CN104501812A - Filtering algorithm based on self-adaptive new target strength - Google Patents

Abstract

The invention discloses a PHD (probability hypothesis density) filtering algorithm based on self-adaptive new target strength. The PHD filtering algorithm is implemented through the following steps: carrying out track initiation and track confirmation; constructing a strength function of a new target random set according to confirmed track initiation position information; merging a new target strength function constructed by current frames with a known new target strength function at previous moment so as to produce an adaptive new target strength function of the current frames; starting a PHD filter by using the adaptive new target strength function so as to complete the prediction and updating of the PHD filter; and carrying out multiple target state extraction. According to the Bayesian multiple target tracking method based on the filtering algorithm, a new target strength function is constructed only according to measurement information completely without using priori information. The algorithm can be used for solving the problem that an original PHD filter lacks a new target tracking function. The algorithm has the advantages that a new target of which the position is unknown can be tracked well, so that the application range is enlarged, the robustness of the PHD filter is improved, and the like.

Description

Based on the filtering algorithm of the newborn target strength of self-adaptation

Technical field

What the present invention relates to is a kind of algorithm of target tracking domain, specifically based on the probability hypothesis density filtering algorithm of the newborn target strength of self-adaptation.

Technical background

Multiple target tracking problem has in military and civilian to be applied very widely, and as air-borne early warning, air attack (Multi-target Attacking) etc. at military aspect, civilian aspect comprises the multiple target trackings such as air traffic control.Application militarily receives various countries and extensively payes attention to.The object of multiple target tracking (MTT) estimates multiobject number and state from the measurement set including clutter.Tradition multiple target tracking algorithm is based upon on the basis of data correlation (Data Association) technology, and being intended to is the tracking problem of multiple single goal the decoupling zero of multiple target tracking problem.But the derivation algorithm of data association technique exists shot array, thus limit the practical application of such algorithm.In recent years, the multi-object tracking method based on stochastic finite collection achieves quantum jump, and obtains and pay close attention to widely.

Track algorithm based on stochastic finite collection is described as multi-objective Bayesian filtering multiple target tracking problem, and this multi-objective Bayesian wave filter propagates the Posterior probability distribution of multiple goal state by prediction and correction two steps.Wherein, multiple goal Posterior distrbutionp function is set up according to finite set statistics (FISST).Be difficult to calculate because this multiple goal Posterior distrbutionp function comprises the Infinite-dimensional of multiple goal state.For this reason, Mahler proposes probability hypothesis density (Probability Hypothesis Density, PHD) wave filter, and the first moment of multiple goal Posterior distrbutionp function only propagated by this wave filter, i.e. PHD.PHD wave filter avoids the data association technique in traditional multiple target tracking algorithm, and is easy to process dense clutter, but there is PHD wave filter number of targets estimates inaccurate problem.In order to address this problem, Mahler also been proposed radix probability hypothesis density (Cardinalized Probability Hypothesis Density, CPHD) wave filter.CPHD obtains number of targets accurately by the strength function of simultaneously spread state random set and radix distribution and estimates.Although PHD wave filter can process multiobject new life, hatching and problem of death, but still there is a shortcoming, namely standard P HD wave filter supposes that the strength function of newborn target is known conditions.But the detection of newborn target or the initial of flight path are the ingredients that modern MTT system is necessary and important.Therefore, this hypothesis constrains the widespread use of PHD wave filter.

In recent years, the people such as B.RISTIC proposes a kind of newborn target strength method of self-adaptation based on particle PHD wave filter.The method surveys the newborn target strength function of design by each frame amount, be divided into measurement can survey subspace p in measurement space and can not survey subspace v with measurement, by likelihood function, newborn target strength is constructed for p, but known prior imformation is still needed for v, the not basic construction problem solving newborn target strength function.The people such as MichaelBeard propose newborn target strength and are distributed as Gaussian Mixture PHD wave filter when being uniformly distributed, but still it is known to suppose that newborn target strength is distributed as.The people such as Ou Yangcheng are for document, a kind of normalized factor modification method is proposed, in order to improve the flight path normalization unbalance of former algorithm, but the method still needs known portions prior imformation, and the method still comprises a large amount of clutter after the filtering, weaken the major advantage that clutter removed by PHD wave filter.

