CN110334472A - Group motion trend assisted set potential probability hypothesis density filtering method - Google Patents

Group motion trend assisted set potential probability hypothesis density filtering method Download PDF

Info

Publication number
CN110334472A
CN110334472A CN201910644486.5A CN201910644486A CN110334472A CN 110334472 A CN110334472 A CN 110334472A CN 201910644486 A CN201910644486 A CN 201910644486A CN 110334472 A CN110334472 A CN 110334472A
Authority
CN
China
Prior art keywords
group
probability
weight
target
density
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201910644486.5A
Other languages
Chinese (zh)
Other versions
CN110334472B (en
Inventor
卢哲俊
田彪
胡卫东
刘永祥
黎湘
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
National University of Defense Technology
Original Assignee
National University of Defense Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by National University of Defense Technology filed Critical National University of Defense Technology
Priority to CN201910644486.5A priority Critical patent/CN110334472B/en
Publication of CN110334472A publication Critical patent/CN110334472A/en
Application granted granted Critical
Publication of CN110334472B publication Critical patent/CN110334472B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Pure & Applied Mathematics (AREA)
  • Mathematical Optimization (AREA)
  • Mathematical Analysis (AREA)
  • Computational Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Software Systems (AREA)
  • Databases & Information Systems (AREA)
  • Algebra (AREA)
  • Computing Systems (AREA)
  • Computer Hardware Design (AREA)
  • Evolutionary Computation (AREA)
  • Geometry (AREA)
  • Radar Systems Or Details Thereof (AREA)

Abstract

The invention relates to the technical field of target tracking, in particular to a population motion trend assisted set potential probability hypothesis density filtering method which is realized by using a Gaussian mixture technology. The method utilizes the group motion model to assist tracking, and predicts the target state by using different models according to the specific gravity in the state transition process by calculating the specific gravity of different motion models, thereby obtaining more stable and accurate target state estimation and target number estimation by only utilizing a single target motion model compared with the traditional method. And the method solves the problem of CPHD legacy PHD distribution, and the tracking performance is obviously improved.

