CN104019816A - Flight track extraction method based on probability hypothesis density filter associated with global time and space - Google Patents

Abstract

The invention discloses a flight track extraction method based on a probability hypothesis density filter associated with global time and space. The flight track extraction method comprises the following steps: S1, extracting a target state by using the probability hypothesis density filter; S2, measuring consistency and calculating consistency confidence; and S3, obtaining a global flight track extraction strategy. By adopting the scheme, the target state is firstly obtained by using the probability hypothesis density filter; then the consistency between a predicted peak value and an estimated peak value is measured by using global time-space information and the consistency confidence is calculated; simultaneously four decision rules based on expert knowledge of the flight track extraction are shown, namely a rule for judging inseparable targets, a rule for judging homology of the targets, a rule for judging disappearance of the targets and a rule for judging novel targets; a global flight track extraction strategy is shown based on the consistency confidence and four decision rules, thereby extracting flight tracks of a plurality of targets, improving the flight track extraction effect, improving the flight track extraction accuracy and playing an important role in multi-target tracking engineering application.

Description

Probability hypothesis density wave filter flight path extracting method based on overall temporal and spatial correlations

Technical field

The present invention relates to a kind of flight path extracting method, particularly relate to a kind of probability hypothesis density wave filter flight path extracting method based on overall temporal and spatial correlations.

Background technology

Multitarget Tracking, in multiple fields extensive application such as communication, radar, biomedicines, is an important and difficult research topic.Wherein, in clutter environment, carrying out multiple target tracking is a more difficult problem, and in the time of target or clutter appearance or disappearance, the observed reading that number of targets and target produce may change in time.The method of traditional processing multiple target tracking problem is taking data correlation as basic multiple target tracking algorithm, such as nearest neighbor method (NN), probabilistic data association (PDA) algorithm, JPDA (JPDA) algorithm and many hypothesis are followed the tracks of (MHT) algorithm etc.This class algorithm, by the association process with flight path to measurement, utilizes the related algorithm in monotrack to upgrade every flight path.But data correlation itself is very complicated, simultaneously general data association algorithm cannot processing target be counted the situation of real-time change.2003, Mahler has proposed the algorithm of another kind of processing multiple target tracking problem, i.e. (the Probabi1ity hypothesis density of the probability hypothesis density based on random finite set theory, PHD) filtering algorithm, the state of multiple targets and measurement are expressed as random set by it, calculate the PHD of the random finite set of dbjective state by recursion, make multiple target tracking problem be converted into single-sensor monotrack problem, thereby avoided data correlation.PHD algorithm is the integral operation that set function integral operation is reduced to single variable, is corresponding number of targets at the PHD integration that monitors spatial domain, the corresponding dbjective state of peak value that PHD distributes.But although PHD filtering algorithm has been avoided data correlation, it can only obtain dbjective state and target number, and continuous targetpath information can not be provided; And in multiple target tracking, obtaining continual and steady targetpath information is its final purpose.

For this problem, L.Lin (Data Association Combined with the Probability Hypothesis Density Filter for Multitarget Tracking.Proceedings of SPIE conference on Signal and Data Processing on Small Targets, 2004.) clap target status information association by filtered PHD continuous two, as assignment problem processing, K.Panta (Probabi1ity Hypothesis Density Fi1ter versus Multiple Hypothesis Tracking.Proceedings of SPIE, Signal Processing, Sensor Fusion & Target Recognition XIII, 2004,5429 (8): 284-295.) traditional many tentation datas correlating method is combined with PHD, obtain the track estimation information of single goal, but increased extra calculated amount, F.Papi (Multitarget tracking via joint PHD filtering and multiscan association.Proceedings of the12th International Conference on Information Fusion, 2009.) form mixing PHD-SDA tracker in conjunction with multiframe association (SDA) algorithm and particle PHD wave filter, the more new particle clustering center obtaining using multiframe particle PHD wave filter and center situation covariance are as the input data of SDA, by carrying out SDA optimized algorithm, remove the pseudo-flight path of standard cancellation part according to flight path, and output residue flight path current state is estimated, thereby simultaneously according to prior probability distribution to current state resampling again initial PHD wave filter, D.E.Clark (Data association for the PHD filter.Proceedings of the International Conference on Intelligent Sensors, Sensor Networks and Information Processing, 2005.) label in each iteration of particle filter PHD each particle, then by comparing the number of same label in cluster, target estimation point is got up to relevant track association, L.Lin (Track 1abeling and PHD filter for multitarget tracking.IEEE Transactiohs on Aerospace and Electronic Systems, 2006,42 (3): 778-794.) flight path labeling acts is combined with PHD, for each target sets up a tracker, and determine peak state and number of targets in conjunction with resolution characteristic, then carry out the association process of peak value and flight path, and association results can feed back to PHD wave filter.Above-mentioned flight path extracting method all based on local spatial information, is not used overall space time information in the time extracting arbitrary flight path simultaneously.And under multiple target tracking environment, between target location, may there is certain time-space relationship, if can use this category information, flight path extraction accuracy perhaps can improve.For this point, the present invention considers overall space time information, given first prediction peak value and estimate consistency metric and the consistance confidence calculations between peak value, then provided overall flight path fetch strategy based on consistance degree of confidence and four decision rules, realize multiobject flight path and extract.

