CN109509207A - The method that a kind of pair of point target and extension target carry out seamless tracking - Google Patents
The method that a kind of pair of point target and extension target carry out seamless tracking Download PDFInfo
- Publication number
- CN109509207A CN109509207A CN201811338016.8A CN201811338016A CN109509207A CN 109509207 A CN109509207 A CN 109509207A CN 201811338016 A CN201811338016 A CN 201811338016A CN 109509207 A CN109509207 A CN 109509207A
- Authority
- CN
- China
- Prior art keywords
- target
- state
- poisson
- covariance
- measurement
- 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
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
-
- 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
- G06F17/15—Correlation function computation including computation of convolution operations
-
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/207—Analysis of motion for motion estimation over a hierarchy of resolutions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/277—Analysis of motion involving stochastic approaches, e.g. using Kalman filters
Abstract
The invention discloses the methods that a kind of pair of point target and extension target carry out seamless tracking, belong to high-resolution sensor target tracking technique field.The variable geometry profile of the Gaussian process GP modeling target of the variable dimension of the method for the invention, each sensor resolution cell that objective contour occupies are denoted as measurement source;If the target relative to sensor observation is smaller, the measurement source number on objective contour is fewer, and vice versa.The present invention utilizes the radius number of the estimated value size on-line tuning GP model of measurement source number.The present invention adapts to the mutual conversion between the profile variation of ET and ET and PT, seamlessly tracks multiple ET and PT, and keeps preferable tracking performance.When target is ET, the position and shape of ET-GP-PMHT tracking and output target;When target is PT, ET-GP-PMHT only keeps track and exports the position of target.In addition, this method computation complexity is related to number, radius number and number of targets is measured.When ET shape becomes smaller, the GP model radius number of use tails off, and computation complexity reduces.
Description
Technical field
The invention belongs to high-resolution sensor target tracking technique fields, and in particular to a kind of pair of point target and extension target
The method for carrying out seamless tracking.
Background technique
In existing track algorithm, target is modeled as point source.As the raising or target of radar resolution are from biography
Sensor is closer, and target accounts for multiple resolution cells, generates multiple measurements, and point target (PT) model is no longer applicable in, thus develops and expand out
Therefore there is the document of more and more research ET tracking (ETT) algorithms in the problem of opening up target (ET).
In existing ETT research, ET state is modeled as motion state and target shape two parts.In recent years, various
ET shape modeling pattern is suggested.Random matrix (RM) model is modeled as an ellipse using symmetrical positive definite matrix, moves
State meets Gaussian Profile, and target shape then meets inverse Wishart and is distributed.Only the shape of simulated target is gone to be unsuitable for ellipse
All targets, therefore non-model of ellipse is more suitable for the ET of arbitrary shape.
One intuitive idea is by Target Modeling into multiple elliptical combinations, and another method is using star convex form side
Method, unknown extension target shape is modeled as limited unknown function of radius by it, the star based on random super surface model (RHM)
Convex form method defines function of radius in frequency domain, is parameterized using the method pair radius of Fourier expansion.In RHM frame
In can use greatest hope (EM) method, advantage of the EM based on recurrence Gauss RHM method also be studied verifying.Based on Gauss mistake
The star convex form method of journey (GP) model can be in spatial domain to radius of target function modelling, i.e., target shape models, and keeps mesh
Mark does not observe the uncertainty of part.It is flexible enough, can be used to indicate various shape.In GP frame, spreading kalman
Filtering (EKF) and particle filter (PF) can track single extension target, in order to track multiple targets, the more Bernoulli Jacob of label
(LMB) filtering and Gaussian-mixture probability assume that density (GM-PHD) filtering is suggested in isomery multisensor scene.
Multiple ET in actual ETT scene may have different shapes, and remote from sensor or close, target may occupy
One resolution cell or multiple resolution cells show as PT or ET.The same target may show as ET, far when close from sensor
When be PT.It may be simultaneously present ET and PT in certain monitoring sections, extension target can be with mesh relative to the size of sensor
The change marking division, merging, deflection and the variation away from sensor distance and occurring.Present GP model E TT algorithm uses shape
The fixed GP model of radius number, is no longer adapted to the extension target of multiple varied appearances.Although existing method can handle mesh
Certain weaker variations of size are marked, but effectively handles without proposition system and becomes between profile variation, especially ET and PT
The method of change.
Summary of the invention
It is an object of the invention in view of the above-mentioned problems, proposing that a kind of pair of point target and extension target carry out seamless tracking
Method, there are clutter and missing inspection, to track the profile variation of ET and while track PT and ET, solution measurement and shape
The data correlation problem of point.
Realizing key problem in technology of the invention is: with the variable geometry profile of the GP modeling target of variable dimension, introducing Poisson
Rate estimates measurement source number, measures source number and reflects outline point number, so that available Poisson rate dynamic adjusts the radius of GP model
Number adapts model to the variation of target shape.And data correlation is solved the problems, such as using PMHT algorithm, to complete multiple expansions
Open up tracking of the target in clutter.