Description

The collection gesture probability hypothesis density filtering method of one population movement tendency auxiliary
Technical field
The present invention relates to target following technical fields, and in particular, to the collection gesture probability of population movement tendency auxiliary is false If density (Group Motion Assisted Cardinalized Probability Hypothesis Density, GMA- CPHD) filtering method.
Background technique
As target following scene is more complicated, the requirement to target following technology is also constantly being promoted.Moment sensor Need to track has been not only single target or a small amount of multiple targets, and tracking target is often gone out in the form of cluster Show, such as the crowd moved in civilian scene, the wagon flow travelled on road；It is just more common in military scene, such as tank, dress The combat formations at different levels of the battle units such as first vehicle, aircraft and combatant's composition, boundling (multiple warhead) guided missile and newest nothing Man-machine battle group etc..In conventional target tracking, single goal motion model is often based on to the prediction of target movement, but it is single The movement of target often has biggish randomness, tends not to accurately grasp to the movement of single goal.But in collective motion In, target often has similar motion model in group, or has the characteristics that cooperative motion.Such as with identical destination The crowd of people's composition, the fleet advanced together, various units of cooperation etc..From the point of view of group, these clusters have group The characteristic of movement, the i.e. associated movement of target form the movement of group's entirety.Therefore, in single goal movable basis, in conjunction with Group's kinetic characteristic, can be by single goal kinematic constraint in a certain range that group moves, thus to mesh with group's motion model auxiliary Mark movement more accurately hold.
Summary of the invention
The purpose of the present invention is to provide collection gesture probability hypothesis density (GMA-CPHD) filters of population movement tendency auxiliary Wave method, and realized using Gaussian Mixture technology.
Realize the technical solution of the object of the invention are as follows: the collection gesture probability hypothesis density filtering side of population movement tendency auxiliary Method, comprising the following steps:
Step 1, it is measured using k moment each target position and calculates group's state, then using B-spline to 1~k moment group's state Track fitting is carried out, group's motion profile is obtained, the specific steps are as follows:
(1.1) collection for setting k moment all target positions measurement compositions is combined intoWherein LkIt is the amount at k moment Survey number, zk,lIt is that target position measures, sensor measurement model are as follows:
zk,l=Hmk,lR
Wherein mk,lIt is dbjective state, H is measurement matrix, and R is to measure noise covariance.
(1.2) group positionTake the mean value that all target positions measure in group:
1~k moment group position composition set can be obtained accordinglyWhen being fitted 1~k using B-spline function Group's geometric locus at quarter, is expressed from the next:
WhereinIt is fitting group position, PiIt is i-th of B-spline curves control point, NPIt is control point number, BI, pIt (k) is B Spline base function, p are matched curve orders, and curve matching number is p-1.
B-spline fit procedure is as follows:
One for providing 1~k moment is evenly dividingNode interval is [ki,ki+1).I-th of p rank B Spline base function is calculated by following recurrence formula:
PiCalculated by least square method: the data vector for enabling needs be fitted isData Covariance matrix be C, B-spline set of basis function is at matrix B；
The then least square solution at B-spline curves control point are as follows:
P=(BTC-1B)-1(BC-1Zgrp)
Wherein
(1.3) each target position is calculated to measure at a distance from fitting group position:
Wherein | | modulus is indicated, if Δ zk,lGreater than three times standard deviation, then it is assumed that zk,lIt is outlier (clutter) and rejects；
(1.4) step (1.2)~(1.3) are repeated until no outlier, last residue L 'kA measurement；
Step 2, group velocity is calculated by B-spline matched curve, and each using the measurement of each target position and group velocity composition Dbjective state, the gesture for obtaining the k moment are distributed ρk(n) and probability hypothesis density Dk(x), and be arranged each target single goal movement mould Type and group's motion model have identical specific gravity, the specific steps are as follows:
(2.1) B-spline basic function B is calculatedI, p(k) derivative:
Then group velocityAre as follows:
Target position is measured into zk,lAnd group velocityCollectively constitute L 'kA dbjective state
(2.3) equally distributed gesture distribution ρ is setk(n)=1/Nmax, wherein n ∈ { 0,1 ..., Nmax, NmaxFor maximum possible Number of targets；Probability hypothesis density Dk(x) are as follows:
WhereinIndicate that mean value isCovariance isLabel isGaussian component, Its weight isCall number j ∈ { 1,2 ..., Jk}；Gaussian component number is equal to remaining measurement number, i.e. J at this timek=L 'k, mean valueLabelWhereinIt indicates the integer greater than 0, assigns each Gaussian component one new label, own Different labels forms setWherein U () expression takes all unduplicated elements, Wherein γk,iIt is different, IkIndicate the quantity of different labels；
(2.4) motion model specific gravity be different motion model weight in the total weight of target proportion, use It indicates label (target)Motion model σ specific gravity, wherein σ=1 and σ=2 respectively indicate single goal motion model and group fortune Movable model, the initial single goal motion model and group's motion model that each label is arranged have identical specific gravity, i.e.,
Step 3, k=k+1 is enabled, has the probability hypothesis density D at last moment (k-1 moment)k-1(x) and gesture is distributed ρk-1 (n), current time (k moment) is predicted, the gesture distribution ρ predictedk|k-1(n) and prediction probability hypothesis density Dk|k-1(x), specific steps are as follows:
(3.1) the gesture distribution at current time is predicted:
Wherein λ (n-s) indicates the gesture distribution of clutter, and s and t are two integers,It is number of combinations,pS It is target survival probability；
(3.2) probability hypothesis density at current time is predicted:
Prediction label in formula WithRespectively indicate the Gaussian component with motion model σ prediction Mean value and covariance are calculated by lower two formula:
Wherein FσAnd QσThe state-transition matrix and process noise covariance matrix of motion model σ are respectively indicated, T indicates square Battle array transposition；
Last Dk|k-1It is (x) writeable are as follows:
Wherein Jk|k-1=2Jk-1,
Step 4, pass through the measurement set at current timeDensity D is assumed to the prediction probability of step 3k|k-1 (x) and prediction gesture is distributed ρk|k-1(n) it is updated, the probability hypothesis density D updatedk(x) and the gesture of update is distributed ρk (n), comprising the following steps:
(4.1) the update gesture distribution at current time are as follows:
Wherein:
<,>indicate that inner product calculates symbol, such as functionAnd ρk|k-1(n) inner product is expressed as
Sum () indicates all elements summing function, i.e.,
Indicate the number of permutations,
pDIndicate detection probability；
C () indicates the probability-density function of clutter；
es() indicates s rank elementary symmetric function, such as the s rank elementary symmetric function e of set Xs(X) is defined as:
And provide e0(X)=1；
(4.2) update probability assumes density are as follows:
Wherein DE,k(x) and DU,k(x；zk,l) respectively indicate missing inspection destination probability hypothesis density and measure update probability and assume Density；DE,k(x) it is given by:
Wherein missing inspection target Gaussian component weightAre as follows:
DU,k(x；zk,l) it is given by:
WhereinMax { } expression is maximized, and is measured and is updated Gaussian component weightAre as follows:
Wherein Zk{zk,lIndicate ZkRemove zk,lSet later,It is calculated by following formula:
Measure zk,lUpdate Gaussian component mean valueAnd covarianceCalculating process are as follows:
Step 5, it calculates the update weight of each label and weight may be left, then to the Gaussian component weight of missing inspection target It is modified, and arranges the call number arranged in update probability hypothesis density, be divided into following steps:
(5.1) global missing inspection target weight W is calculatedE,k:
(5.2) label is calculatedPrediction weight
WhereinIndicate labelGaussian component call number set；
(5.3) label γ is calculatedk,iUpdate weight
(5.4) label γk,iPossibility leave weightIt is calculated by following formula:
(5.5) allocation proportion of missing inspection weight is calculated by following formula:
(5.6) to labelMissing inspection target Gaussian component weightIt is modified, counts again as the following formula It calculates:
Wherein min { } expression is minimized；
(5.7) step (5.2)~(5.6) are repeated to all labels；
(5.8) revised global missing inspection target weightAre as follows:
IfThe two difference is averagely allocated to the Gaussian component of all missing inspection targets, finally are as follows:
(5.9) it arranges update probability and assumes density D 'k(x) call number of Gaussian component in:
Wherein Jk=Jk|k-1+Jk|k-1Lk；
Step 6, each label is calculatedSingle goal motion model specific gravity and group's motion model specific gravity, respectively Are as follows:
Then ownGaussian component haveAnd
Then remove the motion model label of Gaussian component, then Dk(x) it is written as:
Step 7, to the Gaussian component of the probability hypothesis density with same tag carry out beta pruning with merge, estimation number of targets Dbjective state is measured and is extracted, following steps are divided into:
(7.1) Gaussian component weight is less than Gaussian component weight threshold wPWhen it is negligible, i.e., ifBy weightCorresponding Gaussian component is deleted, Gaussian component weight threshold wP=10-5；
(7.2) if the distance between the Gaussian component with same tag is less than distance threshold dM, by these Gausses point Amount merges, distance threshold dM=1m；
(7.3) tag number estimate at the moment is to make to update the integer that gesture distribution is maximized, it may be assumed that
(7.4) it takes with weight limitThen label selects weight maximum from the Gaussian component of each label Gaussian component, mean value are exactly state estimation, such as labelState estimation be
Step 8, assume that density calculates group's state with update probability, calculate newborn destination probability and assume density, and will be newborn Probability hypothesis density is merged into update probability and assumes to be divided into following steps in density:
(8.1) z is calculatedk,lUpdate weight Wk(zk,l):
If Wk(zk,l) it is greater than measurement weight threshold value Wz, then it is assumed that measurement update weight is larger, belongs to already present mesh Mark rejects the measurement rather than from fresh target, last residue L "kA measure assumes that density calculates for newborn destination probability, Measurement weight threshold value Wz=0.5；
(8.2) group's state at this time is calculated by following formula:
By group's stateIn group velocity and each target position measure zk,lIt is combined into L "kA new life dbjective stateThen newborn destination probability assumes density DB,k(x) it is given by:
WhereinFor newborn target covariance,It is new label,For newborn target weight, calculated by following formula:
Wherein NBIt is the expectation of newborn number of targets, wB,max∈ [0,1] is the maximum allowable existing probability of newborn target, is used to LimitationIn a zone of reasonableness；The single goal motion model and group's motion model of new label have identical specific gravity, i.e.,
(8.3) newborn probability hypothesis density is merged into update probability to assume in density, the probability at current time is assumed close Degree is final are as follows:
Wherein Jk=Jk+L″k；
Step 9, step 3~8 are repeated, until track demand terminates.
The present invention has following technical effect that the present invention assists tracking using group's motion model, by calculating different fortune Movable model specific gravity obtains to use different model prediction dbjective states by specific gravity in state migration procedure compared with conventional method Only estimated using the more stable and accurate Target state estimator of single goal motion model and number of targets, and the present invention solves CPHD PHD assignment problem is left, tracking performance is promoted significant.
Detailed description of the invention
Fig. 1 is the flow chart of the collection gesture probability hypothesis density filtering method of population movement tendency auxiliary of the invention；
Fig. 2 is the monitoring scene and target real trace schematic diagram of emulation experiment of the present invention；
Fig. 3 is the multiple targets filter effect figure of emulation experiment of the present invention；
Fig. 4 be GMA-CPHD filter in emulation experiment of the present invention, improve the OSPA of CPHD and classics CPHD filter away from From comparison diagram；
Fig. 5 be GMA-CPHD filter in emulation experiment of the present invention, improve the OSPA of CPHD and classics CPHD filter away from From location error ingredient comparison diagram；
Fig. 6 be GMA-CPHD filter in emulation experiment of the present invention, improve CPHD estimate with classics CPHD filter number of targets Count comparative result figure；
Fig. 7 is GMA-CPHD filter single goal motion model specific gravity variation diagram in emulation experiment of the present invention.
Specific embodiment
In conjunction with Fig. 1, the collection gesture probability hypothesis density filtering method of population movement tendency auxiliary of the invention, including it is following Step:
Step 1, it is measured using k moment each target position and calculates group's state, then using B-spline to 1~k moment group's state Track fitting is carried out, group's motion profile is obtained；
Step 2, group velocity is calculated by B-spline matched curve, and each using the measurement of each target position and group velocity composition Dbjective state, the gesture for obtaining the k moment are distributed ρk(n) and probability hypothesis density Dk(x), and be arranged each target single goal movement mould Type and group's motion model have identical specific gravity；
Step 3, k=k+1 is enabled, has the probability hypothesis density D at last moment (k-1 moment)k-1(x) and gesture is distributed ρk-1 (n), current time (k moment) is predicted, the gesture distribution ρ predictedk|k-1(n) and prediction probability hypothesis density Dk|k-1(x)；
Step 4, pass through the measurement set at current timeDensity D is assumed to the prediction probability of step 3k|k-1 (x) and prediction gesture is distributed ρk|k-1(n) it is updated, the probability hypothesis density D updatedk(x) and the gesture of update is distributed ρk (n)；
Step 5, it calculates the update weight of each label and weight may be left, then to the Gaussian component weight of missing inspection target It is modified, and arranges the call number arranged in update probability hypothesis density；