Summary of the invention

The present invention is directed to above shortcomings in prior art, propose a kind of probability hypothesis density filtering algorithm based on the newborn target strength of self-adaptation.This algorithm does not need prior imformation completely, only constructs newborn target strength function according to measurement information.This method may be used for the newborn Target Tracking Problem of the Location-Unknown of PHD wave filter, and is also applicable to CPHD wave filter.This algorithm has can the newborn target of tracing positional the unknown well, thus improves the advantage such as range of application and robustness of PHD wave filter.

The present invention is achieved by the following technical solutions, the present invention includes following steps:

The first step, any measurement of being arrived by each reception as new goal hypothesis (flight path head), and carries out track initiation with this, and carries out track confirmation by the newborn target strength generating algorithm of self-adaptation.

The newborn target strength generating algorithm of described self-adaptation, specifically, the thought of Delayed Decision in multiple hypotheis tracking is adopted to carry out logic initial, namely do not carry out association at once to measure when initial flight path One-step Extrapolation, but postpone a step decision-making, postpone outside continuous two steps, then choose optimum extrapolation flight path and the measurement of association thereof according to statistical method.

Second step, the initial sum of targetpath confirms, if there is newborn target (flight path of confirmation), then carry out step 3, otherwise carry out step 4.

3rd step, if there is the flight path (i.e. fresh target flight path) confirmed, then according to confirming that the start position information of flight path constructs the strength function of newborn target random set, the newborn target strength function merging newborn target strength function that present frame builds known with previous moment produces present frame self-adaptation new life target strength function.

The strength function of described newborn target random set adopts Gaussian Mixture form, specifically:

${\mathrm{\Σ}}_{j=1}^{{\tilde{J}}_{Y,k}}{\widetilde{\mathrm{\ω}}}_{\mathrm{\γ},k}^{j}N(x;{\tilde{m}}_{\mathrm{\γ},k}^j,{\tilde{P}}_{\mathrm{\γ},k}^j)---\left(1\right)$

The newborn target strength function of described present frame self-adaptation, specifically:

${\mathrm{\γ}}_k\left(x\right)=\underset{i=1}{\overset{J_{\mathrm{\γ},k}}{\mathrm{\Σ}}}{\mathrm{\ω}}_{\mathrm{\γ},k}^{i}N(x;m_{\mathrm{\γ},k}^i,P_{\mathrm{\γ},k}^i)+\underset{j=1}{\overset{{\tilde{J}}_{\mathrm{\γ},k}}{\mathrm{\Σ}}}{\widetilde{\mathrm{\ω}}}_{\mathrm{\γ},k}^{j}N(x;{\tilde{m}}_{\mathrm{\γ},k}^j,{\tilde{P}}_{\mathrm{\γ},k}^j)---\left(2\right)$

Wherein, the place-centric vector of known newborn target respectively, the weighting coefficient of variance matrix and each target; the place-centric vector of each newborn target that step 2 constructs respectively, variance matrix and weighting coefficient.

4th step, uses the newborn target strength function of self-adaptation to start PHD wave filter, thus completes prediction and the renewal of PHD wave filter.

The prediction of described PHD wave filter, specifically: predict to newborn target prediction with to the target of survival.

The described prediction to newborn target, specifically: the newborn target strength function after the renewal obtained in step 3 is predicted, the newborn target that the newborn target comprising known reference position obtains with detection.

The renewal of described PHD wave filter, specifically: when measure lose time prediction Gauss unit renewal and by measure to prediction Gauss unit renewal.

5th step, carry out cut operation, and the expectation value extracting the larger Gauss unit of weights is as state estimation to the Gauss unit after upgrading in step 4.

Compared with prior art, the invention has the beneficial effects as follows: this algorithm does not need prior imformation completely, only construct newborn target strength function according to measurement information.This method may be used for the newborn Target Tracking Problem of the Location-Unknown of PHD wave filter, and is also applicable to CPHD wave filter.This algorithm has can the newborn target of tracing positional the unknown well, thus improves the advantage such as range of application and robustness of PHD wave filter.

Accompanying drawing explanation

Fig. 1 is the PHD filter flow figure detected based on adaptive targets.

Fig. 2 is One-step delay algorithm candidate target flight path expansion schematic diagram.

Fig. 3 be context of methods on x-and y-direction to the location estimation figure of time.

Fig. 4 be standard GM-PHD on x-and y-direction to the location estimation figure of time.

