Specific embodiment
Some vocabulary is used in specification and claims to censure specific component, the skill in fields
Art personnel are, it is to be appreciated that manufacturer may call same component with different nouns.Present specification and claims
Not in a manner that the difference of title is used as and distinguishes component, but it is used as the base of differentiation with the difference of component functionally
It is accurate.The present invention is described in detail with reference to the accompanying drawings and examples.
In many practical engineering applications such as signal processing, communication, radar, sonar, through coming frequently with dynamic space model
Many of which problem is described.Its model is represented by state equation (1) and observational equation (2):
xk=fk(xk-1)+vk (1)
zk=hk(xk)+ek (2)
In formula,WithNonlinear function known to representing respectively;Represent the system mode vector at k moment;It is the observation at k moment;WithRespectively
For process noise and observation noise, and it is independent from each other.
State estimation is exactly in given observation data acquisition system z1:kUnder conditions of estimated state vector xkProbability density function
p(xk|z1:k-1).It is assumed that in k-1 moment probability density function p (xk-1|z1:k-1) it is known that then predicting the original prior probability at k moment
Density function is:
p0(xk|z1:k-1)=∫ p (xk|xk-1)p(xk-1|z1:k-1)dxk-1
Wherein, p0(xk|z1:k-1) represent original priori probability density function, p (xk|xk-1) represent state transition probability density
Function, p (xk-1|z1:k-1) represent k-1 moment probability density functions.
Observation z is obtained in moment kkAfterwards, according to Bayes rule, posterior probability density function can be defined as follows:
p(xk|z1:k)=p (zk|xk)p0(xk|z1:k-1)/p(zk|z1:k-1)
Wherein, p (xk|z1:k) represent posterior probability density function, p0(xk|z1:k-1) represent original priori probability density letter
Number, p (zk|z1:k-1) it is regarded as a constant, p (zk|xk) represent observation likelihood function, it can be by observation model (2) and observation noise
ekIt obtains.
Fig. 1 is that the auxiliary of the embodiment of the present invention blocks the flow chart of particle filter method.If it is noted that have substantially
It is identical as a result, the method for the present invention is not limited with flow shown in FIG. 1 sequence.As shown in Figure 1, this method includes following step
Suddenly:
Step S11:Using original priori probability density function as the first the importance density function carry out particle filter with
Obtain the first mean value corresponding with dbjective state and the first covariance value.
In step s 11, by the use of original priori probability density function as the first the importance density function to dbjective state
It is updated.Please also refer to Fig. 2, Fig. 2 is by the use of original priori probability density function as the first importance density letter in Fig. 1
Number carries out particle filter to obtain the flow chart of the first mean value corresponding with dbjective state and the first covariance value.As shown in Fig. 2,
The flow chart specifically comprises the following steps:
Step S111:The first particle collection is extracted from the first the importance density function.
In step S111, according to state equation (1), N is extracted from the first the importance density functionsA particle forms the
One particle collection, wherein, the first importance density function is by original priori probability density function p0() builds, wherein, the first particle
The value for i-th of the particle concentrated is expressed asWherein, i=1,2,3 ..., NS。
Step S112:It obtains the first particle and concentrates corresponding first weights of each particle.
In step S112, with reference to observational equation (2), the first weights of each particle are obtained according to equation below:
Wherein,Represent corresponding first weights of i-th of particle, zkRepresent the observation at k moment.
Step S113:Each first weights are standardized.
In step S113, the step of being standardized to each first weights, is specially:First to the first particle
Corresponding first weights of each particle is concentrated to carry out summation process, then by corresponding first weights of each particle and each first power
Value carries out the value after summation process and is divided by, wherein, first weights that are divided by that treated are first after standardization
Weights.
Specifically, each first weights are standardized according to equation below:
Wherein,Represent corresponding first weights of i-th of particle,It represents at the corresponding standardization of i-th of particle
The first weights after reason, NsRepresent that the first particle concentrates the sum of particle.
Step S114:According to the first weights after standardization and particle corresponding with the first weights obtains and target
Corresponding first mean value of state and the first covariance value.
In step S114, corresponding first mean value of dbjective state and the first covariance value are obtained according to equation below
It arrives:
Wherein,Represent corresponding first mean value of dbjective state,It represents at the corresponding standardization of i-th of particle
The first weights after reason,Represent the value of i-th of particle,Represent corresponding first covariance value of dbjective state.
Step S12:In original priori probability density function, using blocking, theory introduces Current observation information and target is special
Property information with build correct priori probability density function.
In step s 12, Current observation information and mesh are introduced in original priori probability density function based on blocking theory
Mark characteristic information corrects priori probability density function to build.
Specifically, in particle filter, when object module exist it is uncertain, particle filter due to the accumulation of error,
Filtering performance can decline, and to solve the problems, such as this, the method that the auxiliary of the present invention blocks particle filter (ATPF) may be used.
Auxiliary blocks particle filter and is limited to following two primary conditions:
Condition 1:Nonlinear function h in observational equation (2)k() is continuous dijection;
Condition 2:The probability density function bounded unicom of observation noise, shown in formula specific as follows:
Wherein,Represent nzTie up unicom region, ekFor observation noise.
According to condition 2, the observation likelihood function p (z in formula (4)k|xk) can be defined as follows:
Wherein, p (zk|xk,rk) it is based on blocking the observation likelihood function after theoretical treatment, xkRepresent the system shape at k moment
State vector, zkRepresent the observation at k moment, hkNonlinear function known to () expression,Represent regionOn finger
Show function, rkIt represents comprising c objective attribute target attributeTarget signature scalar.
Meanwhile as the observation z at k momentkWith target signature scalar rkWhen uncorrelated, according to condition 1, formula (8) can become
It is changed to:
Wherein,
Then, using Bayes rule, posterior probability density function can be defined as:
Wherein, ε1For generalized constant.As can be seen that correcting priori probability density function p from formula (12)new(xk|
zk,xk-1,r1:k) introduce current observation information zkWith target signature information rk, therefore, when observation noise variance ratio is relatively low,
Correct priori probability density function pnew(xk|zk,xk-1,r1:k) original priori probability density function p can be effectively reduced0(xk|
xk-1,z1:k-1) variance improve filtering estimation performance.
Therefore, the performance for enhancing particle filter, is blocked in auxiliary in particle filter, and joint corrects priori probability density letter
Number pnew() and original priori probability density function p0(), the importance density function q (xk|x0:k-1,z1:k,r1:k) be defined as follows:
q(xk|x0:k-1,z1:k,r1:k)=αkpnew(xk|zk,xk-1,r1:k)+(1-αk)p0(xk|xk-1,z1:k-1)
=αkp(xk|xk-1,r1:k)χIxk(zk)(xk)+(1-αk)p0(xk|xk-1,z1:k-1)
Wherein, αk∈ [0,1] be variable element, behind will provide definition method.
According to the above discussion, in order to the importance density function q (xk|x0:k-1,z1:k,r1:k) sampled, using two
Particle filter updates dbjective state, the original priori probability density function p of a particle filter0() is important as first
Property density function carry out particle filter estimation update, such as step S11;One particle filter amendment priori probability density function
pnew() carries out particle filter estimation update as the second the importance density function.To correct priori probability density function pnew
() is carried out as the second the importance density function in the implementation process of particle filter, due to calculatingInvolve integration,
Correct priori probability density function pnewThe implementation of () is often infeasible, while only to making an uproar during truncation
Sound has carried out truncation, only to correcting priori probability density function pnewIn () with system mode vector xkIn position have
The part of relationship has an impact, therefore, as system mode vector xkIt is defined as follows:
xk=[ak T,bk T]T (13)
Wherein,Represent system mode vector xkMiddle location components,Represent system mode vector xkIn
Velocity component, and nx=na+nb。
Meanwhile if original priori probability density function p0() is that mean value isCovariance isGaussian function,
And
Wherein,Represent original priori probability density function p0The state component a of target in ()kCorresponding mean value;Represent original priori probability density function p0The state component b of target in ()kCorresponding mean value;Represent original elder generation
Test probability density function p0The state component a of target in ()kCorresponding covariance;Represent original priori probability density letter
Number p0The state component b of target in ()kCorresponding covariance;Represent original priori probability density function p0Mesh in ()
Target state component cross covariance corresponding with velocity component.
It obtains and corrects priori probability density function pnew() is limited to following three condition:(1) observation function hk() is
Local linear;(2) the state component a of targetkEdge prior probability density p0(ak) in regionIt is constant;(3) it sees
Survey noise ekMeetModified noise and real noise have an identical second moment, E [ek]=0 and cov [ek]
=Rk。
According to above three condition, modified priori probability density function pnew() can be approximated as mean value
Covariance isGaussian probability-density function, it is specific as follows:
Wherein,It represents to correct priori probability density function pnewThe state component a of target in ()kCorresponding mean value;It represents to correct priori probability density function pnewThe state component b of target in ()kCorresponding mean value;It represents to correct
Priori probability density function pnewThe state component b of target in ()kCorresponding covariance;Represent that amendment prior probability is close
Spend function pnewThe state component a of target in ()kCorresponding covariance;It represents to correct priori probability density function pnew
The state component a of target in ()kWith bkCorresponding covariance.
Represent akMaximal possibility estimation.As observation function hkDuring () local linear,At this momentRepresent non-thread observation functionJacobian matrix, andJacobi square
Battle array Value.
(such as in Pure orientation maneuvering target tracking) in practical applications, the nonlinear function h in observational equation (2)k
() has nonlinearity, is unsatisfactory for the requirement of bijection, therefore, in the present embodiment, utilizes least square positioning side
Method introduces the maximal possibility estimation that target property carrys out combined calculation targetSo that it is determined that correct priori probability density function
pnew(·)。
Please also refer to Fig. 3, Fig. 3 is to introduce Current observation information and mesh in Fig. 1 in original priori probability density function
Mark characteristic information corrects the flow chart of priori probability density function to build.As shown in figure 3, the flow chart specifically includes following step
Suddenly:
Step S121:Position location and the positioning of target are obtained according to Current observation use of information least square localization method
Variance.
In step S121, least square localization method is specific as follows shown:
N observation information of N number of passive sensor is received simultaneously when the k momentWhen, wherein, θiRepresent orientation
Angle, βiRepresent pitch angle, the position of passive sensor i is (xi,yi,zi), i=1,2 ... N.It can by corresponding geometric knowledge
Know:Angle information (the θ of target that each passive sensor measuresi、βi) it can determine the position line in a space, it is missed in observation
In the case that difference is zero, N position line is met at a bit, which is the position of target.But due to being seen during actual observation
Error is surveyed generally to be not zero, therefore above-mentioned N position line can't be met at a bit.In this regard, can will with a distance from N position line and
As target state estimator position, the process using the position of least square Cross Location Method estimation target is shortest point:
If LiRepresent the position line obtained by passive sensor i, T (xT,yT,zT) be target position, Ai(xi0,yi0,zi0)
Represent target to position line LiIntersection point, then position line LiFormula be:
Wherein, (li,mi,ni) represent position line LiDirection cosines, and
li=sin βicosαi, mi=sin βisinαi, ni=cos βi
According to geometrical relationship and corresponding mathematic(al) manipulation, target is relative to N position line LiSquare distance and can be with table
It is shown as:
In formula,
Above formula is the least-squares estimation value of target location, and non trivial solution is as follows shown in above formula:
D=LMN+2TRS-S in above formula2M-R2L-T2N
The position location of targetIt can be calculated by formula (23).Position location varianceIt may be calculated as:
Step S122:According to mesh described in original priori probability density function, position location and target property acquisition of information
The corresponding maximum likelihood estimator of target location components.
In step S122, if using position locationInstead of maximum likelihood estimatorWork as although introducing
Preceding observation information still cannot effectively solve the degradation problem that object module uncertainty is brought.For this purpose, in the present embodiment
In, target property such as target velocity v, target observation time interval T and target course θ are introduced to maximum likelihood estimatorIn the middle, so as to improve probabilistic processing capacity to target movement model.
Based on this, by anchor pointRegard newest target observation, modified maximal possibility estimation as
Value is acquired according to by equation below:
Wherein,Represent maximum likelihood estimator,Represent original priori probability density function p0Target in ()
State component akCorresponding mean value,For the position location of target, λ is a constant,Represent observation noise variance,Represent new breath covariance, T represents target observation time interval, and v represents target velocity.
Step S123:It is obtained according to maximum likelihood estimator, positioning variances and corrects priori probability density function.
In step S123, according to formula (24) and (25), in formula (18)WithIt can be with approximate calculation such as
Under:
Wherein,It represents to correct priori probability density function pnewThe state component a of target in ()kCorresponding mean value,It represents to correct priori probability density function pnewThe state component a of target in ()kCorresponding covariance,It represents
Maximum likelihood estimator,Represent position location variance.
The technical staff of ability is appreciated that due to correcting priori probability density function pnew() is Gaussian function, onceWithIt determines, formula (16) and formula (17) are assured that, so as to correct priori probability density function pnew(·)
Mean value, which can be approximately, isCovariance isGaussian probability-density function.
Step S13:Using correct priori probability density function as the second the importance density function carry out particle filter with
Obtain the second mean value corresponding with dbjective state and the second covariance value.
In step s 13, by the use of correcting priori probability density function as the second the importance density function to dbjective state
It is updated.Please also refer to Fig. 4, Fig. 4 is by the use of correcting priori probability density function as the second importance density letter in Fig. 1
Number carries out particle filter to obtain the flow chart of the second mean value corresponding with dbjective state and the second covariance value.As shown in figure 4,
The flow chart specifically comprises the following steps:
Step S131:The second particle collection is extracted from the second the importance density function.
In step S131, N is extracted from the second the importance density functionsA particle forms the second particle collection, wherein, the
Two the importance density functions are by amendment priori probability density function pnew() builds, wherein, i-th of the second particle concentration
The value of son is expressed asWherein, i=1,2,3 ..., NS。
Step S132:It obtains the second particle and concentrates corresponding second weights of each particle.
In step S132, obtain the step of the second particle concentrates each particle corresponding second weights and include:Using repairing
Positive priori probability density function build the second the importance density function namely:
qnew(xk|x0:k-1,z1:k,r1:k)=pnew(xk|zk,xk-1,r1:k) (28)
Wherein, pnew(xk|zk,xk-1,r1:k) represent to correct priori probability density function, qnew(xk|x0:k-1,z1:k,r1:k) table
Show the second the importance density function.
Then, according to formula (11) and formula (28), the second weights can be defined as follows:
Wherein,Represent second weights at k moment,Represent second weights at k-1 moment,It represents
Observe likelihood function,Represent the posterior probability density function at k moment,Represent k-1
The posterior probability density function at moment,Represent the third the importance density function at k moment,Represent the third the importance density function at k-1 moment, pnew(xk|zk,xk-1,r1:k) represent to correct first
Test probability density function, qnew(xk|x0:k-1,z1:k,r1:k) represent the second the importance density function.
It will be understood by those skilled in the art that the second weightsWith the second weightsOnly write using different
Method, the two represent identical concept.
Step S133:Each second weights are standardized.
In step S133, the step of being standardized to each second weights, is specially:First to the second particle
Corresponding second weights of each particle is concentrated to carry out summation process, then by corresponding second weights of each particle and each first power
Value carries out the value after summation process and is divided by, wherein, second weights that are divided by that treated are second after standardization
Weights.
Specifically, each second weights are standardized according to equation below:
Wherein,Represent corresponding second weights of i-th of particle,It represents at the corresponding standardization of i-th of particle
The second weights after reason, NsRepresent that the second particle concentrates the sum of particle.
Step S134:According to the second weights after standardization and particle corresponding with the second weights obtains and target
Corresponding second mean value of state and the second covariance value.
In step S134, corresponding second mean value of dbjective state and the second covariance value are obtained according to equation below
It arrives:
Wherein,Represent corresponding second mean value of dbjective state,It represents at the corresponding standardization of i-th of particle
The second weights after reason,Represent the value of i-th of particle,Represent corresponding second covariance value of dbjective state.
Step S14:According to Target state estimator weights respectively to the first mean value and the second mean value, the first covariance value and
Two covariance values are weighted processing to obtain posterior probability density function corresponding with dbjective state, complete particle filter mistake
Journey.
In step S14, Target state estimator weights are acquired according to equation below:
Wherein, akRepresent Target state estimator weights, zkRepresent Current observation value;hkNonlinear riew known to () expression
Survey function;WithIt represents to carry out particle as the first the importance density function using priori probability density function respectively
Filter obtained the first mean value and the first covariance;WithIt represents to utilize respectively and corrects priori probability density function work
The second mean value and the second covariance value that particle filter obtains are carried out for the second the importance density function.
The corresponding mean value of posterior probability density function and covariance value are acquired according to equation below:
Wherein,Represent the corresponding mean value of posterior probability density function, PkRepresent the corresponding association of posterior probability density function
Variance, akRepresent Target state estimator weights,WithIt is represented respectively by the use of priori probability density function as the first weight
The property wanted density function carries out the first mean value and the first covariance that particle filter obtains;WithIt represents to utilize respectively to repair
Positive priori probability density function carries out the second mean value and the second association side that particle filter obtains as the second the importance density function
Difference.
The particle filter method of the present invention blocks particle filter (ATPF) for auxiliary, below will be with two examples to the present invention
The performances of ATPF methods assessed.First example is single argument non-stationary model of growth (UNGM), will be with extending karr
Graceful filtering (EKF), Unscented kalman filtering (UKF), particle filter (PF), spreading kalman particle filter (PF-EKF) and without mark
Particle filter (UPF) scheduling algorithm is compared;Second example is Pure orientation maneuvering target tracking, will be extended with Interactive Multiple-Model
Kalman filtering (IMMEKF) and Interactive Multiple-Model Rao-Blackwellized particle filters (IMMRBPF) carry out performance comparison.
First example --- single argument non-stationary model of growth (UNGM):
State equation and the observational equation difference of single argument non-stationary model of growth are as follows:
Wherein, vkRepresent the non-Gaussian noise of obedience Gamma distribution Γ (2,3), ekRepresent the height that zero-mean variance is 0.01
This noise, α=0.5, β=25, γ=8, φ1=0.2, φ2=0.5.All experiments carry out 100 Monte-Carlo Simulations, state
Practical initial value x0Being uniformly distributed between obedience [0,1].All particle filters use population as 1000.
Root-mean-square error comparison diagrams of the Fig. 5 for tri- kinds of filtering methods of EKF, UKF and PF-EKF, Fig. 6 UPF, PF and ATPF
The root-mean-square error comparison diagram of three kinds of filtering methods.Complex chart 5 and Fig. 6 can be seen that:The filtering performance of ATPF, PF and UPF will
It is much better than the filtering performance of EKF, UKF, PF-EKF, meanwhile, the filtering performance of ATPF is better than the filtering performance of PF and UPF.
Please also refer to table 1, table one gives root-mean-square error and the calculating of EKF, UKF, PF, PF-EKF, UPF and ATPF
Time.
As can be seen from Table I, the filtering performance of ATPF is better than other all algorithms, on operation time, the fortune of EKF
Evaluation time is most short, but performance is worst, and the operation time of operation time of ATPF in all particle filter algorithms is most short, and performance is most
It is good.
Second example --- orientation maneuvering target tracking:
In this example, it will be verified using a collection of radar track data of actual acquisition come the ATPF to the present invention.Boat
Mark data include 40 aperiodic track points, flight time 107s.Due to the aperiodicity of track points, so between sampling
It is also variation, and the time interval of some points reaches more than 30s every T=t (k+1)-t (k), wherein k represents sampling number, t
(k+1) time during k+1 sampling is represented, t (k) represents time during k sampling.In this example, using following target with
Track model:
zk=hk(xk)+ek;
Wherein, dbjective state vector isxk、ykThe position of k moment targets is represented respectively,WithRepresent k moment targets in x respectivelyk、ykSpeed on direction;Process noise vk~N (0, Q), wherein, Q=diag
([0.012km2s4 0.012km2s4]).Observation noise ek~N (0, R), wherein R=diag ([0.152km2 0.152km2]),
(si,x,si,y,si,z), i=1,2 represents the position of two passive sensors respectively.The position of passive sensor observation station 1 for (0,
5km, 0), the position of passive sensor observation station 2 is (0, -5km, 0), and population is 200 in all particle filter methods.
Schematic diagrames of the Fig. 7 for the real motion track of target, the mesh that Fig. 8 ATPF, IMMRBPF two kind filtering method export
Target estimated motion track schematic diagram.
Fig. 9 is the root-mean-square error comparison diagram of two kinds of filtering methods of ATPF, IMMRBPF, wherein, Fig. 9 includes Fig. 9 A, figure
9B, Fig. 9 C and Fig. 9 D, Fig. 9 A are X-direction root-mean-square error, and Fig. 9 B are Y-direction root-mean-square error, and Fig. 9 C are missed for Z-direction root mean square
Difference, Fig. 9 D are position root-mean-square error.
Can be seen that the tracking performance of ATPF from Fig. 7, Fig. 8 and Fig. 9 will be significantly better than IMMRBPF, main reason is that
Information is (such as when ATPF can introduce the sky of current goal observation information and target when building the importance density function:When
Between interval, speed etc.), improve the sampling accuracy of particle, mould moved so as to handle the target brought due to target maneuver
The uncertain problem of type.
Figure 10 is the time comparison diagram of two kinds of filtering methods of ATPF, IMMRBPF.From fig. 10 it can be seen that with population
Increase, two methods calculate the time all with increase, but simultaneously, it can be seen that the calculating time of ATPF will be significantly lower than
IMMRBPF。
The structure diagram that the auxiliary that 1, Figure 11 is the embodiment of the present invention blocks particle filter device is please referred to Fig.1, is such as schemed
Shown in 11, which includes:
First acquisition module 21, for being carried out using original priori probability density function as the first the importance density function
Particle filter is to obtain the first mean value corresponding with dbjective state and the first covariance value.Specifically, the first acquisition module 21
For extracting the first particle collection from the first the importance density function first, then obtain the first particle and each particle is concentrated to correspond to
The first weights, then each first weights are standardized, finally according to the first weights after standardization with
And particle corresponding with the first weights obtains the first mean value corresponding with dbjective state and the first covariance value.
Priori probability density function structure module 22 is corrected, in original priori probability density function, using blocking
Theory introduces Current observation information and target property information and corrects priori probability density function to build.Specifically, it corrects first
Probability density function structure module 22 is tested for obtaining target according to Current observation use of information least square localization method first
Position location and positioning variances, obtained then according to original priori probability density function, position location and target property information
The corresponding maximum likelihood estimator of location components of target is taken, is finally obtained and corrected according to maximum likelihood estimator, positioning variances
Priori probability density function, wherein, original priori probability density function corrects priori probability density approximation to function as gaussian probability
Density function.
Wherein, maximum likelihood estimator is acquired according to by equation below:
Wherein,Represent maximum likelihood estimator,Represent the position point of target in priori probability density function
Measure akCorresponding mean value,For the position location of target, λ is a constant, and T is target observation time interval, and v is target speed
Degree,Represent observation noise variance,Represent new breath covariance.
Second acquisition module 23, for using correct priori probability density function as the second the importance density function carry out
Particle filter is to obtain the second mean value corresponding with dbjective state and the second covariance value.Specifically, the second acquisition module 23
For extracting the second particle collection from the second the importance density function first, then obtain the second particle and each particle is concentrated to correspond to
The second weights, then each second weights are standardized, finally according to the second weights after standardization with
And particle corresponding with the second weights obtains the second mean value corresponding with dbjective state and the second covariance value.
Posterior probability density function acquisition module 24, for according to Target state estimator weights respectively to the first mean value and
It is close to obtain posterior probability corresponding with dbjective state that two mean values, the first covariance value and the second covariance value are weighted processing
Function is spent, completes particle filter process.
Wherein, Target state estimator weights are acquired according to equation below:
Wherein, akRepresent Target state estimator weights, zkRepresent Current observation value;hkNonlinear riew known to () expression
Survey function;WithIt represents to carry out particle as the first the importance density function using priori probability density function respectively
Filter obtained the first mean value and the first covariance;WithIt represents to utilize respectively and corrects priori probability density function work
The second mean value and the second covariance value that particle filter obtains are carried out for the second the importance density function.
Please refer to Fig.1 the flow chart for the method for tracking target that 2, Figure 12 is the embodiment of the present invention.If it is noted that there is reality
Identical as a result, method of the invention is not limited with flow shown in FIG. 1 sequence in matter, as shown in figure 12, this method includes
Following steps:
Step S31:Receive observation data acquisition system.
In the present embodiment, observation data acquisition system is observed before including Current observation time and Current observation time
The observation of target can include the corresponding angle information of target, the speed of target, the observation interval of target etc..
Step S32:Original priori probability density function is built according to observation data acquisition system.
In the present embodiment, original priori probability density function is Gaussian function.
Step S33:Using original priori probability density function as the first the importance density function carry out particle filter with
Obtain the first mean value corresponding with dbjective state and the first covariance value.
Step S34:In original priori probability density function, using blocking, theory introduces Current observation information and target is special
Property information with build correct priori probability density function.
Step S35:Using correct priori probability density function as the second the importance density function carry out particle filter with
Obtain the second mean value corresponding with dbjective state and the second covariance value.
Step S36:According to Target state estimator weights respectively to the first mean value and the second mean value, the first covariance value and
Two covariance values are weighted processing to obtain posterior probability density function corresponding with dbjective state.
In the present embodiment, step S33~step S36 is identical with step S11~step S14 in Fig. 1, includes step
All technology contents disclosed in S11~step S14, for the sake of brief, details are not described herein.
Step S37:Dbjective state is estimated using posterior probability density function, to obtain Target state estimator value.
Step S38:Target state estimator value is exported, to realize the tracking to target.
Please refer to Fig.1 the structure diagram for the target tracker that 3, Figure 13 is the embodiment of the present invention.As shown in figure 13, should
Device includes:
Data reception module 41 is observed, data acquisition system is observed for receiving.
Original priori probability density function builds module 42, close for building original prior probability according to observation data acquisition system
Spend function.
First acquisition module 43, for being carried out using original priori probability density function as the first the importance density function
Particle filter is to obtain the first mean value corresponding with dbjective state and the first covariance value.
Priori probability density function structure module 44 is corrected, in original priori probability density function, using blocking
Theory introduces Current observation information and target property information and corrects priori probability density function to build.
Second acquisition module 45, for using correct priori probability density function as the second the importance density function carry out
Particle filter is to obtain the second mean value corresponding with dbjective state and the second covariance value.
Posterior probability density function acquisition module 46, for according to Target state estimator weights respectively to the first mean value and
It is close to obtain posterior probability corresponding with dbjective state that two mean values, the first covariance value and the second covariance value are weighted processing
Function is spent, completes particle filter process.
State estimation module 47, for being estimated using posterior probability density function dbjective state, to obtain target
State estimation.
Output module 48, for exporting target state estimator value, to realize the tracking to target.
The beneficial effects of the invention are as follows:The present invention auxiliary block particle filter method, device and method for tracking target and
Device as the first the importance density function and corrects priori probability density function conduct by the use of original priori probability density function
Second the importance density function obtains the corresponding mean value of dbjective state and covariance respectively, then according to Target state estimator weights
Two mean values and covariance are weighted respectively to obtain posterior probability density function corresponding with dbjective state, so as to complete
Particle filter process.By the above-mentioned means, the present invention can improve the accuracy and real-time of particle filter, it is non-thread so as to solve
Target maneuver brings the fast-moving target tracking problem under object module uncertain condition under property non-Gaussian environment.
The foregoing is merely embodiments of the present invention, are not intended to limit the scope of the invention, every to utilize this
It is relevant to be directly or indirectly used in other for the equivalent structure or equivalent flow shift that description of the invention and accompanying drawing content are made
Technical field is included within the scope of the present invention.