrix calculation; the ellipse / multi-ellipse contour modeling method based on random matrices can overcome the shortcomings of the state expansion method in terms of high computational requirements, but using one or more ellipses to describe the contour of the target cluster is difficult to meet the tracking accuracy requirements for reliable estimation of the morphological contour for high-precision tracking scenarios, especially when the contour morphology changes rapidly; the star convex contour modeling method based on Gaussian processes can more accurately describe the contour of irregularly shaped target clusters, but the existing methods do not consider the situation where the targets in the cluster are unpredictable. Therefore, how to maintain high-precision and stable tracking when the target cluster is maneuvering and considering the unpredictable maneuvers of the targets in the cluster is a technical problem that needs to be solved urgently. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a mobile target cluster tracking method based on Gaussian process.

[0005] The present invention is achieved through the following technical solutions.

[0006] The present invention provides a Gaussian process-based mobile target cluster tracking method, comprising the following steps:

[0007] S1. Select the cluster centroid, describe the cluster morphological profile by a star convex model, and construct a profile state space model according to the cluster morphological profile and the cluster centroid by a Gaussian process. The profile state space model is used to estimate the target cluster profile by cluster profile, sampling interval and input position.

[0008] S2. constructing a center of mass motion state space model through a jump Markov process, and constructing a direction evolution state space model through a Singer model, wherein the center of mass motion state space model is used to estimate the center of mass motion, and the direction evolution state space model is used to estimate the direction angle;

[0009] S3, constructing a joint state space equation according to the contour state space model, the center of mass motion state space model and the orientation evolution state space model, wherein the joint state space equation is used to jointly estimate the target cluster from multiple aspects, wherein the multiple aspects include the target cluster contour, the center of mass motion and the orientation angle;

[0010] S4, obtaining a target cluster measurement, where the target cluster measurement is an observed value of a state of a corresponding point in the target cluster, constructing a target cluster state estimator according to the target cluster measurement and a joint state space equation through an interactive multi-model framework, and recursively estimating the target cluster state online through the target cluster state estimator;

[0011] S5. Visualize the target cluster status and obtain the target cluster tracking result.

[0012] Preferably, constructing the profile state space model according to the cluster morphological profile and the cluster centroid by using the Gaussian process comprises the following steps:

[0013] Select representative points on the cluster morphology outline;

[0014] The radial distance from the representative point to the cluster centroid is taken as the cluster morphological contour state vector;

[0015] The cluster morphology contour state vector is modeled and characterized through Gaussian process to obtain the contour state space model.

[0016] Preferably, constructing the center of mass motion state space model by using a jump Markov process comprises the following steps:

[0017] Construct the center of mass motion state vector;

[0018] The center of mass motion state vector is modeled and characterized through the jump Markov process to obtain the center of mass motion state space model.

[0019] Preferably, the step of constructing an evolutionary state space model by using the Singer model comprises the following steps:

[0020] Assume angular acceleration towards;

[0021] A heading angular acceleration model is constructed according to the heading angular acceleration through the Singer model, wherein the heading angular acceleration model is an exponentially autocorrelated zero-mean random process;

[0022] Perform a whitening operation on the orientation angular acceleration model to define the orientation estimated state vector;

[0023] A model is constructed according to the maneuvering acceleration model and the estimated orientation state vector to obtain an orientation evolution state space model.

[0024] Preferably, constructing the heading angular acceleration model according to the heading angular acceleration by using the Singer model comprises the following steps:

[0025] Obtain the probability density function of the angular acceleration according to the angular acceleration;

[0026] Calculate the angular acceleration variance according to the angular acceleration probability density function;

[0027] The heading angular acceleration model is constructed based on the heading angular acceleration variance.

[0028] Preferably, a target cluster state estimator is constructed according to the target cluster measurement and the joint state space equation through an interactive multi-model framework, and online recursive estimation of the target cluster state through the target cluster state estimator includes the following steps:

[0029] Construct a global coordinate system;

[0030] Construct measurement equations based on the global coordinate system and target cluster measurements;

[0031] Constructing a joint state space model according to the joint state space equation and the measurement equation;

[0032] The target cluster state estimator is constructed based on the measurement equations and the joint state space model through an interactive multi-model framework;

[0033] The target cluster state is obtained through online recursive estimation of the target cluster state estimator.

[0034] Preferably, obtaining the target cluster state by online recursive estimation by the target cluster state estimator comprises the following steps:

[0035] Computing interactive input data according to the joint state-space model;

[0036] The target cluster state estimator calculates the interactive input data to obtain the expanded state estimation, measurement residual and information covariance;

[0037] Calculate the pattern posterior probability based on the measurement residual and information covariance;

[0038] The target cluster state is calculated based on the expanded state estimate and the pattern posterior probability.

[0039] Preferably, the target cluster state includes target cluster morphological contour state estimation, centroid motion state estimation and cluster orientation state estimation.

[0040] Preferably, the visual processing of the target cluster state and obtaining the target cluster tracking result comprises the following steps:

[0041] Obtain the radial value at the current moment by estimating the target cluster morphological contour state;

[0042] Obtain the center of mass state at the current moment through the center of mass motion state estimation;

[0043] Obtain the current orientation through cluster orientation state estimation;

[0044] Obtain the target cluster contour representative point set according to the current radial value, the current centroid state and the current orientation;

[0045] The target cluster contour is characterized according to the representative point set to obtain the target cluster tracking result.

[0046] The beneficial effects of the present invention are:

[0047] The present invention uses a star convex model to describe the cluster morphological contour, constructs a contour state space model through a Gaussian process, constructs a center of mass motion state space model through a jump Markov process, constructs an orientation evolution state space model through a Singer model, and constructs a joint state space equation. It takes into account multiple aspects such as the target cluster contour, center of mass motion and orientation angle, and solves the problem of inaccurate tracking when unpredictable maneuvers occur in the targets within the cluster; constructs a target cluster state estimator, and recursively estimates the target cluster state online through an interactive multi-model framework, thereby achieving high-precision and stable tracking of the maneuvering target cluster. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is a flow chart of a method provided by an embodiment of the present invention;

[0049] Figure 2 is a schematic diagram of a model interaction structure provided by an embodiment of the present invention;

[0050] Figure 3 is a target cluster measurement schematic diagram provided by an embodiment of the present invention;

[0051] Figure 4 It is a schematic diagram of the interactive multi-model framework structure provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The technical solution of the present invention is further described below, but the scope of protection claimed is not limited to the description.

[0053] like Figure 1 As shown, a mobile target cluster tracking method based on Gaussian process includes the following steps:

[0054] S1. Select the cluster centroid, describe the cluster morphological profile by a star convex model, and construct a profile state space model according to the cluster morphological profile and the cluster centroid by a Gaussian process. The profile state space model is used to estimate the target cluster profile by cluster profile, sampling interval and input position.

[0055] Different from common simple geometric shapes, such as ellipses / ellipsoids, rectangles / cubic parallelepipeds, etc., the present invention describes the cluster morphological contour through a star convex model. The star convex model refers to drawing a straight line from the center point of the model to the contour point, and the straight line is included in the model range. Therefore, the contour of the star convex model can be described by the radial value function r=f(θ) of polar coordinates, where θ represents the polar angle in the polar coordinate system and r represents the polar diameter in the polar coordinate system.

[0056] The constructing of the profile state space model according to the cluster morphological profile and the cluster centroid by using the Gaussian process comprises the following steps:

[0057] Select representative points on the cluster morphology outline;

[0058] The radial distance from the representative point to the cluster centroid is taken as the cluster morphological contour state vector;

[0059] If the radial value function at the representative point position on the contour and the cluster centroid are known, the spatial contour of the cluster can be fitted, and the spatial contour of the cluster is the cluster morphological contour state vector. In other words, the estimation of the cluster morphological contour state vector can be converted into an estimation of the radial value function at a series of representative point positions.

[0060] The cluster morphology contour state vector is modeled and characterized through Gaussian process to obtain the contour state space model.