Summary of the invention

For addressing the above problem, the invention provides a kind of probability hypothesis density wave filter flight path extracting method based on overall temporal and spatial correlations.

For achieving the above object, the technical scheme that the present invention takes is:

Probability hypothesis density wave filter flight path extracting method based on overall temporal and spatial correlations, comprises the following steps:

S1, probability hypothesis density wave filter dbjective state are extracted;

S2, consistency metric and consistance confidence calculations;

S3, overall flight path fetch strategy.

Wherein, described step S1 specifically comprises the steps:

S11, prediction: for original target, sample and calculate the weights of predicting particle; For target that may be newborn, sample and calculate the weights of newborn particle;

S12, renewal: upgrade weights with measuring;

S13, resampling: calculate weights total amount rounding and obtain number of targets, resampling regulates weights.

Wherein, described step S2 specifically comprises the steps:

S21, obtain the estimation peak set in adjacent two moment by step S1;

S22, estimation peak set based on previous moment in two moment, carry out one-step prediction, obtains the prediction peak set in next moment;

S23, calculate relative coordinate: calculate respectively in synchronization prediction peak set between 2 and estimate the directed line segment length between 2 in peak set;

Distance between S24, calculating relative coordinate, calculates by the length difference between step S23 gained directed line segment;

Consistance degree of confidence between S25, calculating twice sweep point, sets up consistance degree of confidence matrix.

Wherein, described step S3 specifically comprises the steps:

S31, based on " judging inseparable goal rule ", set up inseparable object set (ITS, ITS is the abbreviation of Impartibility Target ' s Set), extract inseparable target;

S32, based on " judging target homology rule ", set up targetpath collection (TTS, TTS is the abbreviation of Target ' s Track Set), extract survival target flight path;

S33, based on " judge that target disappear rule ", set up disappearance object set (DTS, DTS is the abbreviation of Disappearing Target ' s Set), extract the target disappearing, flight path correctly terminates;

S34, based on " judge target newborn rule ", set up newborn object set (BTS, BTS is the abbreviation of Birth Target ' s Set), extract newborn target, correct initial flight path.

Beneficial effect of the present invention is as follows:

In such scheme, utilize overall space time information, provide prediction peak value and estimated consistency metric and the consistance confidence calculations between peak value, the expertise extracting based on flight path has provided four decision rules, judge inseparable goal rule, judge target homology rule, judge that target disappears regular, judge the newborn rule of target, provide overall flight path fetch strategy based on consistance degree of confidence and four decision rules simultaneously, thereby realizing multiobject flight path extracts, improve flight path extraction effect, improve flight path extraction accuracy, for multiple target tracking through engineering approaches, application has great importance.

Brief description of the drawings

Fig. 1 is the probability hypothesis density filtering algorithm programming realization flow figure of the probability hypothesis density wave filter flight path extracting method of the embodiment of the present invention based on overall temporal and spatial correlations.

Fig. 2 is consistency metric and consistance confidence calculations process flow diagram in the probability hypothesis density wave filter flight path extracting method of the embodiment of the present invention based on overall temporal and spatial correlations.

Fig. 3 is that the flight path of the probability hypothesis density wave filter flight path extracting method of the embodiment of the present invention based on overall temporal and spatial correlations extracts process flow diagram.

Fig. 4 is that the flight path of the probability hypothesis density wave filter flight path extracting method of the embodiment of the present invention based on overall temporal and spatial correlations extracts result figure.

Embodiment

In order to make objects and advantages of the present invention clearer, below in conjunction with embodiment, the present invention is further elaborated.Should be appreciated that specific embodiment described herein, only in order to explain the present invention, is not intended to limit the present invention.

Embodiments of the invention provide a kind of probability hypothesis density wave filter flight path extracting method based on overall temporal and spatial correlations, comprise the following steps:

S1, probability hypothesis density wave filter dbjective state are extracted;

S2, consistency metric and consistance confidence calculations;

S3, overall flight path fetch strategy.

As shown in Figure 1, described step S1 specifically comprises the steps:

A, prediction.For i=1,2 ..., L _k-1, sampling and the weights of calculating prediction particle ${\widetilde{\mathrm{\ω}}}_{k|k-1}^{\left(i\right)}=\frac{{\mathrm{\Φ}}_{k|k-1}({\widetilde{\mathrm{\ξ}}}_k^{\left(i\right)},{\mathrm{\ξ}}_{k-1}^{\left(i\right)})}{q_k\left({\widetilde{\mathrm{\ξ}}}_k^{\left(i\right)}\right|{\mathrm{\ξ}}_{k-1}^{\left(i\right)},Z_k)}{\mathrm{\ω}}_{k-1}^{\left(i\right)};$ For i=L _k-1+ 1, L _k-1+ 2 ..., L _k-1+ J _k, sampling ${\widetilde{\mathrm{\ξ}}}_k^{\left(i\right)}~p_k(\·|Z_k)$ And calculate the weights of newborn particle in formula, Φ _k|k-1()=d _k|k-1() f _k|k-1(|), wherein, d _k|k-1(), f _k|k-1(|) represents respectively target survival probability, transition probability.Q _k() and p _k() is the suggestion density of importance sampling, d _k() is the PHD of newborn target.

B, renewal.To each z ∈ Z _k, calculate to i=1,2 ..., L _k+1+ J _k, upgrade weights wherein ψ _{k, z}(ξ)=P _df _k(z| ξ), P _dfor detection probability, f _k() is sensor likelihood function.K _k(z)=λ _kc _k(z), λ _kfor each scanning false-alarm average, be assumed to be Poisson distribution, c _k(z) be the probability distribution of false-alarm.

C, resampling.Calculate weights total amount ${\hat{N}}_{k|k}={\mathrm{\Σ}}_{j=1}^{L_{k-1}+J_k}{\widetilde{\mathrm{\ω}}}_k^{\left(j\right)};$ Resampling ${\{({\widetilde{\mathrm{\ω}}}_k^{\left(i\right)}/{\hat{N}}_{k|k}),{\widetilde{\mathrm{\ξ}}}_k^{\left(i\right)}\}}_{i=1}^{L_{k-1}+J_k}$ Obtain with be multiplied by weights and readjust weights acquisition

As shown in Figure 2, described step S2 specifically comprises the steps:

A, obtain two peak set by step S1 with wherein be the estimating target state in k moment, i, j are respectively the order label in k moment and k+1 moment, n _kfor the number of targets in k moment, n _k+1for the number of targets in k+1 moment.

B, based on peak set carry out one-step prediction, obtain the prediction peak set in next moment it is the target of prediction state in k+1 moment;

C, calculate relative coordinate, calculate respectively the k+1 moment predict in peak set between 2 and estimation peak set in directed line segment length between 2;

Distance between D, calculating relative coordinate, i.e. length difference between calculation procedure C gained directed line segment;

E, definition are also calculated the consistency metric between twice sweep point,

$\mathrm{CM}({\tilde{x}}_k^j,{\tilde{x}}_{k+1}^j)=\underset{\stackrel{g\≠i}{\stackrel{(i=\mathrm{1,2}...n_k)}{(j=\mathrm{1,2},...,n_{k+1})}}}{\mathrm{\Σ}}\underset{l\≠j}{\max }\left\{\exp (-\frac{1}{\mathrm{\α}}r({\hat{x}}_{k+1}^g-{\hat{x}}_{k+1}^i,{\tilde{x}}_{k+1}^l-{\tilde{x}}_{k+1}^j))\right\}$

F, definition are also calculated the consistance degree of confidence between twice sweep point,

$\mathrm{\μ}({\tilde{x}}_k^j,{\tilde{x}}_{k+1}^j)=\frac{\mathrm{CM}({\tilde{x}}_k^j,{\tilde{x}}_{k+1}^j)-\underset{i,j}{\min }\left(\mathrm{CM}({\tilde{x}}_k^j,{\tilde{x}}_{k+1}^j)\right)}{\underset{i,j}{\max }\left(\mathrm{CM}({\tilde{x}}_k^i,{\tilde{x}}_{k+1}^j)\right)-\underset{i,j}{\min }\left(\mathrm{CM}({\tilde{x}}_k^i,{\tilde{x}}_{k+1}^j)\right)}$

Set up consistance degree of confidence matrix.

As shown in Figure 3, described step S3 specifically comprises the steps:

A, based on " judging inseparable goal rule ", " for i the peak value in k moment; if the difference that multiple peak values in k+1 moment and the consistance degree of confidence between it are all greater than between certain thresholding ε (0 < < ε < 1) and these consistance degree of confidence is less than certain threshold value η (0 < η≤0.001), these targets in k+1 moment belong to inseparable target.”。Set up inseparable object set (ITS).For certain a line (i is capable) of consistance degree of confidence matrix, if for j arbitrarily ₁=1 ..., n _k+1, j ₂=1 ..., n _k+1and j ₁≠ j ₂, have $\frac{\mathrm{\&mu;}({\tilde{x}}_k^i,{\tilde{x}}_{k+1}^{j_1})-\mathrm{\&mu;}({\tilde{x}}_k^i,{\tilde{x}}_{k+1}^{j_2})}{\mathrm{\&mu;}({\tilde{x}}_k^i,{\tilde{x}}_{k+1}^{j_1})}<\mathrm{\&eta;},\mathrm{\&mu;}({\tilde{x}}_k^i,{\tilde{x}}_{k+1}^{j_1})\&GreaterEqual;\mathrm{\&epsiv;}$ And $\mathrm{\&mu;}({\tilde{x}}_k^i,{\tilde{x}}_{k+1}^{j_2})\&GreaterEqual;\mathrm{\&epsiv;},$ J ₁, j ₂belong to IS (i);

B, based on " judging target homology rule ", " if the consistance degree of confidence between any two peak values in adjacent two moment is greater than certain thresholding ε (0 < < ε < 1); these two peak values come from same target, and these two peak values are same target peak values in adjacent two moment.”。Set up targetpath collection (TTS).For certain a line (i is capable) of consistance degree of confidence matrix, the situation that if there is no A sets, gets one of its element maximum and (and is greater than ε, even has if this element is to be also maximum, corresponding row element in its column (j row) with column element homology is correct associated and put it into corresponding flight path and concentrate by it; If this element is not to be maximum in its column, lie over this journey, continue the investigation of next line;

C, based on " judge that target disappear rule ", " consider decision-making window long be n+1 (n >=2).For a peak value in k moment, if any peak value in k+1 moment and the consistance degree of confidence between it are all less than η, this target is predicted.If the consistance degree of confidence between the peak value of k+1 this prediction of moment and any peak value in k+2 moment is all less than η and this situation lasts till moment k+n, its corresponding target disappearance of k moment.”。Set up disappearance object set (DTS).If for any m=1 ..., n, j _k+m+1=1 ..., n _k+m+1, have the target that i the peak value of moment k is corresponding disappears, and puts it into DTS, and its corresponding flight path terminates;

D, based on " judge target newborn rule ", " if the consistance degree of confidence between peak value in k+1 moment and arbitrary peak value in k moment is all less than ε, it is a newborn target.”。Set up newborn object set (BTS).For all i, if j peak value is newborn target, puts it into BTS, to its initial flight path.

Embodiment

Simulating scenes: acquisition probability is Pd=0.98, survival probability is Pe=0.99.Target is moved in two-dimensional space, and its state equation is

X(k+1)＝F(k)X(k)+v(k)

Wherein, state vector $X\left(k\right)={\left[\begin{array}{cccc}x& \stackrel{\&CenterDot;}{x}& y& \stackrel{\&CenterDot;}{y}\end{array}\right]}^{\text{'}},$ State-transition matrix $F\left(k\right)=\left[\begin{array}{cc}1& T\\ 0& 1\end{array}\right]\&CircleTimes;I_2,$ Process noise covariance is $Q={{\mathrm{\&Phi;}}_Q}^2\left[\begin{array}{cc}\frac{T^4}{4}& \frac{T^3}{2}\\ \frac{T^3}{2}& T^2\end{array}\right]\&CircleTimes;I_2,$ Wherein, the sampling time is T=1s, and process noise is white Gaussian noise and separate with measurement noise.Φ _q=0.25, I ₂it is 2 × 2 unit matrix.

Measurement equation is

Z(k)＝h[X(k)]+ω(k)

Wherein, Z (k)=[ρ (k) θ (k)] ', ω is white Gaussian noise. $h\left[X\left(k\right)\right]=\left[\begin{array}{c}\sqrt{x^2\left(k\right)+y^2\left(k\right)}\\ \arctan [y\left(k\right)/x\left(k\right)]\end{array}\right],$ Measurement noise covariance is σ _ρ=10, σ _θ=0.3 °.Clutter is evenly distributed on monitored area, and each detection moment clutter average is 50.When emulation, α=500, ε=0.8, η=0.001.

In this simulating scenes, comprise that target survives always, the situation that target new life and target disappear.It is X that there are three targets and target original state in initial monitored area ₁(0)=[500 50 100 10] ', X ₂(0)=[0 50 0 0] ', X ₃(0)=[500 50-100-10] '.In the time that emulation the 11st is clapped, newborn two targets of monitor area, two newborn target original states are respectively X ₄(11)=[400 50 100 20] ', X ₅(11)=[400 50-100-20] ', move to emulation and finish.In the time that emulation the 16th is clapped, target 3 disappears.

As shown in Figure 4, each target is all successfully traced into, and all corresponding to different flight paths, can find out that this invention can well stablize extraction flight path, obtains continuous flight path information; For newborn target, this invention can correct initial flight path; For the target disappearing, this invention flight path that can correctly terminate.

The above is only the preferred embodiment of the present invention; it should be pointed out that for those skilled in the art, under the premise without departing from the principles of the invention; can also make some improvements and modifications, these improvements and modifications also should be considered as protection scope of the present invention.

Claims (4)

1. the probability hypothesis density wave filter flight path extracting method based on overall temporal and spatial correlations, is characterized in that, comprises the following steps:

S1, probability hypothesis density wave filter dbjective state are extracted;

S2, consistency metric and consistance confidence calculations;

S3, overall flight path fetch strategy.

2. the probability hypothesis density wave filter flight path extracting method based on overall temporal and spatial correlations according to claim 1, is characterized in that, described step S1 specifically comprises the steps:

S11, prediction: for original target, sample and calculate the weights of predicting particle; For target that may be newborn, sample and calculate the weights of newborn particle;

S12, renewal: upgrade weights with measuring;

S13, resampling: calculate weights total amount rounding and obtain number of targets, resampling regulates weights.

3. the probability hypothesis density wave filter flight path extracting method based on overall temporal and spatial correlations according to claim 1, is characterized in that, described step S2 specifically comprises the steps:

S21, obtain the estimation peak set in adjacent two moment by step S1;

S22, estimation peak set based on previous moment in two moment, carry out one-step prediction, obtains the prediction peak set in next moment;

S23, calculate relative coordinate: calculate respectively in synchronization prediction peak set between 2 and estimate the directed line segment length between 2 in peak set;

Distance between S24, calculating relative coordinate;

Consistance degree of confidence between S25, calculating twice sweep point, sets up consistance degree of confidence matrix.

4. the probability hypothesis density wave filter flight path extracting method based on overall temporal and spatial correlations according to claim 1, is characterized in that, described step S3 specifically comprises the steps:

S31, based on " judging inseparable goal rule ", set up inseparable object set, extract inseparable target;

S32, based on " judging target homology rule ", set up targetpath collection, extract survival target flight path;

S33, based on " judge that target disappear rule ", set up disappearance object set, extract the target disappearing, flight path correctly terminates;

S34, based on " judge target newborn rule ", set up disappearance object set, extract newborn target, correct initial flight path.