CN105844217A - Multi-target tracking method based on measure-driven target birth intensity PHD (MDTBI-PHD) - Google Patents

Abstract

The invention discloses a multi-target tracking method based on measure-driven target birth intensity PHD (MDTBI-PHD), mainly solves the problems in estimating the multi-target motion state and the number of targets when the target birth intensity is unknown. The method comprises the steps of estimating the target state and the clutter state respectively by using an augmented state-space method, thereby avoiding the interference of unknown clutter to the target intensity estimation; constructing a target birth measure set, and estimating the target birth intensity by using a measure-driven method, thereby avoiding dependence on the priori knowledge of the target birth intensity; and implementing the above method using Gaussian mixture probability hypothesis density filter. The method has the advantage of being sensitive to the change in the number of targets, and meanwhile can reduce the computational complexity and significantly improve the tracking accuracy.

Description

A kind of based on measuring the PHD multi-object tracking method driving newborn target strength to estimate

Technical field

The invention belongs to target tracking domain, a kind of based on measuring the PHD driving newborn target strength to estimate Multi-object tracking method.

Background technology

Multiple target tracking is to utilize sensor measurement information to estimate the kinestates such as the position of multiple targets and speed The method of meter, this estimation procedure is typical filtering problem.Due to exist during multiple target tracking noise jamming, The challenges such as target extinction, target appearance, how to be judged by sensor measurement information the actual number of target with And the corresponding kinestate of each target is the difficult point of multiple target tracking problem.

Existing multiple target tracking algorithm mainly includes that multiple target tracking algorithm based on data association is assumed close with probability Degree (Probability Hypothesis Density, PHD) filtering algorithm two class.Wherein, PHD filtering algorithm due to Avoid the multiple target tracking dependence to data association, thus reduce computational complexity.This algorithm can be estimated simultaneously The number of target and state, be the study hotspot of current academia.When discussing the problem of implementation of PHD wave filter, The most representational two kinds of method for solving are sequence MonteCarlo probability hypothesis density (Sequential Monte Carlo Probability Hypothesis Density, SMC-PHD) and Gaussian-mixture probability hypothesis density (Gaussian Mixture Probability Hypothesis Density, GM-PHD).

In tradition PHD filtering algorithm, it was predicted that the newborn target strength that part comprises generally assumes that as known a priori. But, in a lot of actual application scenarios, due to newborn target dynamic change, and complex distribution, the most this hypothesis Bigger estimation difference can be brought, it may appear that " false-alarm " and " false dismissal " phenomenon to target.To unknown newborn mesh Mark intensity carries out On-line Estimation and is capable of the requirement close to actual application scenarios.

For the situation that newborn target strength is unknown, Yan little Xi et al. used Finite mixture model to new life in 2011 Target strength is modeled estimating, application Dirichlet is distributed as prior distribution.Although this method can exist in real time Newborn target strength estimated by line, but owing to needs carry out priori it is assumed that therefore still have certain with practical situation Gap；Han S T et al. used the thought of information feedback in 2014, examined while estimating newborn target strength Having considered the impact of conception of history measurement information, the newborn target strength information obtained is the most accurate.But have and cannot be distinguished by clutter With the limitation of dbjective state, bigger noise jamming can be brought when estimating.Simultaneously because use previous moment to measure letter Cease the newborn target strength to current time to estimate, thus exist when target numbers changes and estimate delayed asking Topic；Ristic B et al. used the newborn target strength method of estimation that measurement information drives in 2012, new by arranging The state strength of newborn target with target is predicted and updates by the mode of raw mark respectively, thus avoid right The dependence of priori.But newborn target is made a distinction the most merely by it with existing target, is not given for new life The recurrence formula form of target strength.

Summary of the invention

Technical problem solved by the invention is to provide a kind of based on measuring the PHD driving newborn target strength to estimate Multi-object tracking method (Measure-driven Target Birth Intensity PHD, MDTBI-PHD).

The technical solution realizing the object of the invention is: first, by measuring to drive, newborn target strength is carried out reality Time estimate, use the method in augmented state space to process dbjective state intensity and clutter state strength respectively simultaneously, thus Avoid interference when newborn target strength is estimated, dependence and the unknown clutter of priori caused, the most preferably Reduce owing to newborn target strength priori arranges improper brought error.Secondly, filter at GM-PHD On the basis of device, propose MDTBI-PHD algorithm filter and realize, realize relative to SMC-PHD wave filter, tool Have without cluster operation, without using the advantage of a large amount of particle, thus reduce the complexity of calculating.