[0061] The present invention models and characterizes the cluster morphological profile state vector through a Gaussian process, that is, a recursive Gaussian process is used to estimate the radial value function. The Gaussian process is a random process and can be regarded as the distribution of the function f(θ), which can be uniquely defined by the mean function m(θ) and the covariance function κ(θ,θ'), and the expression is as follows:

[0062] m(θ)=E[f(θ)]

[0063] κ(θ,θ')=E((f(θ)-m(θ))(f(θ')-m(θ')) T )

[0064] Furthermore, the Gaussian process can be written as:

[0065] f(θ)~GP(m(θ),κ(θ,θ'))

[0066] Among them, θ is the input of the function.

[0067] Consider the following measurement model form:

[0068] z k =f(θ k )+e k , e k ~N(0,R)

[0069] Among them, e k represents Gaussian white noise with zero mean and covariance R.

[0070] The purpose of Gaussian process regression is to use a series of noisy measurements z = [z1…z k ] T and the input at the corresponding position θ=[θ1…θ m ] T To learn other input positions The function value at n is the number of radial value functions at representative points to be learned. According to the properties of the Gaussian process, the joint distribution of the measurement z and the function value f can be described as follows:

[0071]

[0072] in, is the Kronecker product, K represents the Gaussian process covariance matrix;

[0073] The Gaussian process covariance matrix of the m-dimensional vector p and the n-dimensional vector q is as follows:

[0074]

[0075] where κ(·,·) is the covariance function.

[0076] Here, we use a modified square exponential covariance function to measure the correlation between radial value functions at different input positions, with a period of 2π, expressed as,

[0077]

[0078] in, represents the prior variance of the radial distance of the contour, l represents the scale parameter, Represents the variance of the radial distance from the mean.

[0079] The above regression is a batch processing process, which cannot solve the problem of target tracking that requires real-time processing. 1:k ) and make some reasonable assumptions to obtain its recursive form as follows:

[0080] p(f|z 1:k)∝p(z k |f,z 1:k-1 )p(f|z 1:k-1 )∝p(z k |f,z 1:k-1 )…p(f)

[0081] Approximately conditional on the measurement z 1:k-1 ,but:

[0082] p(z k |f,z 1:k-1 )≈p(z k |f)

[0083] Therefore, measuring z k The joint distribution of and function value f is as follows:

[0084]

[0085] Among them, θ k represents the measurement z at time k k The corresponding radial value function input.

[0086] Then, the likelihood function follows the following distribution:

[0087]

[0088] and,

[0089]

[0090] in, Represented as the k-time measurement matrix, Measure z at time k k The corresponding Gaussian white noise covariance matrix.

[0091] By using the above likelihood structure, Gaussian recursion can be performed on the following state space model:

[0092]

[0093] in, Represents the modeling error of the cluster outline, which is often assumed to have a mean of 0 and a variance of Gaussian distribution, represents zero mean and covariance is Gaussian white noise, α is the forgetting factor of the cluster profile, T is the sampling interval, is the function value at the desired input location.

[0094] Through Gaussian process, the contour state space model is obtained.

[0095] It should be noted that the contour state space model is built on the target body, that is, the input angles of the radial value function are obtained with reference to the local coordinate system.

[0096] S2. constructing a center of mass motion state space model through a jump Markov process, and constructing a direction evolution state space model through a Singer model, wherein the center of mass motion state space model is used to estimate the center of mass motion, and the direction evolution state space model is used to estimate the direction angle;

[0097] The method of constructing the center of mass motion state space model by using the jump Markov process comprises the following steps:

[0098] Construct the center of mass motion state vector;

[0099] Considering that the motion characteristics of each target in the target cluster are similar, the maneuvering motion of the target cluster can be uniformly characterized by the maneuvering motion of the center of mass point target. For the modeling of the center of mass maneuvering motion, the state vector of the center of mass motion of the target cluster in the two-dimensional plane rectangular coordinate system o-ξη is

[0100] The center of mass motion state vector is modeled and characterized through the jump Markov process to obtain the center of mass motion state space model.

[0101] If acceleration is considered, the acceleration variables of each axis are expanded to the state vector; if the three-dimensional plane is further considered, the existing state vector is further expanded in the third dimension. On this basis, the jump Markov process is used to characterize the center of mass maneuvering motion as follows:

[0102]

[0103] Among them, Θ k represents the jump Markov parameter under the finite pattern set {1,…,N}, and its model transition probability p ij =P{Θ k+1 =j|Θ k =i}, the initial time mode probability In the i-th mode, the system matrix in the corresponding state evolution equation is F i,k Here, F i,k It can represent target motion nominal models such as uniform linear motion, uniform left turn motion, uniform right turn motion, etc.

[0104] Assume that the model set contains three modes, namely, the uniform linear motion model CV, the left turn model LCT with a fixed and known turning angular velocity, and the right turn model RCT with a fixed and known turning angular velocity. The corresponding state transfer matrix is ​​as follows:

[0105] When the target moves at a constant speed,

[0106]

[0107] When the target makes a uniform turn,

[0108]

[0109] Where T is the sampling interval, ω is the turning rate, when ω>0, it is a left turn, and when ω<0, it is a right turn. Furthermore, process noise Used to describe model uncertainty, it is often assumed to have a mean of 0 and a covariance matrix Gaussian distribution.

[0110] Common The expression is as follows:

[0111]

[0112] in, is the corresponding covariance coefficient and T is the sampling interval.

[0113] The method of constructing an evolutionary state space model through the Singer model comprises the following steps:

[0114] Assume angular acceleration towards;

[0115] The maneuvering motion of the target cluster often leads to unpredictable changes in the cluster's orientation. At this time, it is difficult to obtain accurate orientation angle estimation results if the orientation angle transformation is modeled using uniform speed or uniform acceleration motion.

[0116] Here, the heading angle change can be regarded as an unpredictable maneuvering motion, that is, the heading angle acceleration is assumed.

[0117] Considering the complexity of the algorithm, the angular acceleration is modeled using the Singer model without the need for maneuver detection.

[0118] A heading angular acceleration model is constructed according to the heading angular acceleration through the Singer model, wherein the heading angular acceleration model is an exponentially autocorrelated zero-mean random process;

[0119] The method of constructing the heading angular acceleration model according to the heading angular acceleration by using the Singer model comprises the following steps:

[0120] Obtain the probability density function of the angular acceleration according to the angular acceleration;

[0121] Calculate the angular acceleration variance according to the angular acceleration probability density function;

[0122] Usually, it is assumed that the angular acceleration is in the interval [-aM ,a M ] is approximately uniformly distributed, and the maneuvering acceleration is equal to the maximum value a M or minimum value-a M The probability of M , the probability of being equal to 0 is p0. Through these assumptions, the probability density function of maneuvering acceleration can be obtained, and the variance of maneuvering acceleration can be further obtained as:

[0123]

[0124] The heading angular acceleration model is constructed based on the heading angular acceleration variance.

[0125] The Singer model considers the maneuver model to be a correlated noise model, rather than the commonly assumed white noise model. It models the target acceleration a(t) as a zero-mean random process with exponential autocorrelation, namely:

[0126]

[0127] in, and ρ are the unknown parameters that determine the maneuvering characteristics of the target in the interval [t, t+τ], is the variance of the angular acceleration, and ρ is the inverse of the maneuvering time constant, i.e., the maneuvering frequency.

[0128] Perform a whitening operation on the orientation angular acceleration model to define the orientation estimated state vector;

[0129] Because the Singer model believes that the heading angular acceleration is a correlated random process, but the Kalman filter requires it to be an uncorrelated white noise process, it is necessary to further "whiten" the maneuvering acceleration. On the basis of "whitening", the estimated heading state vector is defined, that is, the heading angle, heading angular velocity, and heading angular acceleration are used as the estimated state vector of the cluster heading.

[0130] A model is constructed according to the maneuvering acceleration model and the estimated orientation state vector to obtain an orientation evolution state space model.

[0131] The discretized state space equation of the corresponding Singer model, that is, the evolutionary state space model, can be expressed as follows:

[0132]

[0133] In the formula,

[0134]

[0135]

[0136] S3, constructing a joint state space equation according to the contour state space model, the center of mass motion state space model and the orientation evolution state space model, wherein the joint state space equation is used to jointly estimate the target cluster from multiple aspects, wherein the multiple aspects include the target cluster contour, the center of mass motion and the orientation angle;

[0137] The joint state space equation is constructed according to the contour state space model, the center of mass motion state space model and the orientation evolution state space model, namely:

[0138]

[0139] in, is the state evolution matrix of the cluster centroid. Similarly, are the state evolution matrices of cluster orientation angle and cluster outline respectively. Q k is the corresponding modeling error w k The covariance matrix of

[0140] Here, since the center-of-mass maneuver is modeled as a jump Markov process, the new expanded-dimensional state transfer matrix Still contains mode uncertainty. If the center of mass maneuvering motion includes uniform linear motion and uniform left / right turning motion with known angular velocity, that is, Then the corresponding set of extended-dimensional state transfer matrices consists of three modes, namely here, Represents the set of extended-dimensional state transfer matrices under different motion modes, where

[0141]

[0142] Here, F CV 、F LCT and F RCT Represent the state transfer matrices of uniform linear motion, uniform left turn motion and uniform right turn motion respectively.

[0143] like Figure 2 As shown, S4, obtaining target cluster measurement, where the target cluster measurement is the observed value of the state of the corresponding point in the target cluster, constructing a target cluster state estimator according to the target cluster measurement and the joint state space equation through an interactive multi-model framework, and recursively estimating the target cluster state online through the target cluster state estimator;

[0144] The target cluster state estimator is constructed based on the target cluster measurement and the joint state space equation through the interactive multi-model framework. The online recursive estimation of the target cluster state through the target cluster state estimator includes the following steps:

[0145] Construct a global coordinate system;

[0146] Construct measurement equations based on the global coordinate system and target cluster measurements;

[0147] Generally speaking, the target cluster measurements are often generated in the radar coordinate system, where the radar is located at the origin of the coordinate system, and the radar coordinate system is denoted as the global coordinate system G. On the other hand, the contour modeling of the target cluster is often established in the target cluster itself, the so-called local coordinate system L. Figure 3 As shown, the global coordinate system G and the local coordinate system L are often not the same. Therefore, the transformation relationship between each measurement in the global coordinate system and the extended dimension state at each moment, that is, the measurement equation, can be described as follows:

[0148]

[0149] Among them, β k represents the orientation angle of the target cluster at time k, c k =(ξ k ,η k ) T represents the centroid of the target cluster, represents the n obtained at time k k A set of measurements, z k,l represents the lth measurement at time k; f(θ) represents the radial value function; and They represent the angles between the line connecting the lth measurement at time k and the centroid of the target cluster in the global and local coordinate systems, respectively; represents the zero mean of the lth measurement at time k, with a covariance of R e Gaussian white noise; Represents a unit direction vector, that is:

[0150]

[0151] and

[0152]

[0153] Because the radial value function cannot be measured in practice, the radial value in the above measurement equation is replaced by the estimated radial distance of the cluster contour points based on the Gaussian process, that is,

[0154]

[0155] in,

[0156] e k,l ~N(0,Y k,l )

[0157]

[0158] Constructing a joint state space model according to the joint state space equation and the measurement equation;

[0159] The constructed joint state space model is as follows:

[0160]

[0161] z k,l =h k,l (x k )+e k,l , e k,l ~N(0,Y k,l )

[0162] The target cluster state estimator is constructed based on the measurement equations and the joint state space model through an interactive multi-model framework;

[0163] like Figure 4 As shown, it is a schematic diagram of the interactive multi-model framework structure. The interactive multi-model framework includes the interaction of extended-dimensional states, nonlinear sub-filter updates, multiplicative likelihood functions, sub-mode posterior probability updates, and interactive output of extended-dimensional states.

[0164] Based on the joint state space model and the measurement model, the pattern uncertainty in the evolution of the extended-dimensional state is considered, and the interactive multiple mode, Interacting Multiple Mode, or IMM, is used to estimate and realize the online recursion of the extended-dimensional state. At the same time, considering the nonlinearity of the measurement model, the sub-filter of the IMM estimator selects a nonlinear filter, such as the Extended Kalman Filter, or EKF, in this embodiment.

[0165] The target cluster state is obtained through online recursive estimation of the target cluster state estimator.

[0166] The method of obtaining the target cluster state by online recursive estimation of the target cluster state estimator includes the following steps:

[0167] Computing interactive input data according to the joint state-space model;

[0168] In this embodiment, the recursive estimation from time k-1 to time k is taken as an example.

[0169] make and is the expanded dimensional state estimate of sub-mode j at time k-1 and its corresponding covariance matrix. In the target cluster state estimator, the interactive input data of filter j at time k is as follows:

[0170]

[0171] Where N is the total number of modes, is the posterior probability of mode i at time k-1.

[0172] The target cluster state estimator calculates the interactive input data to obtain the expanded state estimation, measurement residual and information covariance;

[0173] Enter the interactive data And measurement As the input of the jth filter in the target cluster state estimator at time k, the extended Kalman filter is executed to obtain the output of the jth filter at time k That is, the dimension-expanded state estimation. During the filtering process, the measurement residual and information covariance are obtained at the same time.

[0174] Calculate the pattern posterior probability based on the measurement residual and information covariance;

[0175] According to the measurement residual obtained by the above filtering process and its corresponding information covariance, the likelihood probability corresponding to the measurement l in the jth model is calculated as follows:

[0176]

[0177] in, is the measurement z during filtering k,l The difference between its predicted value and Assuming that different measurements are independent of each other, the overall likelihood function can be written as the product of the likelihood functions corresponding to each measurement, that is,

[0178]

[0179] Based on this, the posterior probability of the jth pattern is calculated as follows:

[0180]

[0181] The target cluster state is calculated based on the expanded state estimate and the pattern posterior probability.

[0182] According to the upper expanded-dimensional state estimation and the mode posterior probability, the final expanded-dimensional state estimation at time k is weighted and comprehensively expressed as follows:

[0183]

[0184] The target cluster state includes target cluster morphological contour state estimation, centroid motion state estimation and cluster orientation state estimation.

[0185] Split Then the target cluster morphological contour state estimation can be obtained Center of mass motion state estimation and cluster orientation state estimation That is, the target cluster status is obtained.

[0186] S5. Visualize the target cluster status and obtain the target cluster tracking result.

[0187] The visualization process of the target cluster state and obtaining the target cluster tracking result comprises the following steps:

[0188] Obtain the radial value at the current moment by estimating the target cluster morphological contour state;

[0189] Obtain the center of mass state at the current moment through the center of mass motion state estimation;

[0190] Obtain the current orientation through cluster orientation state estimation;

[0191] Assume that the current centroid state estimated at time k is The current direction is The current radial value is

[0192] Obtain the target cluster contour representative point set according to the current radial value, the current centroid state and the current orientation;

[0193] Then the coordinates of the mth representative point on the estimated target cluster contour are:

[0194]

[0195] in,

[0196] The target cluster contour is characterized according to the representative point set to obtain the target cluster tracking result.

[0197] The calculation of the above representative point coordinates is completed in the radar coordinate system. Therefore, based on the centroid point of the target cluster and the target cluster contour represents the point set The overall characterization of the target cluster centroid / orientation / contour in the radar coordinate system at time k can be completed, and the target cluster tracking result can be obtained.

[0198] The present invention uses a star convex model as the cluster morphological contour, constructs a contour state space model through a Gaussian process, constructs a center of mass motion state space model through a jump Markov process, constructs an orientation evolution state space model through a Singer model, and constructs a joint state space equation. The present invention takes into account multiple aspects such as the target cluster contour, center of mass motion and orientation angle, and solves the problem of inaccurate tracking when unpredictable maneuvers occur within the cluster. A target cluster state estimator is constructed, and the target cluster state is estimated online recursively through the target cluster state estimator, thereby achieving high-precision and stable tracking of the maneuvering target cluster.

Claims

1. A mobile target cluster tracking method based on Gaussian process, characterized in that: The following steps are involved: S1. Select the cluster centroid, describe the cluster morphological profile by a star convex model, and construct a profile state space model according to the cluster morphological profile and the cluster centroid by a Gaussian process. The profile state space model is used to estimate the target cluster profile by cluster profile, sampling interval and input position. S2. constructing a mass center motion state space model through a jump Markov process, and constructing a direction evolution state space model through a Singer model, wherein the mass center motion state space model is used to estimate the mass center motion, and the direction evolution state space model is used to estimate the direction angle; S3, constructing a joint state space equation according to the contour state space model, the center of mass motion state space model and the orientation evolution state space model, wherein the joint state space equation is used to jointly estimate the target cluster from multiple aspects, wherein the multiple aspects include the target cluster contour, the center of mass motion and the orientation angle; S4, obtaining a target cluster measurement, where the target cluster measurement is an observed value of a state of a corresponding point in the target cluster, constructing a target cluster state estimator according to the target cluster measurement and a joint state space equation through an interactive multi-model framework, and recursively estimating the target cluster state online through the target cluster state estimator; S5. Visualize the target cluster status and obtain the target cluster tracking result; The method of constructing an evolutionary state space model through the Singer model comprises the following steps: Assume angular acceleration towards; A heading angular acceleration model is constructed according to the heading angular acceleration through the Singer model, wherein the heading angular acceleration model is an exponentially autocorrelated zero-mean random process; Perform a whitening operation on the orientation angular acceleration model to define the orientation estimated state vector; A model is constructed according to the maneuvering acceleration model and the estimated orientation state vector to obtain an orientation evolution state space model; Obtaining the target cluster state through online recursive estimation by the target cluster state estimator includes the following steps: Computing interactive input data according to the joint state-space model; The target cluster state estimator calculates the interactive input data to obtain the expanded state estimation, measurement residual and information covariance; Calculate the pattern posterior probability based on the measurement residual and information covariance; The target cluster state is calculated based on the expanded state estimate and the pattern posterior probability.

2. The mobile target cluster tracking method according to claim 1, characterized in that: The constructing of the profile state space model according to the cluster morphological profile and the cluster centroid by using the Gaussian process comprises the following steps: Select representative points on the cluster morphology outline; The radial distance from the representative point to the cluster centroid is taken as the cluster morphological contour state vector; The cluster morphology contour state vector is modeled and characterized through Gaussian process to obtain the contour state space model.

3. The mobile target cluster tracking method according to claim 1, characterized in that: The method of constructing the center of mass motion state space model by using the jump Markov process comprises the following steps: Construct the center of mass motion state vector; The center of mass motion state vector is modeled and characterized through the jump Markov process to obtain the center of mass motion state space model.

4. The mobile target cluster tracking method according to claim 1, characterized in that: The method of constructing the heading angular acceleration model according to the heading angular acceleration by using the Singer model comprises the following steps: Obtain the probability density function of the angular acceleration according to the angular acceleration; Calculate the angular acceleration variance according to the angular acceleration probability density function; The heading angular acceleration model is constructed based on the heading angular acceleration variance.

5. The mobile target cluster tracking method according to claim 1, characterized in that: The target cluster state estimator is constructed according to the target cluster measurement and the joint state space equation through the interactive multi-model framework, and the target cluster state is estimated online recursively through the target cluster state estimator, including the following steps: Construct a global coordinate system; Construct measurement equations based on the global coordinate system and target cluster measurements; Constructing a joint state space model according to the joint state space equation and the measurement equation; The target cluster state estimator is constructed based on the measurement equations and the joint state space model through an interactive multi-model framework; The target cluster state is obtained through online recursive estimation of the target cluster state estimator.

6. The mobile target cluster tracking method according to claim 1, characterized in that: The target cluster state includes target cluster morphological contour state estimation, centroid motion state estimation and cluster orientation state estimation.

7. The method for tracking a mobile target cluster according to claim 6, characterized in that: The visualization process of the target cluster state and obtaining the target cluster tracking result comprises the following steps: Obtain the radial value at the current moment by estimating the target cluster morphological contour state; Obtain the center of mass state at the current moment through the center of mass motion state estimation; Obtain the current orientation through cluster orientation state estimation; Obtain the target cluster contour representative point set according to the current radial value, the current centroid state and the current orientation; The target cluster contour is characterized according to the representative point set to obtain the target cluster tracking result.

Images