Step 6, each label is calculatedSingle goal motion model specific gravity and group's motion model specific gravity, respectively Are as follows:
Then ownGaussian component haveAnd
Then remove the motion model label of Gaussian component, then D 'k(x) it is written as:
Step 7, to the Gaussian component of the probability hypothesis density with same tag carry out beta pruning with merge, estimation number of targets It measures and extracts dbjective state；
Step 8, assume that density calculates group's state with update probability, calculate newborn destination probability and assume density, and will be newborn Probability hypothesis density is merged into update probability and assumes in density；
Step 9, step 3~8 are repeated, until track demand terminates.
Effect of the invention is further illustrated by following emulation experiment:
Emulation experiment environment is 8 core CPU processors of Intel i73.6Hz dominant frequency, and program is write using Matlab language.
In emulation experiment, GMA-CPHD filter and the filter do not utilize group motion model assist in the case of tracking into Row performance compares, and referred to as improved CPHD filter (Improved CPHD), and the two is carried out with classical CPHD filter again Performance compares.
1, simulated conditions
Select transverse direction (X-axis) range from -2000m to 2000m, plane of longitudinal (Y-axis) range from -500m to 500m Region is monitor area, and the place that target occurs in region is unknown, and the quantity of target is unknown.
Multiple targets motion profile is as shown in Fig. 2, multiple targets occur from the left side of monitor area, and transport along certain route It moves on the right side of monitor area, there are small-scale random motions, group's mass motion trend to keep for each target in group in motion process Stablize, initial position and final position identify in figure.
The state of each target is expressed asWherein [xk,yk]TIndicate target lateral and longitudinal position,It is target lateral and longitudinal speed.
The state of target is expressed asWherein [x, y]TIt is target position,It is target velocity
Sampling interval τ=1s, group's motion model and single goal motion model all use uniform rectilinear motion model, then state Transfer matrix are as follows:
Process noise covariance matrix Q1=Q2=Gv, wherein v=[v1 2,v2 2]TAnd v1=v2=3m/s2, G gives by following formula Out:
Newborn target it is expected NB=0.03, maximum allowable existing probability wB,max=0.1, the initial association of fresh target Gaussian component Variance value is diag ([10,10,10,10]T)。
Measurement matrix H are as follows:
Measure noise covariance matrix R=diag ([q1 2,q2 2]T), it is provided with q1=q2=3m and q1=q2=10m two Kind situation；
The mean clutter number at each moment is 5, and clutter is evenly distributed in monitoring region；
Detection probability takes pD=0.9 and pD=0.6 two kind of situation；
Target survival probability is pS=0.99；
Preset several threshold values are respectively wP=105, dE=100m and Wz=0.5；
2, analysis of simulation result
Fig. 3 gives pD=0.9, q1=q2The tracking effect of GMA-CPHD filter when=3m.It can from figure Out, by carrying out track fitting to 1~10 moment group's state come after initialized target track, GMA-CPHD filter tracks quickly Upper all targets, accurately estimate dbjective state, and steadily and surely keep preferable tracking performance, illustrate that the present invention can complete well Target following.
Fig. 4 gives GMA-CPHD filter, improves CPHD filter and classics CPHD filter in three kinds of varying environments The lower Comparative result for carrying out 100 Monte Carlo simulations.Three kinds of environment are: pD=0.9, q1=q2=3m；pD=0.9, q1=q2 =10m；pD=0.6, q1=q2=3m.Performance evaluation norm is used as used here as the average OSPA distance of 100 Monte Carlos. From figure 5 it can be seen that GMA-CPHD filter has optimal performance, GMA-CPHD filter and improvement CPHD filter are all It is more many than doing very well for classical CPHD filter.Illustrate that filter of the invention has good tracking performance to multiple targets.
Fig. 5 gives GMA-CPHD filter and improves CPHD filter being averaged in this 100 Monte Carlo simulations Location error ingredient in OSPA distance.It can be seen that GMA-CPHD can obtain it is smaller compared with single goal motion model is only used Evaluated error illustrates that group moves the validity of auxiliary tracking.
Fig. 6 gives average target number estimated result of three kinds of filters in this 100 Monte Carlo simulations.It can see Out, GMA-CPHD filter and improvement CPHD filter all obtain accurate number of targets estimated result under all circumstances, compared with Classical CPHD is more stable accurate, when especially gesture detection probability is lower, illustrates that new filter can correctly distribute missing inspection target PHD, Stable performance can be also maintained when detection probability is lower.
Fig. 7 gives GMA-CPHD filter single goal motion model specific gravity.It can be seen from the figure that single goal moves mould Type specific gravity is gradually reduced, and filter is intended to be illustrated that group mass motion trend keeps stablizing using group's motion model, transported using group Movable model is predicted more accurate.

Claims (8)

1. the collection gesture probability hypothesis density filtering method of population movement tendency auxiliary, which is characterized in that this method includes following Step:
Step 1, it is measured using k moment each target position and calculates group's state, then 1~k moment group's state is carried out using B-spline Track fitting obtains group's motion profile, the specific steps are as follows:
(1.1) collection for setting k moment all target positions measurement compositions is combined intoWherein LkIt is the measurement number at k moment, zK, lIt is that target position measures, sensor measurement model are as follows:
zK, l=HmK, l+R
Wherein mK, lIt is dbjective state, H is measurement matrix, and R is to measure noise covariance；
(1.2) group positionTake the mean value that all target positions measure in group:
1~k moment group position composition set can be obtained accordinglyUse B-spline function fitting 1~k moment Group's geometric locus, is expressed from the next:
WhereinIt is fitting group position, PiIt is i-th of B-spline curves control point, NPIt is control point number, BI, pIt (k) is B-spline Basic function, p are matched curve orders, and curve matching number is p-1；
(1.3) each target position is calculated to measure at a distance from fitting group position:
Wherein | | modulus is indicated, if Δ zK, lGreater than three times standard deviation, then it is assumed that zK, lIt is outlier and rejects；
(1.4) step (1.2)~(1.3) are repeated until no outlier, last residue L 'kA measurement；
Step 2, group velocity is calculated by B-spline matched curve, and is measured using each target position and forms each target with group velocity State, the gesture for obtaining the k moment are distributed ρk(n) and probability hypothesis density Dk(x), and be arranged each target single goal motion model and Group's motion model has identical specific gravity, the specific steps are as follows:
(2.1) B-spline basic function B is calculatedI, p(k) derivative:
Then group velocityAre as follows:
Target position is measured into zK, lAnd group velocityCollectively constitute L 'kA dbjective state
(2.3) equally distributed gesture distribution ρ is setk(n)=1/Nmax, wherein n ∈ { 0,1 ..., Nmax, NmaxFor maximum possible target Number；Probability hypothesis density Dk(x) are as follows:
WhereinIndicate that mean value isCovariance isLabel isGaussian component, power Weight isCall number j ∈ { 1,2 ..., Jk}；Gaussian component number is equal to remaining measurement number, i.e. J at this timek=L 'k, mean valueLabelWhereinIt indicates the integer greater than 0, assigns each Gaussian component one new label, own Different labels forms setWherein U () expression takes all unduplicated elements, Wherein γK, iIt is different, IkIndicate the quantity of different labels；
(2.4) motion model specific gravity be different motion model weight in the total weight of target proportion, useIt indicates LabelMotion model σ specific gravity, wherein σ=1 and σ=2 respectively indicate single goal motion model and group's motion model；
Step 3, k=k+1 is enabled, has the probability hypothesis density D of last momentk-1(x) and gesture is distributed ρk-1(n), to current time It is predicted, the gesture distribution ρ predictedk|k-1(n) and prediction probability hypothesis density Dk|k1(x), specific steps are as follows:
(3.1) the gesture distribution at current time is predicted:
Wherein λ (n-s) indicates the gesture distribution of clutter, and s and t are two integers,It is number of combinations,pSIt is mesh Mark survival probability；
(3.2) probability hypothesis density at current time is predicted:
Prediction label in formula WithRespectively indicate the Gaussian component mean value with motion model σ prediction And covariance, it is calculated by lower two formula:
Wherein FσAnd QσThe state-transition matrix and process noise covariance matrix of motion model σ are respectively indicated, T representing matrix turns It sets；
Last Dk|k-1It is (x) writeable are as follows:
Wherein Jk|k-1=2Jk-1,
Step 4, pass through the measurement set at current timeDensity D is assumed to the prediction probability of step 3k|k-1(x) ρ is distributed with prediction gesturek|k-1(n) it is updated, the probability hypothesis density D updatedk(x) and the gesture of update is distributed ρk(n), it wraps Include following steps:
(4.1) the update gesture distribution at current time are as follows:
Wherein:
<,>indicate that inner product calculates symbol, such as functionAnd ρk|k-1(n) inner product is expressed as
Sum () indicates all elements summing function, i.e.,
Indicate the number of permutations,
pDIndicate detection probability；
C () indicates the probability-density function of clutter；
es() indicates s rank elementary symmetric function, such as the s rank elementary symmetric function e of set Xs(X) is defined as:
And provide e0(X)=1；
(4.2) update probability assumes density are as follows:
Wherein DE, k(x) and DU, k(x；zK, l) respectively indicate missing inspection destination probability hypothesis density and measure update probability and assume density； DE, k(x) it is given by:
Wherein missing inspection target Gaussian component weightAre as follows:
DU, k(x；zK, l) it is given by:
WhereinMax { } expression is maximized, and is measured and is updated Gaussian component weightAre as follows:
Wherein Zk-{zk-lIndicate ZkRemove zK, lSet later,It is calculated by following formula:
Step 5, it calculates the update weight of each label and weight may be left, then the Gaussian component weight of missing inspection target is carried out Amendment, and the call number arranged in update probability hypothesis density is arranged, it is divided into following steps:
(5.1) global missing inspection target weight W is calculatedE, k:
(5.2) label is calculatedPrediction weight
WhereinIndicate labelGaussian component call number set；
(5.3) label γ is calculatedK, iUpdate weight
(5.4) label γK, iPossibility leave weightIt is calculated by following formula:
(5.5) allocation proportion of missing inspection weight is calculated by following formula:
(5.6) to labelMissing inspection target Gaussian component weightIt is modified, recalculates as the following formula:
Wherein min { } expression is minimized；
(5.7) step (5.2)~(5.6) are repeated to all labels；
(5.8) revised global missing inspection target weightAre as follows:
IfThe two difference is averagely allocated to the Gaussian component of all missing inspection targets, finally are as follows:
(5.9) it arranges update probability and assumes density D 'k(x) call number of Gaussian component in:
Wherein Jk=Jk|k-1+Jk|k-1Lk；
Step 6, each label is calculatedSingle goal motion model specific gravity and group's motion model specific gravity, be respectively as follows:
Then ownGaussian component haveAnd
Then remove the motion model label of Gaussian component, then D 'k(x) it is written as:
Step 7, to the Gaussian component of the probability hypothesis density with same tag carry out beta pruning with merge, estimate destination number simultaneously Dbjective state is extracted, following steps are divided into:
(7.1) Gaussian component weight is less than Gaussian component weight threshold wPWhen it is negligible, i.e., ifBy weight Corresponding Gaussian component is deleted；
(7.2) if the distance between the Gaussian component with same tag is less than distance threshold dM, these Gaussian components are closed And；
(7.3) tag number estimate at the moment is to make to update the integer that gesture distribution is maximized, it may be assumed that
(7.4) it takes with weight limitThen label selects the maximum Gauss of weight from the Gaussian component of each label Component, mean value are exactly state estimation；
Step 8, assume that density calculates group's state with update probability, calculate newborn destination probability and assume density, and by newborn probability Assume to be divided into following steps in density assuming that density is merged into update probability:
(8.1) z is calculatedK, lUpdate weight Wk(zK, l):
If Wk(zK, l) it is greater than measurement weight threshold value Wz, then it is assumed that it is larger that the measurement updates weight, belong to already present target and It is not from fresh target, rejects the measurement, last residue L "kA measure assumes that density calculates for newborn destination probability；
(8.2) group's state at this time is calculated by following formula:
By group's stateIn group velocity and each target position measure zK, lIt is combined into L "kA new life dbjective state Then newborn destination probability assumes density DB, k(x) it is given by:
WhereinFor newborn target covariance,It is new label,For newborn target weight, calculated by following formula:
Wherein NBIt is the expectation of newborn number of targets, wB, max∈ [0,1] is the maximum allowable existing probability of newborn target, for limitingIn a zone of reasonableness；
(8.3) newborn probability hypothesis density is merged into update probability and assumes in density that the probability hypothesis density at current time is most Eventually are as follows:
Wherein Jk=Jk+L″k；
Step 9, step 3~8 are repeated, until track demand terminates.
2. the collection gesture probability hypothesis density filtering method of group's movement tendency auxiliary according to claim 1, it is characterised in that: (1.2) in, B-spline fit procedure is as follows:
One for providing 1~k moment is evenly dividingNode interval is [ki, ki+1)；I-th of p rank B-spline Basic function is calculated by following recurrence formula:
PiCalculated by least square method: the data vector for enabling needs be fitted isThe association of data Variance matrix is C, and B-spline set of basis function is at matrix B；
The then least square solution at B-spline curves control point are as follows:
P=(BTC-1B)-1(BC-1Zgrp)
Wherein
3. the collection gesture probability hypothesis density filtering method of group's movement tendency auxiliary according to claim 1, it is characterised in that: (2.4) in, the initial single goal motion model and group's motion model that each label is arranged have identical specific gravity, i.e.,
4. the collection gesture probability hypothesis density filtering method of group's movement tendency auxiliary according to claim 1, it is characterised in that: (4.2) in, z is measuredK, lUpdate Gaussian component mean valueAnd covarianceCalculating process are as follows:
5. the collection gesture probability hypothesis density filtering method of group's movement tendency auxiliary according to claim 1, it is characterised in that: (7.1) in, Gaussian component weight threshold wP=10-5。
6. the collection gesture probability hypothesis density filtering method of group's movement tendency auxiliary according to claim 1, it is characterised in that: (7.2) in, distance threshold dM=1m.
7. the collection gesture probability hypothesis density filtering method of group's movement tendency auxiliary according to claim 1, it is characterised in that: (8.1) in, measurement weight threshold value Wz=0.5.
8. the collection gesture probability hypothesis density filtering method of group's movement tendency auxiliary according to claim 1, it is characterised in that: (8.2) in, the single goal motion model and group's motion model of new label have identical specific gravity,
CN201910644486.5A 2019-07-15 2019-07-15 Group motion trend assisted set potential probability hypothesis density filtering method Active CN110334472B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910644486.5A CN110334472B (en) 2019-07-15 2019-07-15 Group motion trend assisted set potential probability hypothesis density filtering method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910644486.5A CN110334472B (en) 2019-07-15 2019-07-15 Group motion trend assisted set potential probability hypothesis density filtering method

Publications (2)

Publication Number Publication Date
CN110334472A true CN110334472A (en) 2019-10-15
CN110334472B CN110334472B (en) 2020-05-15

Family

ID=68145552

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910644486.5A Active CN110334472B (en) 2019-07-15 2019-07-15 Group motion trend assisted set potential probability hypothesis density filtering method

Country Status (1)

Country Link
CN (1) CN110334472B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113566821A (en) * 2021-06-28 2021-10-29 江南造船(集团)有限责任公司 Unmanned aerial vehicle course estimation method and system based on interactive filtering and electronic equipment

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101894381A (en) * 2010-08-05 2010-11-24 上海交通大学 Multi-target tracking system in dynamic video sequence
CN106022340A (en) * 2016-05-17 2016-10-12 南京理工大学 Improved Gaussian mixed potential probability hypothesis density filtering method
CN106772401A (en) * 2016-12-23 2017-05-31 浙江大学 Number of fish school method of estimation based on probability hypothesis density particle filter algorithm
CN107831490A (en) * 2017-12-01 2018-03-23 南京理工大学 A kind of improved more extension method for tracking target
CN109525220A (en) * 2018-12-10 2019-03-26 中国人民解放军国防科技大学 Gaussian mixture CPHD filtering method with track association and extraction capability
CN109934810A (en) * 2019-03-08 2019-06-25 太原理工大学 A kind of defect classification method based on improvement population wavelet neural network

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101894381A (en) * 2010-08-05 2010-11-24 上海交通大学 Multi-target tracking system in dynamic video sequence
CN106022340A (en) * 2016-05-17 2016-10-12 南京理工大学 Improved Gaussian mixed potential probability hypothesis density filtering method
CN106772401A (en) * 2016-12-23 2017-05-31 浙江大学 Number of fish school method of estimation based on probability hypothesis density particle filter algorithm
CN107831490A (en) * 2017-12-01 2018-03-23 南京理工大学 A kind of improved more extension method for tracking target
CN109525220A (en) * 2018-12-10 2019-03-26 中国人民解放军国防科技大学 Gaussian mixture CPHD filtering method with track association and extraction capability
CN109934810A (en) * 2019-03-08 2019-06-25 太原理工大学 A kind of defect classification method based on improvement population wavelet neural network

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
ZHEJUN LU: ""A new Cardinalized Probability Hypothesis density Filter with efficient track continue and extraction"", 《2018 1ST INTERNATIONAL CONFERENCE ON INFORMATION FUSION》 *
卢哲俊: ""基于随机有限集的空间碎片群运动状态估计"", 《航空学报》 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113566821A (en) * 2021-06-28 2021-10-29 江南造船(集团)有限责任公司 Unmanned aerial vehicle course estimation method and system based on interactive filtering and electronic equipment

Also Published As

Publication number Publication date
CN110334472B (en) 2020-05-15

Similar Documents

Publication Publication Date Title
CN110608743B (en) Multi-unmanned aerial vehicle collaborative route planning method based on multi-population chaotic grayling algorithm
CN109525220B (en) Gaussian mixture CPHD filtering method with track association and extraction capability
CN105182291B (en) The multi-object tracking method of the PHD smoothers of adaptive targets new life intensity
CN107677997B (en) Extended target tracking method based on GLMB filtering and Gibbs sampling
CN111176334A (en) Multi-unmanned aerial vehicle cooperative target searching method
CN105205313B (en) Fuzzy Gaussian sum particle filtering method and device and target tracking method and device
CN110320512A (en) A kind of GM-PHD smothing filtering multi-object tracking method based on tape label
CN107300698B (en) Radar target track starting method based on support vector machine
CN104156984A (en) PHD (Probability Hypothesis Density) method for multi-target tracking in uneven clutter environment
CN107402381A (en) A kind of multiple maneuver target tracking methods of iteration self-adapting
CN111562571B (en) Maneuvering multi-target tracking and track maintaining method for unknown new-born strength
CN107871156B (en) Ant colony multi-cell tracking system based on pheromone prediction
CN106934324A (en) Based on the radar data correlating methods for simplifying many hypothesis algorithms
CN111121784B (en) Unmanned reconnaissance aircraft route planning method
Li et al. An adaptive online prediction method with variable prediction horizon for future driving cycle of the vehicle
CN114172942A (en) Cooperative task allocation and track optimization method for multi-unmanned aerial vehicle assisted Internet of things
CN103226357A (en) Multiple-unmanned aerial vehicle communication decision method based on target tracking
CN109325128A (en) A kind of tracking and system of maneuvering target
CN109597839A (en) A kind of data digging method based on the avionics posture of operation
CN114167898A (en) Global path planning method and system for data collection of unmanned aerial vehicle
CN110334472A (en) Group motion trend assisted set potential probability hypothesis density filtering method
CN106203526B (en) Goal behavior mode online classification method based on multidimensional characteristic
CN109253727A (en) A kind of localization method based on improvement iteration volume particle filter algorithm
CN104849697B (en) It is a kind of based on the α β filtering methods for going inclined Coordinate Conversion
CN104850856A (en) Multi-extension target tracking method for affinity propagation cluster observation

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant