Multi-target tracking method, device and equipment based on track poisson multi-bernoulli

Abstract

The application relates to a multi-target tracking method, device and equipment based on a track Poisson multi-Bernoulli. The method comprises the following steps: performing discretization processing on a continuous-time multi-target motion model constructed by a Gaussian moment approximation method, obtaining a track Poisson multi-Bernoulli density of a multi-target at a previous time step and sequentially performing prediction and updating on a track Poisson multi-Bernoulli density of a current time step based on obtained discretized new target probability density and single-target track state transition density under non-uniform sampling conditions, and finally constructing a track Poisson multi-Bernoulli posterior density according to an updated value of the track Poisson multi-Bernoulli density of the current time step to estimate a multi-target track of the current time step, so that multi-target track tracking is realized. The method can solve the multi-target tracking problem under the non-uniform sampling condition, has high track estimation precision, and has a good engineering application prospect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar target tracking, in particular to a multi-target tracking method, device and equipment based on track Poisson multi-Bernoulli. BACKGROUND

[0002] The multi-target tracking methods of the classic data association class and the stochastic finite set class are recursively implemented in the Bayesian framework based on the uniformly sampled multi-target generation, motion and survival model. However, in the phased array radar system, the resources are shared and limited by multiple functions (search, tracking and guidance, etc.). Therefore, a certain resource scheduling method must be used to allocate the corresponding beam residence time and sampling interval for the target, that is, to maximize the overall tracking performance while minimizing the resources by effectively using the tracking beam. Specifically, when the target maneuvers, a larger sampling interval is used, and when the target maneuvers, a smaller sampling interval is used. In addition, due to the influence of occlusion, delay and interference, non-uniformly sampled radar tracking data will also be formed.

[0003] When the radar sampling time is not uniform, both the classic data association class and the stochastic finite set class cannot be recursively implemented in real time and effectively, so it has become a technical problem to be solved in the field to study the multi-target tracking under the condition of non-uniform radar sampling interval. SUMMARY

[0004] Therefore, it is necessary to provide a multi-target tracking method, device and equipment based on track Poisson multi-Bernoulli which can realize multi-target tracking under the condition of non-uniform sampling in view of the above technical problems.

[0005] A multi-target tracking method based on track Poisson multi-Bernoulli, the method comprising:

[0006] constructing a continuous-time multi-target motion model, wherein the continuous-time multi-target motion model comprises a target generation model, a target extinction model and a target motion model;

[0007] discretizing the continuous-time multi-target motion model by a Gaussian moment approximation method to obtain a discretized new target probability density and a single target track state transition density;

[0008] Under the condition of non-uniform sampling, the track Poisson multi-Bernoulli density of the multi-target at the previous time step is obtained according to the discretized new target probability density and the single target track state transition density, and the track Poisson multi-Bernoulli density at the current time step is sequentially predicted and updated according to the track Poisson multi-Bernoulli density at the previous time step to obtain the updated value of the track Poisson multi-Bernoulli density at the current time step; wherein the updated value of the track Poisson multi-Bernoulli density at the current time step comprises a Poisson component updated value and a Bernoulli component updated value;

[0009] The track Poisson multi-Bernoulli posterior density is constructed according to the combination of the Poisson component update value and the Bernoulli component update value, and the multi-target track at the current time step is estimated according to the track Poisson multi-Bernoulli posterior density and a preset threshold, so as to obtain the multi-target track estimation at the current time step.

[0010] In one of the embodiments, the target generation model is subject to a Poisson random process with a Poisson intensity , and the target state in the target generation model is subject to a Gaussian distribution with a mean value and a covariance matrix , wherein represents the average position and the average velocity, respectively, and represent the position covariance matrix and the velocity covariance matrix, respectively, represents the position and velocity covariance matrix, and the superscript T represents the matrix transpose.

[0011] In one of the embodiments, the target extinction model is represented as

[0012]

[0013] wherein p(τ) represents that the life cycle τ of the target is independent of each other and subject to an exponential distribution with a rate parameter μ.

[0014] In one of the embodiments, the target motion model is constructed based on a Wiener velocity model and is represented as

[0015] dx(t)=Ax(t)dt+Ldβ(t);

[0016]

[0017] wherein represents the multi-target state, and dx(t) represents the differential of x(t), represents the single-target state space, A represents a first matrix with a dimension of n x ×n x , L represents a second matrix with a dimension of n x ×n β , n x represents the dimension of the single-target state, represents a Wiener process with a diffusion matrix Q β =qI d , n represents an n β -dimensional real number space, n β =d represents the dimension, q represents the parameter of the Wiener velocity model, 0 d represents a zero matrix with a dimension of d, and I d represents a unit matrix with a dimension of d.

[0018] In one embodiment, the continuous-time multi-target motion model is discretized by the Gaussian moment approximation method to obtain a discretized new-born target probability density, denoted as

[0019]

[0020] wherein, denotes the distribution of the number of targets n in the sampling time step interval Δt, denotes the discretized Poisson intensity, j denotes the Gaussian component index, J β,k denotes the number of Gaussian components, denotes the Gaussian component, denotes the discretized track start time, denotes the discretized weight, denotes the discretized mean, is the state variable of the qth arriving target at the kth time step, denotes the discretized covariance matrix.

[0021] In one embodiment, the continuous-time multi-target motion model is discretized by the Gaussian moment approximation method to obtain a single-target track state transition density as

[0022]

[0023] wherein, X = (ι, x 1:ν ) ∈ X k denotes any one of the track variables in the track set X k at the kth time step, ι y denotes the track start time, v denotes the track duration, x 1:ν denotes the state sequence of the track, denotes the single-target state transition density, δ ι [ι y ]、δ v+1 [v y ]、 respectively denote the Kronecker function, the transition matrix and the transition covariance.

[0024] In one embodiment, under the condition of non-uniform sampling, the track Poisson multi-Bernoulli density of the multi-target at the previous time step is obtained according to the discretized new-born target probability density and the single-target track state transition density, including:

[0025] Under the condition of non-uniform sampling, the track Poisson multi-Bernoulli density of the multi-target at the k-1th time step is obtained according to the discretized new-born target probability density and the single-target track state transition density, denoted as

[0026]

[0027] where k'∈{k,k-1} represents the time between k-1 time step and k time step, X represents the total track set, represents the Bernoulli track set, n k'|k-1 represents the number of Bernoulli components, Y represents the Poisson component track set, represents the Poisson component in the track Poisson multi-Bernoulli density at k-1 time step, represents the Bernoulli component in the track Poisson multi-Bernoulli density at k-1 time step.

[0028] In one of the embodiments, the track Poisson multi-Bernoulli density at the current time step is sequentially predicted and updated according to the track Poisson multi-Bernoulli density at the previous time step, to obtain the updated value of the track Poisson multi-Bernoulli density at the current time step, comprising:

[0029] The track Poisson multi-Bernoulli density at k time step is predicted according to the track Poisson multi-Bernoulli density at k-1 time step, to obtain the predicted value of the track Poisson multi-Bernoulli density at k time step, represented as

[0030]

[0031] where, represents the Bernoulli track set at k time step, n k|k-1 represents the number of Bernoulli components at k time step, represents the Poisson component prediction value at k time step, represents the Bernoulli component prediction value at k time step;

[0032] The Poisson component prediction value at k time step and the Bernoulli component prediction value at k time step are updated respectively, to obtain the Poisson component updated value at k time step and the Bernoulli component updated value at k time step.

[0033] In one of the embodiments, the track Poisson multi-Bernoulli posterior density is constructed according to the combination of the Poisson component updated value and the Bernoulli component updated value, and the multi-target track at the current time step is estimated according to the track Poisson multi-Bernoulli posterior density and the pre-set threshold, to obtain the multi-target track estimation at the current time step, comprising:

[0034] The track Poisson multi-Bernoulli posterior density is constructed according to the combination of the Poisson component updated value and the Bernoulli component updated value, and the multi-target track at k time step is estimated according to the track Poisson multi-Bernoulli posterior density and the pre-set threshold Γ d The track estimation at k time step is obtained by estimating the multi-target track at k time step as where ι i represents the track start time of the i-th Bernoulli component, an updated value representing a mean of the i th Bernoulli component at the k th time step, an updated value representing a presence probability of the i th Bernoulli component at the k th time step.

[0035] A multi-target tracking device based on track Poisson multi-Bernoulli, the device comprising:

[0036] a model construction module configured to construct a continuous-time multi-target motion model, wherein the continuous-time multi-target motion model comprises a target generation model, a target death model and a target motion model;

[0037] a discretization processing module configured to discretize the continuous-time multi-target motion model by a Gaussian moment approximation method to obtain a discretized new target probability density and a single-target track state transition density;

[0038] a prediction and update module configured to, under a non-uniform sampling condition, obtain a track Poisson multi-Bernoulli density of the multi-target at a previous time step according to the discretized new target probability density and the single-target track state transition density, and sequentially predict and update the track Poisson multi-Bernoulli density at a current time step according to the track Poisson multi-Bernoulli density at the previous time step to obtain an updated value of the track Poisson multi-Bernoulli density at the current time step, wherein the updated value of the track Poisson multi-Bernoulli density at the current time step comprises a Poisson component updated value and a Bernoulli component updated value;

[0039] a track estimation module configured to construct a track Poisson multi-Bernoulli posterior density according to a combination of the Poisson component updated value and the Bernoulli component updated value, estimate the multi-target track at the current time step according to the track Poisson multi-Bernoulli posterior density and a pre-set threshold, and obtain a multi-target track estimation at the current time step.

[0040] A computer device comprising a memory and a processor, the memory storing a computer program, and the processor implementing the following steps when executing the computer program:

[0041] constructing a continuous-time multi-target motion model, wherein the continuous-time multi-target motion model comprises a target generation model, a target death model and a target motion model;

[0042] discretizing the continuous-time multi-target motion model by a Gaussian moment approximation method to obtain a discretized new target probability density and a single-target track state transition density;

[0043] Under the non-uniform sampling condition, the track Poisson multi-Bernoulli density of the multiple targets at the previous time step is obtained according to the discretized new target probability density and the single-target track state transition density, and the track Poisson multi-Bernoulli density at the current time step is sequentially predicted and updated according to the track Poisson multi-Bernoulli density at the previous time step, to obtain an updated value of the track Poisson multi-Bernoulli density at the current time step; wherein the updated value of the track Poisson multi-Bernoulli density at the current time step includes an updated value of a Poisson component and an updated value of a Bernoulli component;

[0044] The track Poisson multi-Bernoulli posterior density is constructed according to the combination of the updated value of the Poisson component and the updated value of the Bernoulli component, and the multiple-target track at the current time step is estimated according to the track Poisson multi-Bernoulli posterior density and a preset threshold, to obtain the multiple-target track estimation at the current time step.

[0045] The above multiple-target tracking method, device and equipment based on the track Poisson multi-Bernoulli, through the Gaussian matrix approximation method, discretize the constructed continuous-time multiple-target motion model, and based on the obtained discretized new target probability density and single-target track state transition density, under the non-uniform sampling condition, obtain the track Poisson multi-Bernoulli density of the multiple targets at the previous time step and sequentially predict and update the track Poisson multi-Bernoulli density at the current time step, and finally construct the track Poisson multi-Bernoulli posterior density according to the updated value of the track Poisson multi-Bernoulli density at the current time step to estimate the multiple-target track at the current time step, to realize the multiple-target track tracking. The scheme can solve the multiple-target tracking problem under the non-uniform sampling condition, has high track estimation precision, and has good engineering application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 A flowchart of a multiple-target tracking method based on the track Poisson multi-Bernoulli in an embodiment is shown;

[0047] Figure 2 A real track diagram in an embodiment is shown;

[0048] Figure 3 A state estimation diagram of a classical CD-PHD (continuous-discrete PHD) method at k = 1:88 in an embodiment is shown;

[0049] Figure 4 A state estimation diagram of a CD-CPHD (continuous-discrete CPHD) method at k = 1:88 in an embodiment is shown;

[0050] Figure 5 A state estimation diagram of a CD-PMBM (continuous-discrete PMBM) method at k = 1:88 in an embodiment is shown

[0051] Figure 6Fig. 6 is a schematic diagram of real-time tracking results according to the NUS-TPMB method in one embodiment;

[0052] Figure 7 Fig. 7 is a schematic diagram of performance comparison of the CD-PHD, CD-CPHD, CD-PMBM and NUS-TPMB methods in one embodiment;

[0053] Figure 8 Fig. 8 is a schematic diagram of target state estimation results of the CD-PHD, CD-CPHD, CD-PMBM and NUS-TPMB methods in one embodiment;

[0054] Figure 9 Fig. 9 is a schematic diagram of position error comparison of the CD-PHD, CD-CPHD, CD-PMBM and NUS-TPMB methods in one embodiment;

[0055] Figure 10 Fig. 10 is a schematic diagram of missed target cost results of the CD-PHD, CD-CPHD, CD-PMBM and NUS-TPMB methods in one embodiment;

[0056] Figure 11 Fig. 11 is a schematic diagram of false target cost results of the CD-PHD, CD-CPHD, CD-PMBM and NUS-TPMB methods in one embodiment;

[0057] Figure 12 Fig. 12 is a schematic diagram of internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0058] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application.

[0059] First, a unified symbol definition is given: (·) k represents the value at the k time step, (·)0represents the initial value, (·) k|k-1 represents the prediction value of the k-1 time step to the k time step, (·) k|k represents the updated value at the k time step, (·) (i) represents a physical quantity related to the Gaussian component with index number i;

[0060] Second, the definition of a single-track Gaussian component is given: represents the Gaussian component with index number i at the k time step, and the weight is the start time is the mean value is the covariance is the index number i∈{1,2,...,Ik}, where I k This represents the number of Gaussian components at time step k;

[0061] Considering the continuous generation, movement, and disappearance of targets, measurements are obtained in discrete time. The multi-target state at time t∈[0,∞) is... in, For a single-objective state space, Get state x at time 1 k =x(t) k The corresponding measurement, k time steps, corresponds to time t. k Assume the measurement set obtained at time step k is It includes measurements of target generation and clutter, given a single target state x k Each target has a detection probability p D (x) detection, generating the measurement with conditional density l(·|x), or with a false negative probability 1-p D (x) Missed detection, clutter is independent of the target corresponding measurement and follows a Poisson distribution with intensity κ(z).

[0062] In one embodiment, such as Figure 1 As shown, a multi-target tracking method based on Poisson and Bernoulli trajectory is provided, including the following steps:

[0063] Step S1: Construct a continuous-time multi-target motion model, which includes a target generation model, a target disappearance model, and a target motion model.

[0064] Step S2: Discretize the continuous-time multi-target motion model using the Gaussian moment approximation method to obtain the discretized new target probability density and single-target trajectory state transition density.

[0065] Step S3: Under non-uniform sampling conditions, based on the discretized new target probability density and single target track state transition density, obtain the Poisson-Bernoulli density of multiple targets in the previous time step, and predict and update the Poisson-Bernoulli density of the current time step sequentially based on the Poisson-Bernoulli density of the previous time step to obtain the updated value of the Poisson-Bernoulli density of the current time step; wherein, the updated value of the Poisson-Bernoulli density of the current time step includes the updated value of the Poisson component and the updated value of the Bernoulli component.

[0066] Step S4: Construct the Poisson-Bernoulli posterior density of the track based on the combination of the Poisson component update value and the Bernoulli component update value. Estimate the multi-target track at the current time step based on the track Poisson-Bernoulli posterior density and a pre-set threshold to obtain the multi-target track estimate at the current time step.

[0067] In one embodiment, the target generation model is a Poisson random process with Poisson intensity The target state in the target generation model is a Gaussian distribution with mean and covariance matrix where denote the mean position and mean velocity, respectively, and denote the position covariance matrix and velocity covariance matrix, respectively, denote the position and velocity covariance matrix, and the superscript T denotes the matrix transpose.

[0068] In one embodiment, the target death model is represented as

[0069]

[0070] where p(τ) denotes the life cycle τ of the target is mutually independent and subject to an exponential distribution with rate parameter μ, and the average life of the target is 1 / μ.

[0071] In one embodiment, the target motion model is constructed based on a Wiener velocity model and is represented as

[0072] dx(t) = Ax(t)dt + Ldβ(t);

[0073]

[0074] where denotes the multi-target state, and dx(t) denotes the differential of x(t), denotes the single-target state space, A denotes a first matrix with dimension n x × n x , L denotes a second matrix with dimension n x × n β , n x denotes the dimension of the single-target state, denotes a Wiener process with diffusion matrix Q β = qI d , n β denotes an n β -dimensional real space, n d = d denotes the dimension, q denotes the parameter of the Wiener velocity model, 0 d denotes a d-dimensional all-zero matrix, and I k-1 denotes a d-dimensional identity matrix.

[0075] Specifically, the specific steps of constructing the target motion model include:

[0076] First, since the target motion is mutually independent and subject to a linear time-invariant stochastic differential equation, for the multi-target state Its target motion equation is

[0077] dx(t) = Ax(t)dt + Ldβ(t) ;

[0078] Since Poisson process has independent increment property and odd degree, the distribution of target arrival number n in time interval (t, t + Δt) is is expressed as

[0079]

[0080] where, is Poisson intensity.

[0081] Then, calculate t k-1 to t k The target survival probability in time interval Δt k = t k - t k-1 is

[0082] If the target appears in time t k -t', t' ∈ [0, Δt k ], and t' represents the time difference between the target appearance time and t k , its spatial density at time t k -t' is independent of the rest of the targets, and is given by:

[0083]

[0084] Thus, the target state transition density at t k is

[0085]

[0086] Finally, the Wiener velocity model is used to model the single target motion, considering the single target state as

[0087] x(t k ) = [p1(t k ), υ1(t k ), …, p d (t k ), υ d (t k )] ;

[0088] where d = n x / 2, p i (t k ) is the i-th dimensional position at t k , and υ i (t k ) is the i-th dimensional velocity at t kThe velocity of the i-th dimension at time t, the target moves in the respective direction according to the linear velocity model, and the target motion model is constructed as

[0089] dx(t) = Ax(t)dt + Ldβ(t);

[0090]

[0091] Further, according to the target motion model, the transition matrix and the covariance matrix in the target state transition density can be obtained, respectively, as

[0092]

[0093]

[0094] In one embodiment, the continuous-time multi-target motion model is discretized by the Gaussian approximation method to obtain a discretized new target probability density, which is expressed as

[0095]

[0096]

[0097]

[0098] wherein, represents the distribution of the number of targets n arriving within the sampling time step interval Δt, represents the discretized Poisson intensity, j represents the Gaussian component index, and J β,k represents the number of Gaussian components, represents the Gaussian component, represents the discretized track start time, represents the discretized weight, represents the discretized mean, is the state variable of the q-th arriving target at the k-th time step, represents the discretized covariance matrix. Specifically, are respectively expressed as

[0099]

[0100]

[0101] wherein,

[0102]

[0103]

[0104]

[0105]

[0106]

[0107] C[t'] = E[t' 2 ] - (E[t']) 2 ;

[0108]

[0109] where C[t'] is a constant, E[t'], E[t' 2 ], and E[t' 3 ] represent the first, second, and third order moments, respectively.

[0110] In one embodiment, the continuous-time multi-target motion model is discretized by the Gaussian moment approximation method to obtain a single-target track state transition density, denoted as

[0111]

[0112]

[0113] where X = (i, x 1:ν ) ∈ X k represents any track variable in the track set X k existing at the kth time step, i y represents the track start time, v represents the track duration, and x 1:ν represents the state sequence of the track, denotes a single-target state transition density, δ ι [i y ], δ v+1 [v y ], represent the Kronecker function, the transition matrix, and the transition covariance, respectively. The specific formulas of the transition matrix F k and the covariance matrix Q k at the kth time step are as follows:

[0114]

[0115]

[0116] In one embodiment, under the condition of non-uniform sampling, the track Poisson-Multinomial density of the multi-target at the previous time step is obtained according to the discretized new target probability density and the single-target track state transition density, including:

[0117] Under the condition of non-uniform sampling, the track Poisson multi-Bernoulli density of the multi-target at the k-1 time step is obtained by recursively calculating the track Poisson multi-Bernoulli density according to the discretized new-born target probability density and the single-target track state transition density, and is expressed as

[0118]

[0119] wherein k'∈{k,k-1} represents the time between the k-1 time step and the k time step, X represents the total track set, represents the Bernoulli track set, n k'|k-1 represents the number of Bernoulli components, Y represents the Poisson component track set, represents the Poisson component in the track Poisson multi-Bernoulli density at the k-1 time step, represents the Bernoulli component in the track Poisson multi-Bernoulli density at the k-1 time step. Specifically, the single-track density of the i th Bernoulli component at the k time step has the following Gaussian form

[0120]

[0121] The intensity of the Poisson component

[0122]

[0123] wherein is the number of Poisson components, is the weight of the j th Poisson component, is the start time of the j th Poisson component, is the mean of the j th Poisson component, is the covariance of the j th Poisson component.

[0124] In one embodiment, the track Poisson multi-Bernoulli density at the current time step is sequentially predicted and updated according to the track Poisson multi-Bernoulli density at the previous time step, to obtain the updated value of the track Poisson multi-Bernoulli density at the current time step, including:

[0125] The track Poisson multi-Bernoulli density at the k time step is predicted according to the track Poisson multi-Bernoulli density at the k-1 time step, to obtain the predicted value of the track Poisson multi-Bernoulli density at the k time step, and is expressed as

[0126]

[0127] wherein represents the Bernoulli track set at the k time step, n k|k-1 represents the number of Bernoulli components at the k time step, represents the predicted value of the Poisson component at the k time step, ​a Bernoulli component prediction value at the k-th time step;

[0128] updating the Poisson component prediction value at the k-th time step and the Bernoulli component prediction value at the k-th time step respectively to obtain a Poisson component updated value at the k-th time step and a Bernoulli component updated value at the k-th time step.

[0129] Specifically, after the prediction step is performed, the existence probability of the Bernoulli component with the index i is

[0130]

[0131]

[0132] the mean of the Bernoulli component with the index i is

[0133]

[0134] the covariance matrix of the Bernoulli component with the index i is

[0135]

[0136] the state transition matrix is the state dimension is

[0137] the prediction intensity of the Poisson component in the track Poisson multi-Bernoulli density is

[0138]

[0139]

[0140]

[0141]

[0142] when the update step is performed, the update intensity of the Poisson component at the k-th time step is

[0143] D k|k (X) = (1 - p D )D k|k-1 (X);

[0144] wherein, p D is a detection probability; for the Bernoulli component, the update thereof is composed of a missed detection hypothesis, a detection hypothesis and a new Bernoulli component facing the measurement. Wherein, the missed detection hypothesis updates the missed detection hypothesis update weight of the Bernoulli component with the index i, updates the existence probability, and updates the single-track distribution, which are respectively represented as

[0145]

[0146]

[0147]

[0148] The Bernoulli component with index i has a probability of existence and an update weight of

[0149]

[0150] The track distribution is

[0151] The augmented measurement matrix is

[0152]

[0153]

[0154] The predicted and innovation covariances of the measurement are

[0155]

[0156]

[0157] where H is the measurement matrix and R is the measurement noise covariance matrix. The mean and covariance updates of the Bernoulli component with index i at time step k are

[0158]

[0159]

[0160] The measurement The initial newborn Bernoulli component i contains two sets of local hypotheses: hypothesis 1 is

[0161]

[0162] where denotes the measurement index of the Bernoulli component i and the local hypothesis J. For hypothesis 2, first compute the υ q

[0163]

[0164] where υ q denotes the probability hypothesis density component with index q, the predicted mean of the probability hypothesis density component with index q, denotes the likelihood of the measurement being , the augmented measurement matrix and the innovation covariance matrix S​​​q respectively,

[0165]

[0166] wherein, denote the predicted covariance of the probability hypothesis density component with index q, the weight and existence probability of the Bernoulli component, respectively,

[0167]

[0168] wherein, denote the clutter intensity, the single-track distribution of hypothesis 2 with index i,

[0169]

[0170] wherein, denote the start time of the single-track with index q*, the mean and covariance of the single-track of hypothesis 2 with index i, respectively,

[0171]

[0172]

[0173] wherein, denote the predicted mean and covariance of the probability hypothesis density component with index q*, respectively, denote the augmented measurement matrix and innovation covariance of the probability hypothesis density component with index q*, respectively, denote the transpose. The newborn Bernoulli component single-track density is Gaussian mixture, whose start time and length can be different. By Gaussian approximation, the computational efficiency of the filter can be significantly improved. For this purpose, the Gaussian component with the highest weight is taken, whose index is q* = max q (υ q ). This process is called absorption.

[0174] In one embodiment, a track Poisson multi-Bernoulli posterior density is constructed according to a combination of a Poisson component update value and a Bernoulli component update value, a multi-target track at a current time step is estimated according to the track Poisson multi-Bernoulli posterior density and a preset threshold, to obtain a multi-target track estimate at the current time step, comprising:

[0175] A track Poisson multi-Bernoulli posterior density is constructed according to a combination of a Poisson component update value and a Bernoulli component update value, a multi-target track at a current time step is estimated according to the track Poisson multi-Bernoulli posterior density and a preset threshold Γ d The multi-target track at the k time step is estimated, and the track estimate at the k time step is wherein, ι i denotes the track start time of the i-th Bernoulli component, an updated value representing a mean of the i-th Bernoulli component at the k-th time step, an updated value representing a probability of existence of the i-th Bernoulli component at the k-th time step.

[0176] In order to better illustrate the technical solutions of the present application, the present application is further described below in combination with simulation experiments:

[0177] The simulation experiment environment is an Intel i7 3Hz frequency 8-core CPU processor, and the program is written in Matlab language. The simulation experiment compares the method proposed in the present application with a continuous-discrete PHD (CD-PHD) algorithm, a continuous-discrete CPHD (CD-CPHD) algorithm, and a continuous-discrete PMBM algorithm (CD-PMBM).

[0178] 1. Simulation conditions

[0179] A square two-dimensional region with a size of [0, 750] x [0, 350] (m) is selected as a monitoring region;

[0180] A total of 20 targets appear in the simulation process, and the real tracks are as shown in Figure 2 In the figure, the solid line represents the real track, the hollow circle represents the track starting position, the number represents the track starting time, and the black "+" represents the sampling point. As can be seen from Figure 2 , the targets are densely distributed and closely adjacent to each other, and the measurement sampling interval is non-uniform, so the standard multi-target model cannot achieve real-time and effective target tracking.

[0181] The state of the target is represented as wherein is the position of the target in the horizontal and vertical directions, is the speed of the target in the horizontal and vertical directions; the continuous-time model parameters of the multi-target dynamic system are: μ = 0.01 s -1 , d = 2, q = 0.2 m 2 / s 3 , According to the parameter μ, the average life cycle of the target is 100 s, and the number of targets obeys a Poisson distribution with a parameter of Therefore, the average number of surviving targets is 20.

[0182] The multi-target measurement model parameters are: the detection probability is p D = 0.9, the measurement matrix and the measurement noise covariance matrix are

[0183]

[0184] The clutter obeys a strength of κ(z) = λ · uA The Poisson distribution of (z), where u A (z) represents the average density of clutter in region A, with an average of λ = 10 clutter particles generated per scan.

[0185] 2. Simulation Result Analysis

[0186] Figure 3 , Figure 4 and Figure 5 Schematic diagrams of state estimation at time k = 1:88 are given for the classical CD-PHD, CD-CPHD and CD-PMBM methods respectively. Figure 3 , Figure 4 and Figure 5 The dashed line represents the actual trajectory throughout the monitoring process, and "+" indicates sensor measurement. "*" indicates the target's actual position at k=1:88, and "o" indicates the estimated target position at k=1:88. This indicates that the target position is estimated at time k=88. Figure 6 This is a schematic diagram of the real-time tracking results of the NUS-TPMB method. Figure 6 In the figure, "o-" indicates the estimated track at time k=88. (Compare) Figure 3 , Figure 4 , Figure 5 and Figure 6 It is evident that, compared to the classic CD-PHD, CD-CPHD, and CD-PMBM methods, the multi-target tracking method based on Poisson-Dobernoli (NUS-TPMB) proposed in this application can not only generate tracks, but also achieve higher track state estimation accuracy and lower false negative rate.

[0187] Figure 7 A performance comparison diagram of CD-PHD, CD-CPHD, CD-PMBM, and NUS-TPMB methods is presented. The evaluation is based on the root mean square (RMS) metric of the generalized optimal sub-pattern assignment (GOSPA) method, and averaged using the results of 100 Monte Carlo experiments. Figure 7 It can be seen that the NUS-TPMB algorithm has the lowest GOSPA error, indicating that the state estimation accuracy of this method is the highest.

[0188] Figure 8 A comparative diagram of the target potential estimation results of the CD-PHD, CD-CPHD, CD-PMBM, and NUS-TPMB methods is presented. Figure 8 It can be seen that the NUS-TPMB algorithm has the highest potential estimation accuracy, especially when the number of targets is large.

[0189] Figure 9 The position error comparison of the CD-PHD, CD-CPHD, CD-PMBM and NUS-TPMB methods is given. Figure 9 It can be seen that the position error of the NUS-TPMB algorithm is the smallest.

[0190] Figure 10 The missed target cost result comparison diagram and the error target cost result comparison diagram of the CD-PHD, CD-CPHD, CD-PMBM and NUS-TPMB methods are given respectively. 11 Figure 10 11 It can be seen that compared with the CD-PHD, CD-CPHD and CD-PMBM methods, the NUS-TPMB method not only has a low missed detection rate, but also has a low error rate. The running time of the CD-PHD, CD-CPHD, CD-PMBM and NUS-TPMB methods is 0.72s, 2.67s, 11.21s and 2.46s respectively. It can be seen that the NUS-TPMB method takes less time than the CD-CPHD and CD-PMBM methods, which shows that the application has good performance and high computational efficiency, and has good engineering application prospect.

[0191] It should be understood that although each step in the flowchart of Figure 1 is shown in sequence following the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise explicitly stated herein, there is no strict order limitation for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps in may include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or sub-steps or stages of other steps.

[0192] In one embodiment, a multi-target tracking device based on track Poisson multiple Bernoulli is provided, comprising:

[0193] A model construction module is configured to construct a continuous-time multi-target motion model, wherein the continuous-time multi-target motion model comprises a target generation model, a target extinction model and a target motion model.

[0194] A discretization processing module is configured to perform discretization processing on the continuous-time multi-target motion model by a Gaussian moment approximation method to obtain a discretized new target probability density and a single-target track state transition density.

[0195] ​a prediction updating module, configured to obtain a track Poisson multi-Bernoulli density of the multiple targets at a previous time step according to the discretized target birth probability density and the single-target track state transition density under the condition of non-uniform sampling, and sequentially perform prediction and updating on the track Poisson multi-Bernoulli density at a current time step according to the track Poisson multi-Bernoulli density at the previous time step, to obtain an updated value of the track Poisson multi-Bernoulli density at the current time step; wherein the updated value of the track Poisson multi-Bernoulli density at the current time step comprises an updated value of a Poisson component and an updated value of a Bernoulli component;

[0196] a track estimation module, configured to construct a track Poisson multi-Bernoulli posterior density according to the updated value of the Poisson component and the updated value of the Bernoulli component, and estimate the multiple-target track at the current time step according to the track Poisson multi-Bernoulli posterior density and a preset threshold, to obtain a multiple-target track estimation at the current time step.

[0197] Specific limitations of the multi-target tracking device based on track Poisson multi-Bernoulli can be seen from the limitations of the multi-target tracking method based on track Poisson multi-Bernoulli in the foregoing, which will not be repeated here. Each module in the multi-target tracking device based on track Poisson multi-Bernoulli can be realized by software, hardware, and a combination thereof, in whole or in part. Each module can be embedded in or independent of a processor in a computer device in hardware form, or can be stored in a memory in a computer device in software form, so as to be called and executed by a processor to perform the operations corresponding to each module.

[0198] In one embodiment, a computer device is provided, which can be a terminal, and an internal structure diagram thereof can be as shown in Figure 12 The computer device includes a processor, a memory, a network interface, a display screen, and an input device connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The network interface of the computer device is configured to communicate with external terminals through network connection. The computer program is executed by the processor to implement a multi-target tracking method based on track Poisson multi-Bernoulli. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or can be a key, trackball, or touchpad arranged on the shell of the computer device, or can be an external keyboard, touchpad, or mouse, etc.

[0199] Those skilled in the art can understand that Figure 12The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0200] In one embodiment, a computer device is provided, comprising a memory and a processor, the memory storing a computer program, and the processor implementing the following steps when executing the computer program:

[0201] constructing a continuous-time multi-target motion model, wherein the continuous-time multi-target motion model comprises a target generation model, a target death model and a target motion model;

[0202] discretizing the continuous-time multi-target motion model by a Gaussian moment approximation method to obtain a discretized new target probability density and a single-target track state transition density;

[0203] under non-uniform sampling conditions, obtaining a track Poisson multi-Bernoulli density of the multi-target at a previous time step according to the discretized new target probability density and the single-target track state transition density, and sequentially performing prediction and update on a track Poisson multi-Bernoulli density of a current time step according to the track Poisson multi-Bernoulli density of the previous time step to obtain an updated value of the track Poisson multi-Bernoulli density of the current time step; wherein the updated value of the track Poisson multi-Bernoulli density of the current time step comprises an updated value of a Poisson component and an updated value of a Bernoulli component;

[0204] combining and constructing a track Poisson multi-Bernoulli posterior density according to the updated value of the Poisson component and the updated value of the Bernoulli component, and estimating the multi-target track of the current time step according to the track Poisson multi-Bernoulli posterior density and a pre-set threshold to obtain a multi-target track estimation of the current time step.

[0205] The technical features of the above embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not contradict, they should be considered within the scope of the present application.

[0206] The above-described embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the scope of the present application. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A track-while-scan Poisson multi-Bernoulli based multi-target tracking method, characterized in that, The method comprises: constructing a continuous-time multi-target motion model, wherein the continuous-time multi-target motion model comprises a target generation model, a target extinction model and a target motion model; discretizing the continuous-time multi-target motion model by a Gaussian moment approximation method to obtain a discretized new target probability density and a single-target track state transition density; under non-uniform sampling conditions, obtaining a track Poisson multi-Bernoulli density of the multi-target at a previous time step according to the discretized new target probability density and the single-target track state transition density, and sequentially performing prediction and update on a track Poisson multi-Bernoulli density of a current time step according to the track Poisson multi-Bernoulli density of the previous time step to obtain an updated value of the track Poisson multi-Bernoulli density of the current time step; wherein the updated value of the track Poisson multi-Bernoulli density of the current time step comprises a Poisson component updated value and a Bernoulli component updated value; combining the Poisson component updated value and the Bernoulli component updated value to construct a track Poisson multi-Bernoulli posterior density, and estimating a multi-target track of the current time step according to the track Poisson multi-Bernoulli posterior density and a pre-set threshold to obtain a multi-target track estimate of the current time step; The target generation model is subject to a Poisson random process with a Poisson intensity of The target state in the target generation model is subject to a Gaussian distribution with a mean of and a covariance matrix of wherein , represent the average position and average velocity, respectively, and represent the position covariance matrix and velocity covariance matrix, respectively, represents the position and velocity covariance matrix, and the superscript T represents the matrix transpose. the target extinction model is expressed as ； wherein, represents the lifetime of the target are independent and follow an exponential distribution with a rate parameter the target motion model is constructed based on a Wiener velocity model and is expressed as ； ； in, Represents a multi-objective state. express The differential, Represents the state space of a single objective. Indicates dimension as The first matrix, Indicates dimension as The second matrix, This represents the dimension of a single-objective state. The diffusion matrix is ​​represented as Wiener process, express 3D real space, Representing dimension, Indicates the parameters of Wiener's velocity model. The dimension is A matrix of all zeros The dimension is The identity matrix; the discretizing the continuous-time multi-target motion model by the Gaussian moment approximation method to obtain the discretized new target probability density is expressed as ； wherein, denotes the distribution of target arrivals within the sampling time step interval, denotes the discretized Poisson intensity, denotes the Gaussian component index, denotes the number of Gaussian components, denotes the Gaussian components, denotes the discretized track start time, denotes the discretized weight, denotes the discretized mean, is the state variable of the q th arriving target at time step k , and denotes the discretized covariance matrix.​ 2. The method of claim 1, wherein, the discretizing the continuous-time multi-target motion model by the Gaussian moment approximation method to obtain the single-target track state transition density is expressed as ； wherein, denotes k a set of tracks present at a time step any one of the track variables, denotes the track start time, denotes the track duration, denotes the state sequence of the track, denotes the single-target state transition density, , , denote the Kronecker function, the transition matrix and the transition covariance, respectively.

3. The method of claim 1, wherein, under non-uniform sampling conditions, obtaining a track Poisson multi-Bernoulli density of the multi-target at a previous time step according to the discretized new target probability density and the single-target track state transition density, comprising: Under the condition of non-uniform sampling, according to the discretized new-born target probability density and the single-target track state transition density, a track Poisson multi-Bernoulli density of the multi-target at the time step is obtained, expressed as the track Poisson multi-Bernoulli density of the multi-target at the time step is obtained, expressed as ； wherein, denotes the time step, the time step, denotes the total track set, denotes the Bernoulli track set, denotes the number of Bernoulli components, denotes the Poisson component track set, denotes the Poisson component in the track Poisson multi-Bernoulli density at time step, denotes the Bernoulli component in the track Poisson multi-Bernoulli density at time step.

4. The method of claim 3, wherein, sequentially performing prediction and update on a track Poisson multi-Bernoulli density of a current time step according to the track Poisson multi-Bernoulli density of the previous time step to obtain an updated value of the track Poisson multi-Bernoulli density of the current time step, comprising: According to the predicted value of the track Poisson multinomial density at the time step, denoted as the predicted value of the track Poisson multinomial density at the time step, denoted as the predicted value of the track Poisson multinomial density at the time step, denoted as ； wherein, denotes Bernoulli track set at time step denotes number of Bernoulli components at time step denotes Poisson component prediction at time step denotes Bernoulli component prediction at time step respectively, are updated to obtain the Poisson component prediction value at the time step and the Bernoulli component prediction value at the time step are updated to obtain the Poisson component updated value at the time step and the Bernoulli component updated value at the time step.

5. The method of claim 1, wherein, combining the Poisson component updated value and the Bernoulli component updated value to construct a track Poisson multi-Bernoulli posterior density, and estimating a multi-target track of the current time step according to the track Poisson multi-Bernoulli posterior density and a pre-set threshold to obtain a multi-target track estimate of the current time step, comprising: combining the Poisson component update value and the Bernoulli component update value to construct a track Poisson multi-Bernoulli posterior density, and determining a track based on the track Poisson multi-Bernoulli posterior density and a pre-set threshold to k estimate the multi-target track at the time step, to obtain k the track estimate at the time step is wherein, denotes the track start time of the th Bernoulli component, denotes the k update value of the mean of the th Bernoulli component at the time step, denotes the k update value of the existence probability of the th Bernoulli component at the time step.

6. A multi-target tracking device based on track Poisson multi-Bernoulli based on the method of any one of claims 1 to 5, characterized in that, the device comprises: a model construction module configured to construct a continuous-time multi-target motion model, wherein the continuous-time multi-target motion model comprises a target generation model, a target extinction model and a target motion model; a discretization processing module configured to discretize the continuous-time multi-target motion model by a Gaussian moment approximation method to obtain a discretized new target probability density and a single-target track state transition density; under non-uniform sampling conditions, obtaining a track Poisson multi-Bernoulli density of the multi-target at a previous time step according to the discretized new target probability density and the single-target track state transition density, and sequentially performing prediction and update on a track Poisson multi-Bernoulli density of a current time step according to the track Poisson multi-Bernoulli density of the previous time step to obtain an updated value of the track Poisson multi-Bernoulli density of the current time step; wherein the updated value of the track Poisson multi-Bernoulli density of the current time step comprises a Poisson component updated value and a Bernoulli component updated value; a prediction updating module, configured to obtain a track Poisson multi-Bernoulli density of the multiple targets at a previous time step according to the discretized new target probability density and a single-target track state transition density under a non-uniform sampling condition, and sequentially perform prediction and updating on a track Poisson multi-Bernoulli density at a current time step according to the track Poisson multi-Bernoulli density at the previous time step to obtain an updated value of the track Poisson multi-Bernoulli density at the current time step; wherein the updated value of the track Poisson multi-Bernoulli density at the current time step comprises an updated value of a Poisson component and an updated value of a Bernoulli component; a track estimation module, configured to construct a track Poisson multi-Bernoulli posterior density according to a combination of the updated value of the Poisson component and the updated value of the Bernoulli component, and estimate a multiple-target track at the current time step according to the track Poisson multi-Bernoulli posterior density and a preset threshold to obtain a multiple-target track estimation at the current time step. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-6 when the computer program is executed by the processor. The processor implements the steps of the method in any one of claims 1 to 5 when executing the computer program.

Images