Fig. 5 is that context of methods and GM-PHD number of targets estimate comparison diagram.

Fig. 6 is the average Wasserstein distance versus of 100 Monte Carlo of context of methods and GM-PHD method.

Embodiment

Below in conjunction with accompanying drawing, embodiments of the invention are elaborated: the present embodiment is implemented under premised on technical solution of the present invention, give detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to following embodiment.

Embodiment

The present embodiment is two-dimentional maneuver tracking scene, and by radar tracking five targets, target can occur with optional position at any time, and newborn target distribution obeys Poisson distribution.In order to compare with the PHD wave filter of standard, suppose that the reference position of two newborn targets is known, its strength function is:

${\mathrm{\γ}}_k\left(x\right)={\mathrm{\Σ}}_{i=1}^2{w}_{y}N(x;m_y^i,P_y),$ Wherein, w _y=0.03, $m_{\mathrm{\γ}}^1={\left[\begin{array}{ccccc}0& 0& 0& 0& 0\end{array}\right]}^{\′},$ P＝diag([50 10 50 10 0.1]')。The initial intensity function of its excess-three target is unknown.The state of target is taken as: wherein, P _x, P _yrepresent target position coordinates in the x and y direction, represent target speed in the x and y direction.Constant speed (CV) model chosen by the state model of target:

x _k＝Fx _k-1+Gv _k-1(3)

Wherein:

$F=\left[\begin{array}{cccc}1& 0& 1& 0\\ 0& 1& 0& 1\\ 0& 0& 1& 0\\ 0& 0& 0& 1\end{array}\right]---\left(4\right)$

$G=\left[\begin{array}{cc}\frac{1}{2}& 0\\ 0& \frac{1}{2}\\ 0& 0\\ 0& 1\end{array}\right]---\left(5\right)$

standard deviation is s _v=10m/s ².The measurement of target is for containing noisy positional information.Measurement model is

Z _k＝Hx _k+w _k(6)

Wherein: $H=\left[\begin{array}{ccccc}1& 0& 0& 0& 0\\ 0& 1& 0& 0& 0\end{array}\right],$ Measuring noise square difference is here measuring standard difference is s _v=50 meters.The survival probability of hypothetical target is 0.99, aimless division in flight course.Clutter obeys poisson process model, and intensity is K _{k (z)}=λ cVu (z), wherein u () obeys equally distributed density function in monitor area V, volume V=4 × 10 of monitor area ⁸rice ², λ _c=1 × 10 ^-7rice ²it is the mean clutter number in unit volume.The optimum configurations of GM-PHD wave filter is: it is T=10 that thresholding is deleted by Gauss unit ^-5, it is u=4m that Gauss unit merges thresholding, and each moment uses J at the most _max=100 Gauss units.

The first step, any measurement of being arrived by each reception as new goal hypothesis (flight path head), and carries out track initiation with this, and carries out track confirmation operation by the newborn target strength generating algorithm of self-adaptation.

It is initial that the present embodiment adopts the thought of Delayed Decision in multiple hypotheis tracking to carry out logic, namely do not carry out association at once to measure when initial flight path One-step Extrapolation, but postpone a step decision-making, postpone outside continuous two steps, optimum extrapolation flight path and the measurement of association thereof is chosen again according to statistical method, thus quick initial flight path.

Second step, the initial sum of targetpath confirms, if there is newborn target (flight path of confirmation), then carry out step 3, otherwise carry out step 4.

According to this paper part 2 newborn target strength function computing method, measure initial sum expansion by present frame and suppose flight path temporarily, and judge whether to there is the flight path confirmed, namely whether there is newborn target, judge next step thus.

3rd step, if there is the flight path (i.e. fresh target flight path) confirmed, then according to confirming that the start position information of flight path constructs the strength function of newborn target random set, the newborn target strength function merging newborn target strength function that present frame builds known with previous moment produces this current frame adaptive new life target strength function.

The strength function of described newborn target random set adopts Gaussian Mixture form, specifically:

${\mathrm{\Σ}}_{j=1}^{{\tilde{J}}_{\mathrm{\γ},k}}{\widetilde{\mathrm{\ω}}}_{\mathrm{\γ},k}^{j}N(x;{\tilde{m}}_{\mathrm{\γ},k}^j,{\tilde{P}}_{\mathrm{\γ},k}^j)---\left(7\right)$

The newborn target strength function of described present frame self-adaptation, specifically

${\mathrm{\γ}}_k\left(x\right)=\underset{i=1}{\overset{J_{\mathrm{\γ},k}}{\mathrm{\Σ}}}{\mathrm{\ω}}_{\mathrm{\γ},k}^{i}N(x;m_{\mathrm{\γ},k}^i,P_{\mathrm{\γ},k}^i)+\underset{j=1}{\overset{{\tilde{J}}_{\mathrm{\γ},k}}{\mathrm{\Σ}}}{\widetilde{\mathrm{\ω}}}_{\mathrm{\γ},k}^{j}N(x;{\tilde{m}}_{\mathrm{\γ},k}^j,{\tilde{P}}_{\mathrm{\γ},k}^j)---\left(8\right)$

Wherein, the place-centric vector of known newborn target respectively, the weighting coefficient of variance matrix and each target; the place-centric vector of each newborn target that step 2 constructs respectively, variance matrix and weighting coefficient.

4th step, uses the newborn target strength function of self-adaptation to start PHD wave filter, thus completes prediction and the renewal of PHD wave filter.

The prediction of described PHD wave filter, specifically: predict to newborn target prediction with to the target of survival.

The described prediction to newborn target, specifically: the newborn target strength function after the renewal obtained in step 3 is predicted, the newborn target that the newborn target comprising known reference position obtains with detection.

The renewal of described PHD wave filter, specifically: when measure lose time prediction Gauss unit renewal and by measure to prediction Gauss unit renewal.

5th step, carry out cut operation, and the expectation value extracting the larger Gauss unit of weights is as state estimation to the Gauss unit after upgrading in step 4.

In the present embodiment, self-adaptation newborn target strength generating algorithm specific operation process is:

1) initial flight path is set up.Use k reception to each measurement set up track initiation point.

2) two step extrapolations.As shown in Figure 2, for the candidate target flight path a in k+1 moment, carry out single order polynomial extrapolation, form the confirmation region of k+2 moment corresponding to this flight path, wherein b is forecast position, and b1, b2 fall into the measurement confirming region; Utilize b1 respectively, b2 carries out second order polynomial extrapolation, obtain the k+3 moment measure confirm region.

3) the new breath of two step accumulations is calculated.Combination l{z is measured for each confirmation in region _{(l, k-1)}, z _{(l, k)}, defining its combination new breath normalization variance is:

$\mathrm{\γ}\left(l\right)=\underset{i=k-1}{\overset{k}{\mathrm{\Σ}}}{[z_{(l,i)}\left(i\right)-H\left(i\right){\hat{x}}_l\left(i\right)]}^{T}R{\left(i\right)}^{-1}[z_{(l,i)}\left(i\right)-H\left(i\right){\hat{x}}_l\left(i\right)]---\left(9\right)$

Wherein, z _{(l, k)}represent that the k moment belongs to the measurement of combination l.X _il () combines Target state estimator corresponding to l with measurement.

4) targetpath expansion.The minimum measurement corresponding with γ (l) is selected to carry out targetpath expansion as effective measurement in k moment;

5) track confirmation.

Track confirmation logic adopts the m/n logical approach based on " 4 sliding windows ", namely initial flight path remains that 4 scanning measures, if the measurement of continuous 3 scan periods till a targetpath expansion is included in current time, just judge that this flight path is as confirming flight path, the state x (k) of flight path carries out linear extrapolation by least square method and obtains:

$x(k+1)=x\left(k\right)+\frac{1}{2}(x(k-1)-x(k-3))---\left(10\right)$

The specific operation process that in the present embodiment, the Gaussian Mixture of the newborn target strength PHD algorithm of self-adaptation realizes is:

The Gaussian Mixture of the newborn target strength PHD algorithm of described self-adaptation realizes, for GM-PHD filtering.GM-PHD algorithm steps then based on the newborn target detection of self-adaptation is as follows, wherein, supposes that the measurement collection in known k moment is the flight path of being survived by the k-1 moment is assumed to be

1) initial sum of targetpath confirms.According to this paper part 2 newborn target strength function computing method, measure initial sum expansion by present frame and suppose flight path temporarily, and judge whether to there is the flight path confirmed, namely whether there is newborn target.If there is newborn target (flight path of confirmation), then carry out step 2, otherwise carry out step 3.

2) structure is with the strength function of newborn target.If the k moment is confirmed by step 1 individual newborn target, then according to the strength function of newborn target location structure Gaussian Mixture form and merge with a upper moment known fresh target strength function and form newborn target strength function gamma _k(x), namely

${\mathrm{\γ}}_k\left(x\right)=\underset{i=1}{\overset{J_{\mathrm{\γ},k}}{\mathrm{\Σ}}}{\mathrm{\ω}}_{\mathrm{\γ},k}^{i}N(x;m_{\mathrm{\γ},k}^i,P_{\mathrm{\γ},k}^i)+\underset{j=1}{\overset{{\tilde{J}}_{\mathrm{\γ},k}}{\mathrm{\Σ}}}{\widetilde{\mathrm{\ω}}}_{\mathrm{\γ},k}^{j}N(x;{\tilde{m}}_{\mathrm{\γ},k}^j,{\tilde{P}}_{\mathrm{\γ},k}^j)---\left(11\right)$

Wherein, the place-centric vector of known newborn target respectively, the weighting coefficient of variance matrix and each target; the place-centric vector of each newborn target that step 2 constructs respectively, variance matrix and weighting coefficient.

3) probability hypothesis density of Gaussian Mixture method to multiple goal state random set is used to predict.If k-1 moment state posteriority strength function is by J _k-1individual Gauss unit composition, the posteriority intensity of its Gaussian Mixture form is,

$v_{k-1}\left(x\right)=\underset{j=1}{\overset{J_{k-1}}{\mathrm{\Σ}}}{\mathrm{\ω}}_{k-1}^{j}N(x;m_{k-1}^j,P_{k-1}^j)---\left(12\right)$

Then the k moment predicts that the Gaussian Mixture form of PHD is as follows,

V _k|k-1(x)=γ _k(x)+v _{s, k|k-1}(x) (13) wherein γ _kx () is fresh target prediction Gauss unit strength function, v _{s, k|k-1}(x) for survival target prediction Gauss unit strength function, then respectively to newborn target strength and survival target predict as follows.

(1) newborn target prediction.Newborn target strength function after the renewal obtained in step 2 is predicted, the newborn target that the newborn target comprising known reference position obtains with detection.

To k moment J in step 2 _{γ, k}individual known location newly produces target and predicts, namely to i=1 ..., J _{γ, k}, ${\mathrm{\ω}}_{k|k-1}^i={\mathrm{\ω}}_{\mathrm{\γ},k}^i,m_{k|k-1}^i=m_{\mathrm{\γ},k}^i,P_{k|k-1}^i=P_{\mathrm{\γ},k}^i,$ Wherein, be respectively the weights of k moment newborn target prediction Gauss unit, state is expected and covariance;

For the k moment in step 2 the fresh target of the Location-Unknown of individual structure is predicted, for $j=J_{\mathrm{\γ},k}+1,...,J_{\mathrm{\γ},k}+{\tilde{J}}_{\mathrm{\γ},k},$ Have ${\mathrm{\ω}}_{k|k-1}^j={\widetilde{\mathrm{\ω}}}_{\mathrm{\γ},k}^{j-J_{\mathrm{\γ},k}},m_{k|k-1}^j={\tilde{m}}_{\mathrm{\γ},k}^{j-J_{\mathrm{\γ},k}},P_{k|k-1}^j={\tilde{P}}_{\mathrm{\γ},k}^{j-J_{\mathrm{\γ},k}},$ Wherein be respectively the weights of the newborn target prediction Gauss unit of unknown position, state is expected and covariance;

(2) target of survival is predicted.

The survival probability of hypothetical target is p _s, then to i=1 ..., J _k-1, survival target is predicted have ${\mathrm{\ω}}_{k|k-1}^i=p_S{\mathrm{\ω}}_{k-1}^i,m_{k|k-1}^i=F_{k-1}{m}_{k-1}^i,P_{k|k-1}^i=Q_{k-1}^i+F_{k-1}{P}_{k-1}^i{\left(F_{k-1}\right)}^T,$ Wherein for the weights of survival target prediction Gauss unit, average and covariance, for process-noise variance battle array.

The prediction PHD calculated by above two steps is the Gauss unit sum of newborn Gauss unit and survival, supposes that the total number of prediction Gauss unit is J _k|k-1.

4) by current measurement set, the Gauss unit of multi-objective predictive PHD is upgraded.Suppose that the k moment predicts that PHD has Gaussian Mixture form:

$v_{k|k-1}\left(x\right)=\underset{i=1}{\overset{J_{k|k-1}}{\mathrm{\Σ}}}{\mathrm{\ω}}_{k|k-1}^{i}N(x;m_{k|k-1}^i,P_{k|k-1}^i)---\left(14\right)$

Then k moment posteriority PHD has following Gaussian Mixture form:

$v_K\left(x\right)=(1-P_D)v_{k|k-1}\left(x\right)+\underset{z\∈Z_k}{\mathrm{\Σ}}{v}_{D,k}(x;z)---\left(15\right)$

Wherein

$v_{D,k}(x;z)=\underset{i=1}{\overset{J_{k|k-1}}{\mathrm{\Σ}}}{\mathrm{\ω}}_k^{i}N(x;m_{k|k}^i\left(z\right),P_{k|k}^i)---\left(16\right)$

${\mathrm{\ω}}_k^i=\frac{p_{D,k}{\mathrm{\ω}}_{k|k-1}^i{q}_k^i\left(z\right)}{{\mathrm{\κ}}_k\left(z\right)+P_{D,k}{\mathrm{\Σ}}_{j=1}^{J_{k|k-1}}{\mathrm{\ω}}_{k|k-1}^j{q}_k^j\left(z\right)}---\left(17\right)$

$q_k^i\left(z\right)=N(z;H_k{m}_{k|k-1}^j,\mathrm{nR}+H_k{P}_{k|k-1}^j{H}_k^T)---\left(18\right)$

Wherein, H _kfor state-transition matrix.

According to formula (15), the renewal of PHD comprises two parts, and namely monovalent surveys the renewal of prediction Gauss unit when losing, and two by measurement Z _kto the renewal of prediction Gauss unit.

(1) the prediction Gauss unit measured when losing upgrades:

To i=1 ..., J _k+1|kupgrade Gauss unit according to the following formula, have $P_{k+1}^i=P_{k+1|k}^i.$

(2) measurement Z is used _k+1prediction Gauss unit is upgraded:

To each z ∈ Z _k+1, have

${\mathrm{\ω}}_{k+1}^i=p_D{\mathrm{\ω}}_{k+1|k}^{i}N(z;{\mathrm{\η}}_{k+1|k}^i,S_{k+1}^i)---\left(19\right)$

$m_{k+1}^i=m_{k+1|k}^i+K_{k+1}^i(z-{\mathrm{\η}}_{k+1|k}^i)---\left(20\right)$

$P_{k+1}^i=[I-K_k^i{H}_k]P_{k+1|k}^i---\left(21\right)$

$K_{k+1}^i=P_{k+1|k}^i{H}_k^T{[R+H_k{P}_{k+1|k}^i{H}_k^T]}^{-1}---\left(22\right)$

In formula with average and the variance of a jth prediction Gauss unit respectively.

5) multiple goal state extracts.Cut operation is carried out to the Gauss unit after upgrading in step 4, and the expectation value extracting the larger Gauss unit of weights is as state estimation.

Implementation result

Fig. 3 gives context of methods and estimates the location status of time on x-and y-direction.As shown in Figure 3, context of methods not only accurately can estimate the newborn target of two known reference positions, and accurately can estimate the newborn target of three unknown reference positions.Fig. 4 uses former GM-PHD wave filter to the state estimation result of same multiple target tracking scene.As seen from Figure 4, GM-PHD wave filter only can estimate the newborn target of two known reference positions, has occurred losing with problem to the newborn target of three unknown reference positions.This is because GM-PHD wave filter cannot the unknown fresh target of reference position, and the fresh target of Location-Unknown is caused as clutter process.

Fig. 5 gives real goal number, and the number of targets of context of methods and GM-PHD wave filter estimates comparison diagram, and in figure, number of targets estimation is the result of 100 Monte Carlo arithmetic averages.Contrast visible, algorithm gives more correct target numbers estimated value herein, and GM-PHD wave filter only gives the number estimated result of known reference position target, and number of targets estimated result is inaccurate.

Wasserstein distance is suitable for the performance evaluation of multi-object tracking method, and this index had both punished the Target state estimator error of algorithm, also punishes the error that multiple goal number is estimated.Fig. 6 compares context of methods and the average Wasserstein distance of former GM-PHD wave filter 100 Monte Carlo.Can be drawn by Fig. 6, in the time step of 50 to 130, the Wasserstein distance of former GM-PHD wave filter is greater than algorithm herein.The decline of these performance index is primarily of former GM-PHD wave filter losing with what cause the newborn target of Location-Unknown, and the number of targets giving mistake in this time step is estimated.Visible, this implementation system can the newborn target of tracing positional the unknown well, thus improves range of application and the robustness of PHD wave filter.

Claims (6)

1., based on a probability hypothesis density filtering algorithm for the newborn target strength of self-adaptation, it is characterized in that, comprise the following steps:

The first step, any measurement of being arrived by each reception as new goal hypothesis (flight path head), and carries out track initiation with this, and carries out track confirmation operation by the newborn target strength generating algorithm of self-adaptation;

Second step, the initial sum of targetpath confirms, if there is newborn target (flight path of confirmation), then carry out step 3, otherwise carry out step 4;

3rd step, if there is the flight path (i.e. fresh target flight path) confirmed, then according to confirming that the start position information of flight path constructs the strength function of newborn target random set, the newborn target strength function merging newborn target strength function that present frame builds known with previous moment produces this current frame adaptive new life target strength function;

4th step, uses the newborn target strength function of self-adaptation to start PHD wave filter, thus completes prediction and the renewal of PHD wave filter;

5th step, carry out cut operation, and the expectation value extracting the larger Gauss unit of weights is as state estimation to the Gauss unit after upgrading in step 4.

2. the probability hypothesis density filtering algorithm based on the newborn target strength of self-adaptation according to claim 1, it is characterized in that, the newborn target strength generating algorithm of described self-adaptation, specifically, adopt the thought of Delayed Decision in multiple hypotheis tracking to carry out logic initial, namely do not carry out association when initial flight path One-step Extrapolation at once and measure, but delay one step decision-making, postpone outside continuous two steps, then choose optimum extrapolation flight path and the measurement of association thereof according to statistical method.

3. the probability hypothesis density filtering algorithm based on the newborn target strength of self-adaptation according to claim 1, is characterized in that, the strength function of described newborn target random set, specifically, if the k moment confirmed by step 1 individual newborn target, then according to the strength function of newborn target location structure Gaussian Mixture form, namely

${\mathrm{\Σ}}_{j=1}^{{\tilde{J}}_{\mathrm{\γ},k}}{\widetilde{\mathrm{\ω}}}_{\mathrm{\γ},k}^{j}N(x;{\tilde{m}}_{\mathrm{\γ},k}^j,{\tilde{P}}_{\mathrm{\γ},k}^j)---\left(2\right)$

4. the probability hypothesis density filtering algorithm based on the newborn target strength of self-adaptation according to claim 1, it is characterized in that, the newborn target strength function of described present frame self-adaptation, specifically, merges with fresh target strength function known in claim 3 and forms newborn target strength function gamma _k(x), namely

${\mathrm{\γ}}_k\left(x\right)=\underset{i=1}{\overset{J_{\mathrm{\γ},k}}{\mathrm{\Σ}}}{\mathrm{\ω}}_{\mathrm{\γ},k}^{i}N(x;m_{\mathrm{\γ},k}^i,P_{\mathrm{\γ},k}^i)+\underset{j=1}{\overset{{\tilde{J}}_{\mathrm{\γ},k}}{\mathrm{\Σ}}}{\widetilde{\mathrm{\ω}}}_{\mathrm{\γ},k}^{j}N(x;{\tilde{m}}_{\mathrm{\γ},k}^j,{\tilde{P}}_{\mathrm{\γ},k}^j)---\left(3\right)$

Wherein, the place-centric vector of known newborn target respectively, the weighting coefficient of variance matrix and each target; the place-centric vector of each newborn target that step 2 constructs respectively, variance matrix and weighting coefficient.

5. the probability hypothesis density filtering algorithm based on the newborn target strength of self-adaptation according to claim 1, is characterized in that, the prediction of described PHD wave filter, specifically, predicts to newborn target prediction with to the target of survival.

6. the probability hypothesis density filtering algorithm based on the newborn target strength of self-adaptation according to claim 1, it is characterized in that, the renewal of described PHD wave filter, specifically, when measure lose time prediction Gauss unit renewal and by measure to prediction Gauss unit renewal.