The present invention realizes the technical method of above-mentioned purpose and comprises the following steps:

Step 1, the real goal intensity initializing the acquisition k-1 moment and noise intensity, be denoted as Weighted Coefficients Gaussian component form.Target strength D in k-1 moment_k-1|k-1X () can be expressed as the Gaussian component of Weighted CoefficientsNoise intensityCan be expressed as

Step 2, newborn target strength are estimated.Differ not according to the metric data produced due to survival target adjacent moment Greatly, and the position that newborn target occurred in the k moment typically and the position of k-1 moment survival target to have certain away from From, i.e. the newborn aim parameter measured value in k moment and the measuring value of k-1 moment survival target typically want difference relatively big, then K moment new life can be obtained and measure collection Z_k ^(newborn), then collected by measurement to be given respectively and target and clutter new life intensity estimated Meter.

Step 3, according to PHD predictor formula prediction real goal intensity and noise intensity.

Step 4, update real goal intensity and noise intensity according to PHD more new formula.

Step 5, according to pruning thresholding and the pruning item less with merging weights in Gaussian component that merge thresholding.

The present invention compared with prior art has the advantage that

1., for the situation that Clutter Model when utilizing PHD to estimate target strength is unknown, use augmented state space side Method, processes with clutter state respectively by true time dbjective state, and then avoids clutter and do target strength estimation Disturb.

2. proposing a kind of based on measuring the filtering algorithm driving newborn target strength to estimate, the method is in augmented state space Framework under application measure driving method the newborn target strength of target and clutter is estimated have target numbers The advantage of sensitive.

3. on the basis of Gaussian-mixture probability assumes density filtering, propose MDTBI-PHD filtered method, ensureing While the tracking accuracy of filtering algorithm, reduce computational complexity.

Below in conjunction with the accompanying drawings the present invention is described in further detail.

Accompanying drawing explanation

Fig. 1 is the flow chart of the inventive method.

Fig. 2 is target actual motion trajectory diagram.

Fig. 3 is target state estimator trajectory diagram.

Fig. 4 is target state estimator number figure.

Fig. 5 is OSPA distance comparison diagram.

Detailed description of the invention

According to accompanying drawing, technical scheme is illustrated.

The dbjective state in k moment and aim parameter measurement information are expressed as random manner collection X by the present invention_kWith random quantity Survey collection Z_k:

X_k={ x_{K, 1}, x_{K, 2}..., x_{K, i}..., x_{K, M (k)}}=S_k|k-1(X_k-1)∪B_k|k-1(X_k-1)∪Γ_k (1)

$Z_k=\{z_{k,1},z_{k,2},...,z_{k,j},...,z_{k,N\left(k\right)}\}=K_k\∪\[\underset{x\∈X_k}{\∪}{\mathrm{\Θ}}_k\left(x\right)\]---\left(2\right)$

Wherein, x_{K, i}Representing the state of k moment i-th target, M (k) represents the tracking target numbers in k moment observation area； z_{K, j}Represent k moment jth aim parameter measured value, the aim parameter detecting number that N (k) the expression k moment observes.S_k|k-1(X_k-1) The dbjective state set that expression k-1 moment to the k moment still survives, B_k|k-1(X_k-1) represent that the target in k-1 moment is when k Carve the fresh target state set derived, Γ_kThe fresh target state set that the expression k moment occurs；K_kRepresent that target is subject to Measurement noise set when clutter and false-alarm interference,Represent by dbjective state set X_kThe measurement collection produced Close.

Described based on measuring the PHD multi-object tracking method driving newborn target strength to estimate, comprise the following steps:

Step 1, the real goal intensity initializing the acquisition k-1 moment and noise intensity.

Target strength D in k-1 moment_k-1|k-1X () can be expressed as the Gaussian component of Weighted CoefficientsNoise intensityCan be expressed asThen have:

$D_{k-1|k-1}\left(x\right)=\underset{l=1}{\overset{J_{k-1}^t}{\Σ}}{\mathrm{\ω}}_{k-1,t}^{\left(l\right)}N\left(x\right|m_{k-1,t}^{\left(l\right)},P_{k-1,t}^{\left(l\right)})---\left(3\right)$

Wherein,WithRepresent the l component weights in the k-1 moment of target and clutter respectively,With Represent the l component Gaussian mean in the k-1 moment of target and clutter respectively,WithRepresent target respectively With the l component of clutter at the Gauss covariance in k-1 moment,WithRepresent that target is divided with clutter Gauss respectively Amount number.

Step 2, estimation real goal and clutter new life intensity.

The metric data produced due to survival target adjacent moment is more or less the same, and the position that newborn target occurred in the k moment Put general with position a certain distance to be had of k-1 moment survival target, i.e. the newborn aim parameter measured value in k moment and The measuring value of k-1 moment survival target typically wants difference relatively big, then can remember that k moment new life measures collection and is:

Z_k ^(newborn)={ z_{I, k}|z_{I, k}∈Z_k, | | z_{I, k}-Z_k-1| | ＞ ε } (5)

Here,d(z_{I, k}, z_{J, k-1}) represent z_{I, k}With z_{J, k-1}Euclidean distance, ε is Threshold value, below by measuring collection Z_k ^(newborn)Be given respectively γ_k(x) andEstimation.

Because target possibly be present at the optional position of monitoring region S, so, obtaining newborn target metric data Z_k ^(newborn)Before, it was predicted that newborn target is evenly distributed on S, i.e.Being rational, | S | is here Monitored area S estimates.In like manner,Obtained by formula (5),

Now,It is exactly to measure under data-driven in newborn target, to target new life intensity γ_kThe possibility predication of (x), I.e.

In like manner, clutter new life intensityIt is represented by

Step 3, prediction real goal intensity and noise intensity.

Target prediction intensity is:

$D_{k|k-1}\left(x\right)={\mathrm{\γ}}_k\left(x\right)+\underset{l=1}{\overset{J_{k-1}^t}{\Σ}}{\mathrm{\ω}}_{k|k-1,t}^{\left(l\right)}N\left(x\right|m_{k|k-1,t}^{\left(l\right)},P_{k|k-1,t}^{\left(l\right)})---\left(9\right)$

Wherein,

$m_{k|k-1,t}^{\left(l\right)}={\mathrm{Fm}}_{k-1,t}^{\left(l\right)}---\left(10\right)$

$P_{k|k-1,t}^{\left(l\right)}=Q+{\mathrm{Fm}}_{k-1,t}^{\left(l\right)}{F}^T---\left(11\right)$

${\mathrm{\ω}}_{k|k-1,t}^{\left(l\right)}=p_s{\mathrm{\ω}}_{k-1,t}^{\left(l\right)}---\left(12\right)$

Wherein, F represents state-transition matrix, F^TThe transposition of representing matrix F.

Clutter predicted intensity is:

Wherein,

$m_{k|k-1,c}^{\left(l\right)}={\mathrm{Fm}}_{k-1,c}^{\left(l\right)}---\left(14\right)$

$P_{k|k-1,c}^{\left(l\right)}=Q+{\mathrm{Fm}}_{k-1,c}^{\left(l\right)}{F}^T---\left(15\right)$

${\mathrm{\ω}}_{k|k-1,c}^{\left(l\right)}=p_s^0{\mathrm{\ω}}_{k-1,c}^{\left(l\right)}---\left(16\right)$

Step 4, renewal real goal intensity and noise intensity.

Target update intensity:

$D_{k|k}\left(x\right)=(1-p_{D,k})D_{k|k-1}\left(x\right)+\underset{z\∈Z_k}{\Σ}{D}_{D,k}(x;z)---\left(17\right)$

Wherein,

$D_{D,k}(x;z)=\underset{l=1}{\overset{J_{k-1}^t}{\Σ}}{\mathrm{\ω}}_{k,t}^{\left(l\right)}N\left(x\right|m_{k,t}^{\left(l\right)},P_{k,t}^{\left(l\right)})---\left(18\right)$

Here,

${\mathrm{\ω}}_{k,t}^{\left(l\right)}=\frac{p_{D,k}{\mathrm{\ω}}_{k|k-1,t}^{\left(l\right)}N\left({\hat{z}}_{k|k-1,t}^{\left(l\right)},S_{k|k-1,t}^{\left(l\right)}\right)}{L\left(z\right)}---\left(19\right)$

$m_{k,t}^{\left(l\right)}=m_{k|k-1,t}^{\left(l\right)}+K_k^{\left(l\right)}(z-{\hat{z}}_{k|k-1,t}^{\left(l\right)})---\left(20\right)$

$P_{k,t}^{\left(l\right)}=\[I-K_k^{\left(l\right)}H\]P_{k|k-1,t}^{\left(l\right)}---\left(21\right)$

${\hat{z}}_{k|k-1,t}^{\left(l\right)}={\mathrm{Hm}}_{k|k-1,t}^{\left(l\right)}---\left(22\right)$

$S_{k|k-1,t}^{\left(l\right)}={\mathrm{HP}}_{k|k-1,t}^{\left(l\right)}{H}^T+R---\left(23\right)$

$K_k^{\left(l\right)}=P_{k|k-1,t}^{\left(l\right)}{H}^T{\[S_{k|k-1,t}^{\left(l\right)}\]}^{-1}---\left(24\right)$

$L\left(z\right)=p_{D,k}^0{\mathrm{\ω}}_{k|k-1,c}^{\left(l\right)}N({\hat{z}}_{k|k-1,c}^{\left(l\right)},S_{k|k-1,c}^{\left(l\right)})+\underset{l=1}{\overset{J_{k-1}^t}{\mathrm{\Σ}}}{p}_{D,k}{\mathrm{\ω}}_{k|k-1,t}^{\left(l\right)}N({\hat{z}}_{k|k-1,t}^{\left(l\right)},S_{k|k-1,t}^{\left(l\right)})---\left(25\right)$

${\hat{z}}_{k|k-1,c}^{\left(l\right)}={\mathrm{Hm}}_{k|k-1,c}^{\left(l\right)}---\left(26\right)$

$S_{k|k-1,c}^{\left(l\right)}={\mathrm{HP}}_{k|k-1,c}^{\left(l\right)}{H}^T+R---\left(27\right)$

Wherein, I representation unit matrix；H represents observation model matrix；Represent Kalman gain；With Represent the Gauss covariance matrix of target and clutter respectively；WithRepresent that target measures with clutter prediction respectively Value；R represents process noise covariance matrix.

Clutter renewal intensity:

Wherein,

Wherein,

${\mathrm{\ω}}_{k,c}^{\left(l\right)}=\frac{p_{D,k}^0{\mathrm{\ω}}_{k|k-1,c}^{\left(l\right)}N({\hat{z}}_{k|k-1,c}^{\left(l\right)},S_{k|k-1,c}^{\left(l\right)})}{L\left(z\right)}---\left(30\right)$

$m_{k,c}^{\left(l\right)}=m_{k|k-1,c}^{\left(l\right)}+K_k^{\left(l\right)}(z-{\hat{z}}_{k|k-1,c}^{\left(l\right)})---\left(31\right)$

$P_{k,c}^{\left(l\right)}=\[I-K_k^{\left(l\right)}H\]P_{k|k-1,c}^{\left(l\right)}---\left(32\right)$

$K_k^{\left(l\right)}=P_{k|k-1,c}^{\left(l\right)}{H}^T{\[S_{k|k-1,c}^{\left(l\right)}\]}^{-1}---\left(33\right)$

Step 5, prune and merge Gaussian component relatively event.

In order to reduce the number of Gaussian component, and then the complexity of reduction algorithm, can be by arranging pruning thresholding Except the item that weights in the Gaussian component after updating are less, merge, by arranging, the item that thresholding combined distance is closer to.

Below the method for the present invention is carried out simulating, verifying,

The method that emulation uses Gaussian-mixture probability to assume that density filter (GM-PHD) and the present invention propose compares Relatively, the tracking performance of the inventive method is described under unknown newborn target strength environment.

The motion model of target is:

X_k=FX_k-1+Gω_k (34)

Wherein, X_k-1={ x_x, v_x, x_y, v_y, representing the Position And Velocity component in x, the y direction of target, F is that state shifts square Battle array, G is noise inputs matrix, and both are respectively as follows:

$F=\left[\begin{array}{cccc}1& T& 0& 0\\ 0& 1& 0& 0\\ 0& 0& 1& T\\ 0& 0& 0& 1\end{array}\right]---\left(35\right)$

$G={\left[\begin{array}{cccc}{T}^2/2& T& 0& 0\\ 0& 0& T^2/2& T\end{array}\right]}^T---\left(36\right)$

Wherein, T represents the sampling period；Process noise ω_kFor white Gaussian noise.

The observational equation of system is:

Z_k=HX_k+V_k (37)

Wherein, H is observing matrix；V_kFor process noise matrix, meet Gauss distribution.

$H=\left[\begin{array}{cccc}1& 0& 0& 0\\ 0& 0& 1& 0\end{array}\right]---\left(38\right)$

Optimum subpattern is used to assign (Optimal Subpattern Assignment, OSPA) distance as tracking accuracy Evaluation index, its expression-form is:

${\stackrel{\‾}{d}}_p^{\left(c\right)}(X,\hat{X})={\left(\frac{1}{n}\right(\underset{\mathrm{\π}\∈{\mathrm{\Π}}_n}{min}\underset{i=1}{\overset{m}{\Σ}}{d}^{\left(c\right)}{\left(x_i,{\hat{x}}_{\mathrm{\π}\left(i\right)}\right)}^p+c^p\left(n-m\right)\left)\right)}^{\frac{1}{p}}---\left(39\right)$

Wherein,Represent OSPA distance；N, m be respectively X withDimension；C represents interrupt threshold；p The exponent number of representing matrix.

The distance parameter p=2 of OSPA, c=5, sensor detection probability p_d=0.98, target survival probability p_s=0.99, It is 10 that Gauss prunes thresholding^-5, it is 4 that Gauss merges thresholding, and in GM-PHD and the inventive method, population is set to 500, Carry out 50 Monte Carlo experiments altogether.Selecting 4 moving targets to emulate, wherein first aim is at 1s Time occur, during 80s disappear；Second target occurs when 20s, disappears during 70s；3rd target goes out when 61s Existing, disappear during 70s；4th target occurs when 46s, disappears during 90s.Experimental result is as Figure 2-Figure 5.

Fig. 2 Yu Fig. 3 respectively show the actual path of target and estimate track.In Fig. 2,4 curves represent 4 respectively The actual path of individual moving target, the estimation position of particle point trace description target in Fig. 3.According to simulation result, The pursuit path that the inventive method calculates can be preferably identical with the actual path of multiple targets.

Fig. 4 represents target true number and target state estimator number.It can be seen that be the unknown in noise intensity In the case of, the method that the present invention proposes is higher to the accuracy of estimation of target numbers for GM-PHD, There is the advantage to target numbers sensitive simultaneously.

Fig. 5 represents target OSPA distance under GM-PHD and the inventive method, as can be seen from the figure this Bright method is less compared with the OSPA of GM-PHD distance, illustrates that the method is estimated and in tracking accuracy at multiple target number Accuracy higher than GM-PHD, it was demonstrated that the effectiveness of the inventive method.

In sum, the PHD multi-object tracking method energy driving newborn target strength estimation based on measurement of the present invention Enough being the most effectively tracked multiple targets during newborn target strength the unknown, have target numbers sensitive is excellent Gesture, and reduce computation complexity, there is the most superior tracking accuracy.The present invention drives newborn mesh based on measuring Have in terms of the PHD multi-object tracking method that mark intensity the is estimated multiple target tracking when processing newborn target strength the unknown There is positive meaning.

Claims (4)

1. the PHD multi-object tracking method driving newborn target strength to estimate based on measurement, it is characterised in that comprise the following steps:

The first step, initializes real goal intensity and the noise intensity obtaining the k-1 moment, is denoted as the Gaussian component shape of Weighted Coefficients Formula.

Second step, newborn target strength is estimated.Metric data according to producing due to survival target adjacent moment is more or less the same, and newborn The position that target occurred in the k moment is general and position a certain distance to be had of k-1 moment survival target, the i.e. newborn target in k moment Measuring value and the measuring value of k-1 moment survival target typically want difference relatively big, then can obtain k moment new life and measure collection Z_k ^(newborn), then Collected by measurement and be given respectively target and the estimation of clutter new life intensity.

3rd step, according to PHD predictor formula prediction real goal intensity and noise intensity.

4th step, updates real goal intensity and noise intensity according to PHD more new formula.

5th step, prunes the item less with merging weights in Gaussian component according to pruning thresholding with merging thresholding.

The most according to claim 1 based on measuring the PHD multi-object tracking method driving newborn target strength to estimate, its feature exists Collecting in described newborn measurement, specifically, k moment new life measures collection and is:

Z_k ^(newborn)={ z_{I, k}|z_{I, k}∈Z_k, | | z_{I, k}-Z_k-1| | ＞ ε } (1)

Wherein, z_{I, k}Represent k moment i-th aim parameter measured value, Z_kRepresent and measure collection at random, d(z_{I, k}, z_{J, k-1}) represent z_{I, k}With z_{J, k-1}Euclidean distance, ε is threshold value.

It is the most according to claim 1 based on measuring the PHD multi-object tracking method driving newborn target strength to estimate, it is characterised in that Described newborn target strength calculating formula, specifically, k moment new life target strength γ_k(x) be:

Wherein, | S | is estimating of monitored area S, p_{D, k}X () represents acquisition probability,Represent clutter detection probability, g_k(z |) table Show joint likelihood function.

The most according to claim 1 based on measuring the PHD multi-object tracking method driving newborn target strength to estimate, its feature exists In described clutter new life intensitometer formula, specifically, k moment new life target strengthFor: