CN106772357A - AI PHD wave filters under signal to noise ratio unknown condition - Google Patents

Abstract

The invention discloses the AI PHD wave filters under a kind of signal to noise ratio unknown condition, belong to radar data process field.Existed based on the multi-object tracking method that PHD is filtered and do not make full use of target measurement information, it is impossible to the problems such as estimating target RCS, and be difficult in adapt to dense clutter environment.AI PHD wave filters under signal to noise ratio unknown condition proposed by the present invention are based on solution problems.The present invention by being filtered in PHD on the basis of combining target amplitude information, the characteristics of being easy to extension using particle filter, the signal to noise ratio variable relevant with target RSC is introduced in dbjective state vector, target RCS is estimated while target number and Target state estimator, its amount of calculation is with the increase linear increase for measuring number, it is specially adapted to the multiple target tracking under dense clutter environment, the limitation of general PHD wave filters is overcome, therefore with stronger engineering application value and promotion prospect.

Description

AI-PHD wave filters under signal to noise ratio unknown condition

Technical field

The present invention relates to a kind of radar data processing method, more particularly to a kind of multiple target tracking filtering method is adapted to In tracking of the radar to multiple target under glint model.

Background technology

Multiple target tracking is one of difficulties of current radar target tracking domain.In the case of dense clutter, one Aspect, because the appearing and subsiding of target has randomness, the number of target is often time-varying and unknown；On the other hand, receive The number that the interference of clutter and noise, data correlation and detection have uncertainty, measurement also has randomness.In such case Lower utilization radar is tracked to multiple target, it is necessary to estimate the uncertain each target of target number from the measurement of number time-varying Multiple target tracking under state, therefore dense clutter environment is particularly difficult.Various measurement informations how are made full use of, is realized intensive Effective tracking of the radar to multiple target under clutter environment, the detection tracking performance to improving radar is significant.Current Multiple target tracking algorithm mainly has Joint Probabilistic Data Association (JPDA) to filter, multiple hypotheis tracking (MHT) filtering and probability are false If density (PHD) filtering etc., wherein the multi-object tracking method based on PHD is low due to algorithm complex, amount of calculation is individual with measuring Number linear increase, the advantages of can estimating target number and dbjective state simultaneously, is subject to wide in multiple target tracking field General concern.The method is mainly realized by following steps：

(1) filter initialization；

(2) particle collection prediction；

(3) particle collection updates；

(4) target number and Target state estimator.

Multiple target tracking based on PHD filtering has two defects：(1) without the amplitude information using target, caused With the loss of measurement information, it is difficult to adapt to dense clutter environment；(2) RCS of target can not be estimated.

The content of the invention

The purpose of the present invention is to propose to the AI-PHD wave filters under a kind of signal to noise ratio unknown condition, general PHD filters are solved Ripple does not make full use of target measurement information, it is impossible to which target RCS is estimated, and is difficult in adapt to dense clutter environment etc. and asks Topic.

The technical scheme of the AI-PHD wave filters under signal to noise ratio unknown condition proposed by the present invention is comprised the following steps：

Step 1：Initialization of variable

(1) K is total simulation time, and T is spaced for radar sampling；

(2)γ₀For there is initial number, L in target₀To represent the population required for a target, J₀During for search fresh target Assign the population of each measurement；

(3)SNR_minIt is the possible minimum signal to noise ratio of target, SNR_maxIt is the possible maximum signal to noise ratio of target,It is false-alarm Probability, τ is false-alarm probabilityCorresponding detection threshold；

(4)γ₀X there is initial distribution, κ for target in ()_kZ () is distributed for clutter；

(5)It is expansion process noise covariance,For extension measures noise covariance；

(6)C₀It is the binding occurrence of clustering class；

Step 2：K=0 is made, the initialization of device is filtered, primary collection is obtainedSpecially to any i ∈{1,2,…,L₀}

(1) initial distribution γ is occurred according to target₀X () generates particleWhereinComprising target Positional informationAnd velocity informationSymbol T represents transposition；

(2) in interval [SNR_min,SNR_max] on according to being uniformly distributed random generation target signal to noise ratioAnd according to

It is calculated the signal-tonoise information of target

(3) makeIt is rightIt is augmented, is obtained new particleAnd assign the particle weightsStep Rapid 3：K=k+1 is made, the radar measurement at k moment is obtained

The signal that radar is received carries out A/D conversion, and the extension for obtaining current time measures collectionData handling system is sent, whereinFor j-th extension that k moment radar is obtained Measure,Range information comprising targetAnd azimuth information It is the amplitude information of target, N_kIt is the k moment Measurement number；

Step 4：Generation prediction particle collection

To any i ∈ 1,2 ..., L_k-1, according to state transition equation to particleIt is predicted, the grain predicted Son

And assign the particle weightsObtain predicting particle collectionWhereinIt is zero-mean White Gaussian noise, its covariance is

Step 5：Generation search fresh target particle collection

(1) to any j ∈ 1,2 ..., N_k, according to measurementWith error in measurement covariance R_kSampling particleThen In interval [SNR_min,SNR_max] on according to being uniformly distributed random generation target signal to noise ratioAnd it is calculated target signal Compare informationFinally makeAnd assign the particle weights

Wherein i=1,2 ..., J₀, measuredSearch particle collection

(2) will the corresponding search particle collection of current time all measurementsIt is merged into a search Fresh target particle collectionWherein J_k=J₀×N_kTo search for the total number of particles of fresh target；

Step 6：Particle collection weight updates

(1) particle collection will be predictedWith search fresh target particle collectionMerging obtains new Particle collection

(2) to any i ∈ 1,2 ..., L_k-1+J_kAnd any j ∈ 1,2 ..., N_k, calculate particleAnd measurementIt Between statistical distance

Wherein

For prediction is measured, (x_s,y_s) it is the coordinate of radar, ifMake particleAnd measurementBetween Likelihood scoreOtherwise

Wherein

(3) to any j ∈ 1,2 ..., N_k, calculate and measureWith particle collectionLikelihood score

(4) to any i ∈ 1,2 ..., L_k-1+J_k, calculate particle weights

Wherein

And

Step 7：Target number is estimated and particle collection resampling

(1) calculate all particles weight and

And take withImmediate integer obtains the estimation of target number

(2) total number of particles needed for calculating current time

(3) in interval [0,1] according to being uniformly distributed generation L_kIndividual random number

(4) weight to particle collection is normalized, and obtains normalized particle weights

(5) calculate particle weights accumulation and

(6) to any j ∈ 1,2 ..., L_k, if there is i ∈ { 1,2 ..., L_k-1+J_kSo that

Then make particle

And assign the particle weightsObtain new particle collection

Step 8：Particle collection point group

(1) to anyMake population numberWith

ObtainIndividual group center

(2) to any i ∈ 1,2 ..., L_k, calculate particleAnd group centerDistance

Then make

By particleIt is divided into j-th group；

(3) to anyOrder

ObtainIndividual new group centerThen calculate new, old group center distance and

OrderIf D_kMore than binding occurrence C₀Turn (2), otherwise go to step 9；

Step 9：Multiple target state and signal-to-noise ratio (SNR) estimation

To anyTake group centerThe 1st dimension to the 4th dimension obtain j-th state estimation of targetTake group centerThe 5th dimension obtain j-th signal-to-noise ratio (SNR) estimation of targetAnd pushed away according to the relational expression of the signal to noise ratio of RCS Calculate the RCS of target_k,j；

Step 10：3~step 9 of repeat step, until radar switching-off.

Compared with background technology, the beneficial effect of the AI-PHD wave filters under signal to noise ratio unknown condition proposed by the present invention is said It is bright：

(1) solve the problems, such as that general PHD wave filters do not make full use of measurement information, be specially adapted to dense clutter Multiple target tracking under environment and in the case of bobbing, effectively increases detection tracking performance of the radar to multiple target； (2) target RCS can be estimated while target number and state estimation is carried out.

Brief description of the drawings

Fig. 1 is the overall flow figure of the AI-PHD wave filters under signal to noise ratio unknown condition of the invention, each symbol in accompanying drawing Implication it is identical with the implication of Summary respective symbol；

Fig. 2 is that the multiple target state that general PHD filtering methods are estimated is contrasted with actual value and measuring value, wherein scheming A () and figure (b) are respectively the contrast of x coordinate value and y-coordinate value contrast；

Fig. 3 be the multiple target state estimated of the AI-PHD wave filters in the embodiment of the present invention under signal to noise ratio unknown condition with it is true Real-valued and measuring value contrast, wherein figure (a) and figure (b) are respectively the contrast of x coordinate value and y-coordinate value contrast；

Fig. 4 is the target signal to noise ratio that the AI-PHD wave filters in the embodiment of the present invention under signal to noise ratio unknown condition are estimated and true Real target signal to noise ratio contrast.

Specific embodiment

The AI-PHD wave filters under unknown condition of the invention are described in detail below in conjunction with the accompanying drawings.

Without loss of generality, it is assumed that at any time, target is all in the two-dimensional observation of S=[- 200,200] × [- 200,200] Moved in region, and target can be in the random appearing and subsiding in the region, total simulation time is K=50s, sampling interval T= 1s；Initially there is obedience Poisson model in target, its density function γ_k(x)=0.2N (x | x₀,Q_b), N (| x₀,Q_b) represent average It is x₀, covariance is Q_bGaussian Profile, wherein x₀=[0 2 0-2]^TAnd Q_b=diag ([10 5 10 5]), target may Minimum signal to noise ratio snr_min=10dB, possible maximum signal to noise ratio SNR_max=40dB, its echo obeys Swerling forcers Type；Radar be located at point (0, -100), it is possible to provide target apart from R_k, azimuth angle theta_kWith amplitude information a_k, the amount at distance and bearing angle Survey noise criteria difference be respectively 2 and 0.05, measure noise it is separate with process noise, clutter be evenly distributed on [0,2 π] × In the observation space of [0,200], and the average noise points per frame are μ=1000, radar false alarm probabilityIt is corresponding Detection threshold τ=2.146；Population N needed for representing a target₀=800, assign each measurement during search fresh target Population J₀=20；The binding occurrence of clustering class is 0.01.Its step is as shown in Figure 1.

Step 1：Initialization of variable is carried out according to above simulated conditions

(1) total simulation time K=50s, radar sampling interval T=1s；

(2) there is initial number γ in target₀=0.2, the population L required for representing a target₀=800, search for fresh target When assign each measurement population J₀=20；

(3) the possible minimum signal to noise ratio snr of target_min=10dB, the possible maximum signal to noise ratio SNR of target_max=40dB, it is empty Alarm probabilityCorresponding detection threshold τ=2.146；

(4) there is initial distribution γ in target_kX () and clutter are distributed κ_kZ () is respectively

γ_k(x)=0.2N (x | x₀,Q_b)

(5) expansion process noise covarianceAnd extension measures error covarianceRespectively

(6) the binding occurrence C of clustering class₀=0.01；

Step 2：Method as described in Summary step 2 is filtered device initialization；

Step 3：Method as described in Summary step 3 obtains the measurement at current time；

Step 4：Method generation prediction particle collection as described in Summary step 4；

Step 5：Method generation search fresh target particle collection as described in Summary step 5；

Step 6：Method as described in Summary step 6 carries out the renewal of particle collection weight；

Step 7：Method as described in Summary step 7 carries out target number and estimates and particle collection resampling；

Step 8：Method as described in Summary step 8 carries out particle collection point group；

Step 9：Method as described in Summary step 9 carries out multiple target state and signal-to-noise ratio (SNR) estimation；

Step 10：Circulation performs Summary step 3~step 9, until radar switching-off.

The average noise points per frame are μ=1000, radar false alarm probability in the present embodimentAverage every frame Clutter number λ=μ P_FA=100, belong to the situation of dense clutter, general PHD wave filters can hardly realize to multiple target with Track (see accompanying drawing 2), and pass through the amplitude information of effective combining target, the AI-PHD wave filters under unknown condition of the invention can be with The effective tracking (see accompanying drawing 3) to multiple target is realized, the defect instant invention overcomes general PHD wave filters is illustrated；Additionally, this hair AI-PHD wave filters under bright unknown condition can enter while target number and state estimation is carried out to target signal to noise ratio Row is estimated (see accompanying drawing 4), and then can realize the estimation to target RCS.

Claims (1)

1. AI-PHD wave filters under signal to noise ratio unknown condition, its feature is comprised the following steps：

Step 1：Initialization of variable

(1) K is total simulation time, and T is spaced for radar sampling；

(2)γ₀For there is initial number, L in target₀To represent the population required for a target, J₀Assigned during for search fresh target Each population for measuring；

(3)SNR_minIt is the possible minimum signal to noise ratio of target, SNR_maxIt is the possible maximum signal to noise ratio of target,It is false-alarm probability, τ It is false-alarm probabilityCorresponding detection threshold；

(4)γ₀X there is initial distribution, κ for target in ()_kZ () is distributed for clutter；

(5)It is expansion process noise covariance,For extension measures noise covariance；

(6)C₀It is the binding occurrence of clustering class；

Step 2：K=0 is made, the initialization of device is filtered, primary collection is obtainedSpecially to any i ∈ 1, 2,…,L₀}

(1) initial distribution γ is occurred according to target₀X () generates particleWhereinPosition comprising target InformationAnd velocity informationSymbol T represents transposition；

(2) in interval [SNR_min,SNR_max] on according to being uniformly distributed random generation target signal to noise ratioAnd according to

$d_0^i={10}^{{\mathrm{SNR}}_0^i/10}$

It is calculated the signal-tonoise information of target

(3) makeIt is rightIt is augmented, is obtained new particleAnd assign the particle weights

Step 3：K=k+1 is made, the radar measurement at k moment is obtained

The signal that radar is received carries out A/D conversion, and the extension for obtaining current time measures collectionData handling system is sent, whereinFor j-th extension that k moment radar is obtained Measure,Range information comprising targetAnd azimuth information It is the amplitude information of target, N_kIt is the k moment Measurement number；

Step 4：Generation prediction particle collection

To any i ∈ 1,2 ..., L_k-1, according to state transition equation to particleIt is predicted, the particle predicted

${\tilde{x}}_{k|k-1}^i=\left[\begin{array}{ccccc}1& T& 0& 0& 0\\ 0& 1& 0& 0& 0\\ 0& 0& 1& T& 0\\ 0& 0& 0& 1& 0\\ 0& 0& 0& 0& 1\end{array}\right]{\tilde{x}}_{k-1}^i+\left[\begin{array}{ccc}{T}^2/2& 0& 0\\ T& 0& 0\\ 0& T^2/2& 0\\ 0& T& 0\\ 0& 0& 1\end{array}\right]{\tilde{v}}_k$

And assign the particle weightsObtain predicting particle collectionWhereinIt is the Gauss of zero-mean White noise, its covariance is

Step 5：Generation search fresh target particle collection

(1) to any j ∈ 1,2 ..., N_k, according to measurementWith error in measurement covariance R_kSampling particleThen in area Between [SNR_min,SNR_max] on according to being uniformly distributed random generation target signal to noise ratioAnd it is calculated target signal to noise ratio letter BreathFinally makeAnd assign the particle weights

${\tilde{w}}_{k|k-1}^i={\mathrm{\γ}}_0/\left(N_k{J}_0\right)$

Wherein i=1,2 ..., J₀, measuredSearch particle collection

(2) will the corresponding search particle collection of current time all measurementsIt is merged into a new mesh of search Mark particle collectionWherein J_k=J₀×N_kTo search for the total number of particles of fresh target；

Step 6：Particle collection weight updates

(1) particle collection will be predictedWith search fresh target particle collectionMerging obtains new particle Collection

(2) to any i ∈ 1,2 ..., L_k-1+J_kAnd any j ∈ 1,2 ..., N_k, calculate particleAnd measurementBetween Statistical distance

$D_k^{ij}(x_{k|k-1}^i,z_k^j)={(z_k^j-z_{k|k-1}^i)}^T{R}_k^{-1}(z_k^j-z_{k|k-1}^i)$

Wherein

$z_{k|k-1}^i=\left[\begin{array}{c}{\mathrm{\ρ}}_{k|k-1}^i\\ {\mathrm{\θ}}_{k|k-1}^i\end{array}\right]=\left[\begin{array}{c}\sqrt{{\left(x_{k|k-1}^i-x_s\right)}^2+{\left(y_{k|k-1}^i-y_s\right)}^2}\\ \arctan \frac{x_{k|k-1}^i-x_s}{y_{k|k-1}^i-y_s}\end{array}\right]$

For prediction is measured, (x_s,y_s) it is the coordinate of radar, ifMake particleAnd measurementBetween seemingly So spendOtherwise

${\widetilde{\mathrm{\ψ}}}_{ij}({\tilde{x}}_{k|k-1}^i,{\tilde{z}}_k^j)=g_k\left(z_k^j\right|x_{k|k-1}^i)g_a^{\mathrm{\τ}}\left(a_k^j\right|d_{k|k-1}^i)$

Wherein

$g_k\left(z_k^j\right|x_{k|k-1}^i)=\frac{1}{2\mathrm{\π}\sqrt{R_k}}\exp \{-D_k^{ij}(x_{k|k-1}^i,z_k^j)/2\}$

$g_a^{\mathrm{\τ}}\left(a_k^j\right|d_{k|k-1}^i)=\frac{a_k^j}{1+d_{k|k-1}^i}\exp \left(-\frac{a_k^j-{\mathrm{\τ}}^2}{2(1+d_{k|k-1}^i)}\right)$

(3) to any j ∈ 1,2 ..., N_k, calculate and measureWith particle collectionLikelihood score

${\tilde{C}}_k^j\left({\tilde{z}}_k^j\right)=\underset{i=1}{\overset{L_{k-1}+J_k}{\Σ}}{\widetilde{\mathrm{\ψ}}}_{ij}({\tilde{x}}_{k|k-1}^i,{\tilde{z}}_k^j){\tilde{w}}_{k|k-1}^i$

(4) to any i ∈ 1,2 ..., L_k-1+J_k, calculate particle weights

${\tilde{w}}_{k|k}^i=\[\tilde{v}\left({\tilde{x}}_{k|k-1}^i\right)+\underset{j=1}{\overset{N_k}{\Σ}}\frac{{\widetilde{\mathrm{\ψ}}}_{ij}({\tilde{x}}_{k|k-1}^i,{\tilde{z}}_k^j)}{{\widetilde{\mathrm{\κ}}}_k\left({\tilde{z}}_k^j\right)+{\tilde{C}}_k^j\left({\tilde{z}}_k^j\right)}\]\×{\tilde{w}}_{k|k-1}^i$

Wherein

$\tilde{v}\left({\tilde{x}}_{k|k-1}^i\right)=1-\exp \left(-\frac{{\mathrm{\τ}}^2}{2(1+d_{k|k-1}^i)}\right)$

And

${\widetilde{\mathrm{\κ}}}_k\left({\tilde{z}}_k^j\right)={\mathrm{\κ}}_k\left(z_k^j\right)c_a^{\mathrm{\τ}}\left(a_k^j\right)$

$c_a^{\mathrm{\τ}}\left(a_k^j\right)=\frac{1}{P_{FA}^{\mathrm{\τ}}}{a}_k^j\exp \left\{-\frac{{\left(a_k^j\right)}^2}{2}\right\}$

Step 7：Target number is estimated and particle collection resampling

(1) calculate all particles weight and

${\hat{N}}_{k|k}=\underset{i=1}{\overset{L_{k-1}+J_k}{\Σ}}{\tilde{w}}_{k|k}^i$

And take withImmediate integer obtains the estimation of target number

(2) total number of particles needed for calculating current time

$L_k={\hat{N}}_k\×L_0$

(3) in interval [0,1] according to being uniformly distributed generation L_kIndividual random numberJ=1,2 ..., L_k；

(4) weight to particle collection is normalized, and obtains normalized particle weights

$w_k^i=\frac{{\tilde{w}}_{k|k}^i}{{\mathrm{\Σ}}_{i=1}^{L_{k-1}+J_k}{\tilde{w}}_{k|k}^i},i=1,2,\mathrm{...},L_{k-1}+J_k$

(5) calculate particle weights accumulation and

${\mathrm{CSW}}_k^i=\underset{j=1}{\overset{i}{\Σ}}{w}_k^i,i=1,2,\mathrm{...},L_{k-1}+J_k$

(6) to any j ∈ 1,2 ..., L_k, if there is i ∈ { 1,2 ..., L_k-1+J_kSo that

${\mathrm{CSW}}_k^{i-1}\≤u_k^j\≤{\mathrm{CSW}}_k^i$

Then make particle

${\tilde{x}}_k^j={\tilde{x}}_{k|k-1}^i$

And assign the particle weightsObtain new particle collection

Step 8：Particle collection point group

(1) to anyMake population numberWith

${\tilde{y}}_k^j={\tilde{x}}_{k|k-1}^i,i=\frac{L_0}{2}(2j-1)$

ObtainIndividual group center

(2) to any i ∈ 1,2 ..., L_k, calculate particleAnd group centerDistance

$D_k^{ij}\left({\tilde{x}}_k^i,{\tilde{y}}_k^j\right)=\sqrt{{\left({\tilde{x}}_k^i\[1\]-{\tilde{y}}_k^j\[1\]\right)}^2+{\left({\tilde{x}}_k^i\[3\]-{\tilde{y}}_k^j\[3\]\right)}^2},j=1,2,\mathrm{...},{\hat{N}}_k$

Then make

$j=\arg \underset{j\∈\{1,2,\mathrm{...},{\hat{N}}_k\}}{min}{D}_k^{ij}({\tilde{x}}_k^i,{\tilde{y}}_k^j)$

$N_k^j=N_k^j+1$

${\tilde{x}}_{k,j}^{N_j}={\tilde{x}}_k^i$

By particleIt is divided into j-th group；

(3) to anyOrder

${\tilde{y}}_k^{\′j}=\frac{1}{N_k^j}\underset{i=1}{\overset{N_k^j}{\Σ}}{\tilde{x}}_{k,j}^i,i=1,2,\mathrm{...},N_k^j$

ObtainIndividual new group centerThen calculate new, old group center distance and

$D_k=\underset{j=1}{\overset{{\hat{N}}_k}{\Σ}}\sqrt{{({\tilde{y}}_k^j\[1\]-{\tilde{y}}_k^{\′j}\[1\])}^2+{({\tilde{y}}_k^j\[3\]-{\tilde{y}}_k^{\′j}\[3\])}^2}$

OrderIf D_kMore than binding occurrence C₀Turn (2), otherwise go to step 9；

Step 9：Multiple target state and signal-to-noise ratio (SNR) estimation

To anyTake group centerThe 1st dimension to the 4th dimension obtain j-th state estimation of targetTake Group centerThe 5th dimension obtain j-th signal-to-noise ratio (SNR) estimation of targetAnd target is calculated according to the relational expression of the signal to noise ratio of RCS RCS_k,j；

Step 10：3~step 9 of repeat step, until radar switching-off.