Technical problem proposed by the invention solves in this way:
The method that a kind of pair of point target and extension target carry out seamless tracking, comprising the following steps:
Step 1. initializes ET-GP-PMHT algorithm parameter:
Step 1a. init state context parameter and GP model parameter: state-transition matrix, state-noise covariance, just
Beginning state covariance, hyper parameter etc.;
Step 1b. initializes observing environment parameters: observation noise variance R, clutter density, sampling interval Δ t, monitoring
SPACE V, sensor position, detection probability Pd;
Step 1c. imports observation information: including sharing T frame data, having T in each sliding windowbFrame data, 1 in sliding window
~TbMetric data set Z, the t frame amount measured data set Z of framet, t frame amount survey collective number Mt, 1≤t≤Tb;
Step 2. initializes the T in sliding windowbFrame data and metric data set Z, enabling current iteration number i is 1;
Step 2a. removes the measurement for being associated with existing track in measuring space, by its surplus in measuring space
Euclidean distance is lower than the measurement stacking of distance threshold between survey, at the beginning of carrying out two o'clock difference method with the measurement heap center of preceding 2 frame
Beginningization fresh target track;New targetpath if it exists, then init state environment parameters: number of targets Nt, in first frame
The GP model radius of result initialized target track of the element number compared with detection probability in set is measured obtained by each stacking
NumberTwo o'clock difference obtains target original stateInit state transfer matrix, original state covariance, state-noise
Covariance, Poisson rate and Poisson parameter;
Step 2b. is importedThe corresponding outline point of a radiusKind measurement model;
The posterior probability calculation formula of step 3. construction ET-GP-PMHT t frame:
Step 3a. Poisson velocity vectors:
Wherein, n=1 ..., Nt,Poisson velocity vectors λ is 1 to TbThe Poisson velocity vectors set of frame,
Present invention assumes that the measurement number of target and the clutter number measured in section meet Poisson distribution, therefore, Poisson speed can be used
Rate replaces the prior probability in original PMHT, and Poisson rate can also react target and generate the mean number measured;λ0, tIt represents miscellaneous
It is λ that wave number, which obeys mean value,0, tPoisson distribution, be set as constant in this invention;Poisson rate λN, l, tDistribution obey with
αN, l, t | tAnd βN, l, t | tλ is distributed for the gamma of Poisson parameterN, l, t=γ (λN, l, t;αN, l, t | t, βN, l, t | t), αN, l, t | tFor shape ginseng
Number, βN, l, t | tFor scale parameter;
The calculation formula of step 3b. likelihood are as follows:
Assuming that clutter is space uniform distribution, then likelihood value:
Wherein, zJ, tIt is jth (j=1 ..., the M of t framet) a measurement, xN, tIt is the dbjective state of t frame n target,Indicate withFor mean value, with RN, l, tFor the Gaussian probability-density function of covariance,hL, t() indicates the measurement function of measurement model corresponding to first of outline point of n target when t,
RN, l, tThe covariance matrix of measurement model is corresponded to for it,It is n-th of t frame, first of target outline point in world coordinates axis
On angle (such as Fig. 1), the measurement model of different target is identical;
Step 3c. posterior probability formula:
Wherein, ωJ, l, n, tIndicate that moment t measures zJ, tIt is derived from target xN, tFirst of outline point posterior probability;
Step 3d. Poisson rate equation:
αN, l, t-1 | t=exp {-Δ t/ τ } αN, l, t | t βN, l, t-1 | t=exp {-Δ t/ τ } βN, l, t | t
βN, l, t | t=βN, l, t | t-1+1
Wherein, exp is exponent, αN, l, t | t-1For predicting shape parameter, βN, l, t | t-1To predict scale parameter, τ is one
A time constant refers to the response speed that estimation changes evolution Poisson rate;
Step 4. calculates comprehensive measurement and comprehensive covariance:
It is comprehensive to measureWith comprehensive covarianceFormula be respectively as follows:
So far, it extends to measure under target scene and solved with target shape point Interconnected Fuzzy problem, for each target
An outline point only one synthesis measure and comprehensive covariance;
Step 5. judges t=TbIt is whether true, if set up, perform the next step;Otherwise t=t+1 is returned to step
3;
Step 6. spreading kalman is smooth:
Because the measurement function of extension target be it is nonlinear, state is realized using spreading kalman smoothing algorithm
With new estimation.Stacking method can be used to stack measurement matrix, comprehensive measurement and comprehensive covariance, then with extension karr
Graceful smoothing algorithm.
Because measurement function is nonlinear function, need to ask Jacobian matrix as measurement matrix measurement function:
Then, respectively measurement matrix, comprehensive measurement and comprehensive covariance are stacked to obtain:
Wherein, diag () indicates diagonalizable matrix;Finally, to target xN, tThe spreading kalman smoothing algorithm of execution is calculated
Method step is consistent with conventional Extension Kalman smoothing algorithm;
Step 7. judgement is whether number of iterations i meets the loop iteration condition of convergence, return step 3 if being unsatisfactory for;Convergence
Then execute 8;
Step 8. judges that track terminates: defining averaged power spectrum rateIf ξ is less than thresholding ξTH, the boat
Mark terminates, otherwise the track continues;
The adaptive dynamic object shape of step 9. adjusts target shape point number:
Step 9a. measures source number with target Poisson rate estimates:
Estimate the difference of measurement source number and outline point number:
If step 9b. var > 0, the big radius of target of var Poisson parameter, adds beside these radiuses before finding out
Add the identical new radius of Poisson parameter;
If var < 0, the smallest-var radiuses of Poisson parameter are deleted;
If var ≠ 0, state-transition matrix, state-noise covariance Q are updatedN, t, state covariance PN, t;
If var=0 thens follow the steps 10;
Step 10. judges whether sliding window includes the last T of T frame data collectionbFrame data, if not provided, sliding window is to advancing slip
Dynamic TsA moment forms T in new windowbFrame data and metric data set Z, return to step 2;Otherwise algorithm terminates.
The method of the invention measures source number with Poisson rate estimates, adjusts the radius number of GP model, objective contour
The each sensor resolution cell occupied is denoted as measurement source.If the target relative to sensor observation is smaller, on objective contour
Measurement source number is fewer, estimates that the radius number of GP is fewer, vice versa.When target be PT when, ET-GP-PMHT only keep track and
Export the position of target.The track algorithm of target uses PMHT algorithm, and one target of hypotheses of the algorithm can produce more
A measurement meets the actual conditions of extension target.
The beneficial effects of the present invention are:
(1) ET-GP-PMHT can track the profile variation of ET and the mutual conversion of seamless tracking ET and PT, when table is outer
When the constant ET of shape, the GP radius number that algorithm estimates is almost unchanged, therefore is able to maintain preferable tracking performance;When target is PT
When, ET-GP-PMHT only keeps track the position of target.
(2) computation complexity of ET-GP-PMHT algorithm is related to number, radius number and number of targets is measured.When ET shape becomes smaller
When, the GP model radius number of use tails off, and computation complexity reduces.
Detailed description of the invention
Fig. 1 is the ET of 16 outline point in GP model, and there are two coordinates in figure --- it local coordinate and global sits
Mark, the point of black are 16 outline points;
Fig. 2 is the estimation tracking of four targets and real goal in single Monte-Carlo Simulation, and wherein real goal is with black
Line indicates that estimation target indicates that targetpath and shape illustrate out with red line, and four plus siges indicate four targetpaths
Initial position, sense in origin;
Fig. 3 is the RMSE of four targets, 100 Monte Carlos;
Fig. 4 is Fig. 5, and 6,7,8 legend, wherein A1 indicates ET-GP-PMHT, A2 indicates ET-GP-PMHT-FBP26, A3
Indicate ET-GP-PMHT-FBP10, A4 indicates ET-RM-PMHT.ET-GP-PMHT-FBP26 is the ET-GP- for having 26 outline points
PMHT-FBP algorithm;
Fig. 5 is to track mesh with ET-GP-PMHT-FBP26, ET-GP-PMHT-FBP10, ET-RM-PMHT, ET-GP-PMHT
Mark 1, in the shape tracking situation of 121s, 181s, 241s and 691s, the estimation shape of the true profile of target and four algorithms is equal
It is showed in figure;
Fig. 6 is to track mesh with ET-GP-PMHT-FBP26, ET-GP-PMHT-FBP10, ET-RM-PMHT, ET-GP-PMHT
Mark 2 tracks situation in the shape of 95s, 125s, 335s and 635s;
Fig. 7 is to track mesh with ET-GP-PMHT-FBP26, ET-GP-PMHT-FBP10, ET-RM-PMHT, ET-GP-PMHT
Mark 3 tracks situation in the shape of 25s, 155s;
Fig. 8 is to track mesh with ET-GP-PMHT-FBP26, ET-GP-PMHT-FBP10, ET-RM-PMHT, ET-GP-PMHT
Mark 4 tracks situation in the shape of 31s, 91s.
Specific embodiment
The present invention is further detailed with reference to the accompanying drawings and examples.
The present embodiment provides a kind of pair of point target and the extension target method that carries out seamless tracking, emulation is there are clutters
It is carried out in scene, as shown in Figure 1, the target of four movements of tracking, calculates root-mean-square error RMSE verification algorithm performance.And it will
ET-GP-PMHT is compared with ET-GP-PMHT-FBP26, ETGP-PMHT-FBP10, ET-RM-PMHT.ET-GP-PMHT-
FBP indicates that the fixed ET-GP-PMHT algorithm of outline point number, ET-GP-PMHT-FBP26 indicate that outline point number is 26, base
ET-RM-PMHT is denoted as in the PMHT algorithm of RM model.
State equation uses uniform rectilinear's model, and 1 range sensor of target arrive closely again from the near to the remote, and target is from extending target
Becoming point target again becomes extending target, run duration 1-701s;2 range sensor of target is by as far as closely again to remote, target
Become extension target from point target becomes point target, run duration 5-701s again.Target 1,2 is steady circular in 230-481s
Movement, that is, turn, angular speed 1/80rad/s.Range sensor farther out, shows as point target to target 3 always, and run duration is
21-230s;Range sensor is closer always for target 4, shows as the constant extension target of size, run duration 31-230s.Mesh
Mark actual measurements source numberPosition S between sensor of interest is related:
The present embodiment the described method comprises the following steps:
Step 1. initializes ET-GP-PMHT algorithm parameter:
Step 1a. init state context parameter and GP model parameter: state-transition matrix, state-noise covariance, just
Beginning state covariance, hyper parameter etc.;
Dbjective state is For motion state,For profile state,Be byIt is a outer
The vector of the corresponding shape radius composition of form point, Table
Show target's center in the position of two-dimensional space,Indicate its corresponding speed, ψN, tIndicate the angle of target rotational
Degree, i.e., the angle (as shown in Figure 1) between world coordinates and local coordinate,For angular speed, the angle of initialized target rotation
It is respectively 0rad, 0rad/s with angular speed;State-transition matrix isThe wherein transfer of motion state
MatrixFor the state-transition matrix of uniform rectilinear (CV) model, the transfer matrix of profile state Table
Show that dimension isUnit matrix, wherein frame time at intervals Δ t be 1s, forgetting parameter alpha=0.0001 of state space;
State-noise covarianceMotion state noiseFor the state-noise of CV model,
The state-noise standard deviation of middle position and angle is respectively σq=0.05, σψ=0.001, the noise of profile state For the basis vector that target shape point is formed in the angle of local coordinate, outline point
It is also referred to as basic point (such as Fig. 1) in the angle of local coordinate, the angle between adjacent outline point is equal, i.e., outline point is in angle
It is evenly distributed in target shape, outline pointFor the covariance matrix of GP model:
Covariance is modified square of index (SE) function, and the period is that 2 π, u and u ' are k argument of function
The hyper parameter of GP model is arranged are as follows: σr=0.3, σf=0.5, another hyper parameter --- length scale is according to following
Rule adjustment:
The original state of four targets is respectively as follows:
The state covariance of four targets isMotion state covariance is diagonal matrix diag
([0.01,0.001,0.01,0.001,0.001,0.0001]), profile state covariance are
Step 1b. initializes observing environment parameters: observation noise variance R, clutter density, sampling interval Δ t, monitoring
SPACE V, sensor position, detection probability Pd;
Sensor position is (0m, 0m)T, the range of two-dimensional detection area x, y are [0,450] × [0,450] m2, in region
Clutter is uniformly distributed, and quantity obeys Poisson distribution, and the mean clutter number at per moment is 20;The signal-to-noise ratio of scene is 21dB, mesh
Target detection probability is Pd=0.92;
Measuring noise isExp {-Δ t/ τ }=0.9;It is criticized every time in PMHT
The duration T of processingbFor 3 frame moment, sliding length TsFor 2 frame moment.Used in every batch processing fixed cycles the number of iterations for
5 times;
Step 1c. imports observation information: including sharing T frame data, having T in each sliding windowbFrame data, 1 in sliding window
~TbMetric data set Z, the t frame amount measured data set Z of framet, t frame amount survey collective number Mt, 1≤t≤Tb;
Step 2. initializes the T in sliding windowbFrame data and metric data set Z, enabling current iteration number i is 1;
Step 2a. is removed in measuring space has been associated with the measurement of existing track, with some distance threshold divide with
Judge measure whether be associated with existing track, in measuring space by remaining measure between Euclidean distance lower than range gate
The measurement stacking for limiting 15m carries out two o'clock difference method initialization fresh target track with the measurement heap center of preceding 2 frame;It is new if it exists
Targetpath, then init state environment parameters: number of targets Nt, element in set is measured in first frame obtained by each stacking
The GP model radius number of result initialized target track of the number compared with detection probabilityAt the beginning of two o'clock difference obtains target
Beginning stateOriginal state covariance, state-noise covariance, Poisson rate and Poisson parameter lambdaN, l, t=0.7, αN, l, t | t=
8, βN, l, t | t=10;
Step 2b. is importedThe corresponding outline point of a radiusKind measurement model;
After ET-GP-PMHT algorithm environment parameter determines, observation model is also predefined.From state space to observation space
Mapping, the measurement model of the corresponding outline point of first of radius:
Wherein, direction vector:
The angle (such as Fig. 1) for being n-th of t frame, first of target outline point on local coordinate axis, eT, j~N (0, R)
Indicate that mean value is the Gaussian Profile that 0 covariance is R;
The posterior probability calculation formula of step 3. construction ET-GP-PMHT t frame:
Step 3a. Poisson velocity vectors:
Wherein, n=1 ..., Nt,Poisson velocity vectors λ is 1 to TbThe Poisson velocity vectors set of frame,
Present invention assumes that the measurement number of target and the clutter number measured in section meet Poisson distribution, therefore, Poisson speed can be used
Rate replaces the prior probability in original PMHT, and Poisson rate can also react target and generate the mean number measured;λ0, tIt represents miscellaneous
It is λ that wave number, which obeys mean value,0, tPoisson distribution, be set as constant in this invention;Poisson rate λN, l, tDistribution obey with
αN, l, t | tAnd βN, l, t | tλ is distributed for the gamma of Poisson parameterN, l, t=γ (λN, l, t;αN, l, t | t, βN, l, t | t), αN, l, t | tFor shape ginseng
Number, βN, l, t | tFor scale parameter;
The calculation formula of step 3b. likelihood are as follows:
Assuming that clutter is space uniform distribution, then likelihood value:
Wherein, zJ, tIt is jth (j=1 ..., the M of t framet) a measurement, xN, tIt is the dbjective state of t frame n target,Indicate withFor mean value, with RN, l, tFor the Gaussian probability-density function of covariance,hL, t() indicates the measurement function of measurement model corresponding to first of outline point of n target when t,
RN, l, tThe covariance matrix of measurement model is corresponded to for it,It is n-th of t frame, first of target outline point in world coordinates axis
On angle (such as Fig. 1), the measurement model of different target is identical;
Step 3c. posterior probability formula:
Wherein, ωJ, l, n, tIndicate that moment t measures zJ, tIt is derived from target xN, tFirst of outline point posterior probability;
Step 3d. Poisson rate equation:
αN, l, t-1 | t=exp {-Δ t/ τ } αN, l, t | t βN, l, t-1 | t=exp {-Δ t/ τ } βN, l, t | t
βN, l, t | t=βN, l, t | t-1+1
Wherein, exp is exponent, αN, l, t | t-1For predicting shape parameter, βN, l, t | t-1To predict scale parameter, τ is one
A time constant refers to the response speed that estimation changes evolution Poisson rate;
Step 4. calculates comprehensive measurement and comprehensive covariance:
It is comprehensive to measureWith comprehensive covarianceFormula be respectively as follows:
So far, it extends to measure under target scene and solved with target shape point Interconnected Fuzzy problem, for each target
An outline point only one synthesis measure and comprehensive covariance;
Step 5. judges t=TbIt is whether true, if set up, perform the next step;Otherwise t=t+1 is returned to step
3;
Step 6. spreading kalman is smooth:
Because the measurement function of extension target be it is nonlinear, state is realized using spreading kalman smoothing algorithm
With new estimation.By T in sliding windowbThe stacking method of=3s stacks measurement matrix, comprehensive measurement and comprehensive covariance, then
With spreading kalman smoothing algorithm.
Because measurement function is nonlinear function, need to ask Jacobian matrix as measurement matrix measurement function:
Then, respectively measurement matrix, comprehensive measurement and comprehensive covariance are stacked to obtain:
Wherein, diag () indicates diagonalizable matrix;Finally, to target xN, tThe spreading kalman smoothing algorithm of execution is calculated
Method step is consistent with conventional Extension Kalman smoothing algorithm;
Step 7. judgement is whether current iteration number i is equal to 5, if otherwise return step 3;If so then execute 8;
Step 8. judges that track terminates: defining averaged power spectrum rateIf ξ is less than thresholding ξTH=0.2,
The track terminates, otherwise the track continues;
The adaptive dynamic object shape of step 9. adjusts target shape point number:
Step 9a. measures source number with target Poisson rate estimates:
Estimate the difference of measurement source number and outline point number:
If step 9b. var > 0, the big radius of target of var Poisson parameter, adds beside these radiuses before finding out
Add the identical new radius of Poisson parameter;
If var < 0, the smallest-var radiuses of Poisson parameter are deleted;
If var ≠ 0, transfer matrix, state-noise covariance Q are updatedN, t, state covariance PN, t;
If var=0 thens follow the steps 10;
Step 10. judges whether sliding window includes the last T of T=701s frame data collectionbFrame data, if not provided, sliding window
Forward slip TsAt=2s the moment, form T in new windowbFrame data and metric data set Z, return to step 2;Otherwise it calculates
Method terminates.
In this example implementation, Fig. 2 is demonstrated by single Monte-Carlo Simulation, and real goal and estimation target, every 10 frame are drawn
Target true profile and estimation shape, point target are indicated with five-pointed star out.Fig. 2 confirm ET-GP-PMHT can initialize PT and
ET can track PT and ET, and the mutual conversion between seamless tracking PT to ET simultaneously.When PT is converted to ET, target measures number
Become more, ET-GP-PMHT can continue to track the position even ET shape of target, while can track the deflection of target.
Fig. 3 is demonstrated by the position RMSE of four targets, when target 1,2 is converted between ET and PT, dynamic model mismatch,
There is peak value in RMSE, and target 4 is the constant ET of shape, its RMSE is less than PT target 3, because target 4 can detect more
It measures.
For the performance of better check algorithm, in addition to RMSE, we can also calculate following performance indicator: target
Average track number;Average initialization time delay starts the time difference for tracking target and real goal starting;Average track
Terminate delay.It is as shown in the table:
Table 1
Performance indicator | Average track number | Average initialization time delay | Average track terminates delay |
Target 1 | 2.08 | 0s | 0s |
Target 2 | 1.49 | 1.62s | 0.1s |
Target 3 | 1.2 | 1.02s | 10.88s |
Target 4 | 1 | 0.02s | 12.38s |
ET-GP-PMHT is compared with ET-GP-PMHT-FBP26, ETGP-PMHT-FBP10, ET-RM-PMHT.Mesh
Mark 1 has 10 measurement sources in 121s, and such as Fig. 5, ET-GP-PMHT and ET-GP-PMHT-FBP10 can preferably track target shape,
The target shape that ET-GP-PMHT-FBP26 is estimated is greater than realistic objective shape;When target 1 181s only have 2 measurement sources and
When 241s has 1 measurement source, ET-GP-PMHT judges target for PT, and other algorithm still estimates a closed curve
Target shape;Become more when target measures source, when such as 691s, ET-GP-PMHT and ET-GP-PMHT-FBP26 have better shape
Estimation.
Therefore, ET-GP-PMHT-FBP is suitable only for tracking a certain size target, regardless of true ET shape, very
To when target is PT, the target shape that ET-RM-PMHT is estimated all is oval (such as Fig. 5,6,7,8).And ET-GP-PMHT energy
When target shape is constant in the stable big target of tracking, Small object and PT, Fig. 8 simultaneously, preferable tracking accuracy is kept, in Fig. 7
PT also can preferably be tracked.
Finally, it is stated that the above implementation is only used to illustrate the technical scheme of the present invention and not to limit it, it is all according to Shen of the present invention
Please the scope of the patents equivalent change and modification done, be all covered by the present invention.
Claims (5)
1. the method that a kind of pair of point target and extension target carry out seamless tracking, which comprises the following steps:
Step 1. initializes ET-GP-PMHT algorithm parameter:
Step 1a. init state context parameter and GP model parameter: state-transition matrix, state-noise covariance, initial shape
State covariance, hyper parameter etc.;
Step 1b. initializes observing environment parameters: observation noise variance R, clutter density, sampling interval Δ t, monitoring space
V, sensor position, detection probability Pd;
Step 1c. imports observation information: including sharing T frame data, having T in each sliding windowbFrame data, 1~T in sliding windowbFrame
Metric data set Z, t frame amount measured data set Zt, t frame amount survey collective number Mt, 1≤t≤Tb;
Step 2. initializes the T in sliding windowbFrame data and metric data set Z, enabling current iteration number i is 1;
Step 2a. is removed in measuring space has been associated with the measurement of existing track, between measuring remaining in measuring space
Euclidean distance is lower than the measurement stacking of distance threshold, carries out two o'clock difference method initialization with the measurement heap center of preceding 2 frame
Fresh target track;New targetpath if it exists, then init state environment parameters: number of targets Nt, each point in first frame
The GP model radius number of result initialized target track of the element number compared with detection probability in set is measured obtained by heapTwo o'clock difference obtains target original stateInit state transfer matrix, original state covariance, state-noise association
Variance, Poisson rate and Poisson parameter;
Step 2b. is importedThe corresponding outline point of a radiusKind measurement model;
The posterior probability calculation formula of step 3. construction ET-GP-PMHT t frame:
Step 3a. Poisson velocity vectors:
Wherein, n=1 ..., Nt,Poisson velocity vectors λ is 1 to TbThe Poisson velocity vectors set of frame, uses Poisson
Rate replaces the prior probability in original PMHT, and Poisson speed response target generates the number measured;λ0, tRepresent clutter number clothes
It is λ from mean value0, tPoisson distribution;Poisson rate λN, l, tDistribution obey with αN, l, t | tAnd βN, l, t | tFor the gamma of Poisson parameter
It is distributed λN, l, t=γ (λN, l, t;αN, l, t | t, βN, l, t | t), αN, l, t | tFor form parameter, βN, l, t | tFor scale parameter;
The calculation formula of step 3b. likelihood are as follows:
Assuming that clutter is space uniform distribution, then likelihood value:
Wherein, zJ, tIt is j-th of measurement of t frame, j=1 ..., Mt, xN, tIt is the dbjective state of t frame n target,
Indicate withFor mean value, with RN, l, tFor the Gaussian probability-density function of covariance,hL, t() table
The measurement function of measurement model corresponding to first of outline point of n target, R when showing tN, l, tFor the covariance of corresponding measurement model
Matrix,For angle of n-th of t frame, first of the target outline point on world coordinates axis, the measurement model phase of different target
Together;
Step 3c. posterior probability formula:
Wherein, ωJ, l, n, tIndicate that moment t measures zJ, tIt is derived from target xN, tFirst of outline point posterior probability;
Step 3d. Poisson rate equation:
αN, l, t-1 | t=exp {-Δ t/ τ } αN, l, t | t βN, l, t-1 | t=exp {-Δ t/ τ } βN, l, t | t
Wherein, exp is exponent, αN, l, t | t-1For predicting shape parameter, βN, l, t | t-1To predict that scale parameter, τ are that the time is normal
Amount;
Step 4. calculates comprehensive measurement and comprehensive covariance:
It is comprehensive to measureWith comprehensive covarianceFormula be respectively as follows:
Step 5. judges t=TbIt is whether true, if set up, perform the next step;Otherwise t=t+1 is enabled, returns to step 3;
Step 6. spreading kalman is smooth:
Ask Jacobian matrix as measurement matrix to function is measured:
Respectively measurement matrix, comprehensive measurement and comprehensive covariance are stacked to obtain:
Wherein, diag () indicates diagonalizable matrix;
Finally, to target xN, tThe spreading kalman smoothing algorithm of execution;
Step 7. judgement is whether number of iterations i meets the loop iteration condition of convergence, return step 3 if being unsatisfactory for;Convergence is then held
Row 8;
Step 8. judges that track terminates: defining averaged power spectrum rateIf ξ is less than thresholding ξTH, the track knot
Beam, on the contrary the track continues;
The adaptive dynamic object shape of step 9. adjusts target shape point number:
Step 9a. measures source number with target Poisson rate estimates:
Estimate the difference of measurement source number and outline point number:
If step 9b. var > 0, the big radius of var Poisson parameter before finding out, in the side of these radiuses addition Poisson ginseng
The identical new radius of number;
If var < 0, the smallest-var radiuses of Poisson parameter are deleted;
If var ≠ 0, state-transition matrix, state-noise covariance Q are updatedN, t, state covariance PN, t;
If var=0 thens follow the steps 10;
Step 10. judges whether sliding window includes the last T of T frame data collectionbFrame data, if not provided, sliding window forward slip TsIt is a
Moment forms T in new windowbFrame data and metric data set Z, return to step 2;Otherwise algorithm terminates.
2. the method according to claim 1 for carrying out seamless tracking to point target and extension target, which is characterized in that step
The detailed process of 1a. init state context parameter and GP model parameter are as follows:
Dbjective state is For motion state,For profile state,Be byA outline point pair
The vector for the radius value composition answered,Dimension be to be adjusted by the measurement source number dynamic estimated, Indicate target's center in the position of two-dimensional space,Indicate its corresponding speed, ψN, tIndicate the angle of target rotational,For angular speed, initialized target rotation
Angle and angular speed be respectively 0rad, 0rad/s;State-transition matrix isWherein motion state
Transfer matrixFor the state-transition matrix of uniform rectilinear's model, the transfer matrix of profile state Table
Show that dimension isUnit matrix, wherein frame time at intervals Δ t be 1s, forgetting parameter alpha=0.0001 of state space;
State-noise covarianceMotion state noiseFor the state-noise of CV model, wherein position
Setting with the state-noise standard deviation of angle is respectively σq=0.05, σψ=0.001, the noise of profile state For the basis vector that target shape point is formed in the angle of local coordinate, adjacent shape
Angle between point is equal, i.e. outline point is evenly distributed in target shape in angle, outline pointFor GP model
Covariance matrix:
Covariance is a modified square of exponential function, and the period is that 2 π, u and u ' are k argument of function
The hyper parameter of GP model is arranged are as follows: σr=0.3, σf=0.5, another hyper parameter, that is, length scale is adjusted according to following rule
Section:
The original state of four targets is respectively as follows:
The state covariance of four targets isMotion state covariance be diagonal matrix diag ([0.01,
0.001,0.01,0.001,0.001,0.0001]), profile state covariance is
3. the method according to claim 2 for carrying out seamless tracking to point target and extension target, which is characterized in that step
The detailed process of 1b. initialization observing environment parameters are as follows:
Sensor position is (0m, 0m)T, the range of two-dimensional detection area x, y are [0,450] × [0,450] m2, clutter in region
It is uniformly distributed, quantity obeys Poisson distribution, and the mean clutter number at per moment is 20;The signal-to-noise ratio of scene is 21dB, target
Detection probability is Pd=0.92;
Measuring noise isExp {-Δ t/ τ }=0.9;Each batch processing in PMHT
Duration TbFor 3 frame moment, sliding length TsFor 2 frame moment;Use fixed cycles the number of iterations for 5 times in every batch processing.
4. the method according to claim 1 for carrying out seamless tracking to point target and extension target, which is characterized in that Poisson
Rate and Poisson parameter: λN, l, t=0.7, αN, l, t | t=8, βN, l, t | t=10.
5. the method according to claim 3 for carrying out seamless tracking to point target and extension target, which is characterized in that step
In 2bThe corresponding outline point of a radiusKind measurement model are as follows:
Mapping from state space to observation space, the measurement model of the corresponding outline point of first of radius:
Wherein, direction vector:
For angle of n-th of t frame, first of the target outline point on local coordinate axis, eT, j~N (0, R) indicates that mean value is 0
Covariance is the Gaussian Profile of R.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811338016.8A CN109509207B (en) | 2018-11-09 | 2018-11-09 | Method for seamless tracking of point target and extended target |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811338016.8A CN109509207B (en) | 2018-11-09 | 2018-11-09 | Method for seamless tracking of point target and extended target |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109509207A true CN109509207A (en) | 2019-03-22 |
CN109509207B CN109509207B (en) | 2021-07-06 |
Family
ID=65747968
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811338016.8A Active CN109509207B (en) | 2018-11-09 | 2018-11-09 | Method for seamless tracking of point target and extended target |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109509207B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111077494A (en) * | 2019-11-27 | 2020-04-28 | 江苏理工学院 | Tracking method and device |
DE102020121064A1 (en) | 2020-08-11 | 2022-02-17 | Valeo Schalter Und Sensoren Gmbh | Method for operating an assistance system in a motor vehicle, computer program product, computer-readable storage medium and assistance system |
CN116736286A (en) * | 2023-05-24 | 2023-09-12 | 兰州理工大学 | Progressive Bayes extended target tracking method and system based on random hypersurface |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7212652B1 (en) * | 2003-07-07 | 2007-05-01 | The United States Of America As Represented By The Secretary Of The Navy | Method for tracking targets with hyper-spectral data |
CN101354786A (en) * | 2007-07-23 | 2009-01-28 | 中国科学院计算技术研究所 | Analysis method of sports video case |
CN103759732A (en) * | 2014-01-14 | 2014-04-30 | 北京航空航天大学 | Angle information assisted centralized multi-sensor multi-hypothesis tracking method |
CN104237880A (en) * | 2014-09-18 | 2014-12-24 | 中国人民解放军海军航空工程学院 | Variable structure joint probability data interconnection formation target tracking method |
US9851461B1 (en) * | 2012-04-04 | 2017-12-26 | The United States Of America As Represented By The Secretary Of The Navy | Modular processing system for geoacoustic sensing |
CN107526070A (en) * | 2017-10-18 | 2017-12-29 | 中国航空无线电电子研究所 | The multipath fusion multiple target tracking algorithm of sky-wave OTH radar |
CN108363054A (en) * | 2018-02-07 | 2018-08-03 | 电子科技大学 | Passive radar multi-object tracking method for Single Frequency Network and multipath propagation |
-
2018
- 2018-11-09 CN CN201811338016.8A patent/CN109509207B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7212652B1 (en) * | 2003-07-07 | 2007-05-01 | The United States Of America As Represented By The Secretary Of The Navy | Method for tracking targets with hyper-spectral data |
CN101354786A (en) * | 2007-07-23 | 2009-01-28 | 中国科学院计算技术研究所 | Analysis method of sports video case |
US9851461B1 (en) * | 2012-04-04 | 2017-12-26 | The United States Of America As Represented By The Secretary Of The Navy | Modular processing system for geoacoustic sensing |
CN103759732A (en) * | 2014-01-14 | 2014-04-30 | 北京航空航天大学 | Angle information assisted centralized multi-sensor multi-hypothesis tracking method |
CN104237880A (en) * | 2014-09-18 | 2014-12-24 | 中国人民解放军海军航空工程学院 | Variable structure joint probability data interconnection formation target tracking method |
CN107526070A (en) * | 2017-10-18 | 2017-12-29 | 中国航空无线电电子研究所 | The multipath fusion multiple target tracking algorithm of sky-wave OTH radar |
CN108363054A (en) * | 2018-02-07 | 2018-08-03 | 电子科技大学 | Passive radar multi-object tracking method for Single Frequency Network and multipath propagation |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111077494A (en) * | 2019-11-27 | 2020-04-28 | 江苏理工学院 | Tracking method and device |
DE102020121064A1 (en) | 2020-08-11 | 2022-02-17 | Valeo Schalter Und Sensoren Gmbh | Method for operating an assistance system in a motor vehicle, computer program product, computer-readable storage medium and assistance system |
CN116736286A (en) * | 2023-05-24 | 2023-09-12 | 兰州理工大学 | Progressive Bayes extended target tracking method and system based on random hypersurface |
CN116736286B (en) * | 2023-05-24 | 2024-02-06 | 兰州理工大学 | Progressive Bayes extended target tracking method and system based on random hypersurface |
Also Published As
Publication number | Publication date |
---|---|
CN109509207B (en) | 2021-07-06 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN103729859B (en) | A kind of probability nearest neighbor domain multi-object tracking method based on fuzzy clustering | |
CN107402381B (en) | Iterative self-adaptive multi-maneuvering target tracking method | |
CN107066806B (en) | Data Association and device | |
CN107102295A (en) | The multisensor TDOA passive location methods filtered based on GLMB | |
CN107526070A (en) | The multipath fusion multiple target tracking algorithm of sky-wave OTH radar | |
CN109509207A (en) | The method that a kind of pair of point target and extension target carry out seamless tracking | |
CN108333569A (en) | A kind of asynchronous multiple sensors fusion multi-object tracking method based on PHD filtering | |
CN110503071A (en) | Multi-object tracking method based on the more Bernoulli Jacob's Additive Models of variation Bayes's label | |
CN108734725A (en) | Probabilistic contractor couple based on Gaussian process extends method for tracking target | |
CN104730511A (en) | Tracking method for multiple potential probability hypothesis density expansion targets under star convex model | |
CN103729637A (en) | Extended target probability hypothesis density filtering method based on cubature Kalman filtering | |
CN104715154B (en) | Core K average Data Associations based on KMDL criterion criterions | |
CN101980044A (en) | Method for tracking multiple targets under unknown measurement noise distribution | |
CN107797106A (en) | A kind of PHD multiple target tracking smooth filtering methods of the unknown clutter estimations of acceleration EM | |
CN108344981A (en) | Multisensor Asynchronous towards clutter detects TSBF multi-object tracking methods | |
CN110865343A (en) | LMB-based particle filter tracking-before-detection method and system | |
CN106403953A (en) | Method for autonomous navigation and positioning underwater | |
CN106054167A (en) | Intensity filter-based multi-extended target tracking method | |
CN111711432B (en) | Target tracking algorithm based on UKF and PF hybrid filtering | |
Güntürkün | Sequential reconstruction of driving-forces from nonlinear nonstationary dynamics | |
CN111736144A (en) | Maneuvering turning target state estimation method only using distance observation | |
CN115169136A (en) | Rapid UK-GMPHD multi-target tracking method in three-dimensional space | |
CN111708013B (en) | Target tracking filtering method for distance coordinate system | |
CN104880708B (en) | A kind of variable number maneuvering target tracking method | |
CN108469604B (en) | TWS radar space-time joint digital simulation method and system |
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 |