CN102707278A - Multi-target tracking method for singular value decomposition - Google Patents

Abstract

The invention discloses a multi-target tracking method for singular value decomposition. The multi-target tracking method is used for solving the technical problem that the conventional target tracking method is unstable in numerical value structure so as to cause target tracking loss in the radar tracking process. The technical scheme is that an evaluated error covariance matrix is subjected to singular value decomposition, a numerical value stabilizing structural model is established, two positive semi-definite matrixes in a covariance matrix without error evaluation are subtracted, and a symmetric matrix containing a negative eigenvalue is not generated in a word length limited processing system. The evaluated error covariance matrix is subjected to singular value decomposition, so that a multi-target tracking structural model with stable numerical value is established, the two positive semi-definite matrixes in an evaluated error covariance matrix are prevented from being subtracted, and numerical divergence is avoided in the word length limited processing system; and therefore, the reliability of the target tracking system is guaranteed, the target tracking loss in the radar tracking process and the whole radar system error are avoided.

Description

The multi-object tracking method of svd

Technical field

The present invention relates to a kind of Radar Multi Target tracking, particularly a kind of multi-object tracking method of svd belongs to areas of information technology.

Background technology

The multiple target tracking technology all is widely used at military and civil area, can be used for aerial target and detects, follows the tracks of and attack, Air Missile defence, air traffic control, harbour and marine surveillance etc.In recent years, along with the change of battlefield surroundings, the development of antagonism and the anti-technology of opposition has produced a series of problems such as the strong clutter of background, low signal-to-noise ratio, low detection probability and high false alarm rate, and the precision and the accuracy of multi-object tracking method are had higher requirement.

The purpose of multiple target tracking is the corresponding different information source of measurement that detector is received, forms different observation set or track, estimates the number of tracked target and the kinematic parameter of each target according to track, realizes the tracking to a plurality of targets.The basic filtering method that is used for the multiple goal state estimation has that alpha-beta filtering, alpha-beta-γ filtering, Kalman filtering, EKF, gaussian sum are approximate, optimum nonlinear filtering, particle filter and auto adapted filtering etc.Alpha-beta and alpha-beta-γ wave filter is because simple in structure, and calculated amount is little, uses very wide when computer resource is short in early days.Kalman filtering is a kind of basic skills of multiple target tracking, but need know the mathematical models of system, and only is applicable to linear system, has limited algorithm application.EKF expands to non-linear field with kalman filtering theory, is similar to the conditional probability distribution of state with a Gaussian distribution; And when approximate condition did not satisfy, the gaussian sum wave filter then was similar to the conditional probability distribution of state with the weighted sum of a Gaussian distribution.Optimum nonlinear filtering uses the Makov transition probability to describe the dynamic process of target, have good characteristic, but calculated amount is bigger, therefore never is used widely.Particle filter adopts stochastic sampling, because calculated amount is too big and the particle degenerate problem, is not suitable for practical application.In order to improve particle filter, Unscented kalman filtering adopts the determinacy sampling, makes sampled particle point number reduce, and avoided the particle point degenerate problem in the particle filter, so its application is very wide.Adaptive filter method is adjusted the state of filter parameter or increase wave filter in real time through the detection to target maneuver, makes wave filter adapt to target travel in real time, is particularly suitable for the tracking to maneuvering target; At present; Actual radar tracking system the most frequently used still be JPDA (Joint Probabilistic Data Association, JPDA) method (James A.Roecker, A Class of Near Optimal JPDA Algorithms; IEEE TRANSACTIONS ONAEROSPACE AND ELECTRONIC SYSTEMS; 1994, VOL.30 (2): 504-51O), other method great majority are to simplification of JPDA method etc.Yet, there are two positive semidefinite matrixs to subtract each other in the variance battle array that method errors such as JPDA are estimated, in the disposal system of limited wordlength, can produce the symmetric matrix that contains positive and negative eigenwert, cause the radar tracking enabling objective to lose and follow and whole radar property mistake.

Summary of the invention

Cause the radar tracking enabling objective to lose the technological deficiency of following in order to solve existing method for tracking target value structure instability; The present invention provides a kind of multi-object tracking method of svd; During this method was upgraded with measurement in the time of multiple target tracking; Through svd to the estimation error variance battle array, set up the numerical stability structural model, not having has two positive semidefinite matrixs to subtract each other in the variance battle array of estimation of error; In the disposal system of limited wordlength, can guarantee can not produce the symmetric matrix that contains negative eigenwert, can avoid the radar tracking enabling objective to lose and follow and whole radar property mistake.

The present invention solves the technical scheme that its technical matters adopts, a kind of multi-object tracking method of svd, and its characteristic may further comprise the steps:

1, the discretization model of i target does in N target following of definition

x _i(k+1)＝Φ(k+1，k)x _i(k)+Λω _i(k)，

Wherein:

Be state vector, (x, y z) are the position coordinates of target under the ground reference rectangular coordinate system, ω _i(k) the expression variance is Q _i(k) process noise vector, and Φ (k+1, k)=Φ=diag [Φ ₁, Φ ₁, Φ ₁] be state-transition matrix, $\mathrm{\Λ}={\∫}_{\mathrm{KT}}^{(k+1)T}\mathrm{\Φ}(k+1,\mathrm{\τ})\mathrm{\Γ}\left(\mathrm{\τ}\right)\mathrm{D\τ}=\left[\begin{array}{ccc}{\mathrm{\Λ}}_i& 0& 0\\ 0& {\mathrm{\Λ}}_i& 0\\ 0& 0& {\mathrm{\Λ}}_i\end{array}\right],$ Γ (t) is a matrix of coefficients, $\mathrm{\Γ}=\left[\begin{array}{ccc}{\mathrm{\Γ}}_1& 0& 0\\ 0& {\mathrm{\Γ}}_1& 0\\ 0& 0& {\mathrm{\Γ}}_1\end{array}\right],$ Γ ₁=[0 0 1] ^T, ${\mathrm{\Φ}}_1=\left[\begin{array}{ccc}1& T& \frac{1}{2}{T}^2\\ 0& 1& T\\ 0& 0& 1\\ & & \end{array}\right],$ ${\mathrm{\Λ}}_1={\left[\begin{array}{ccc}\frac{1}{6}{T}^3& \frac{1}{2}{T}^3& T\end{array}\right]}^T,$ T is the sampling period;

The time of i target is updated to:

x _i(k/k-1)＝Φx _i(k-1/k-1)

Obtain through calling singular value decomposition algorithm

$V_i(k/k-1)D_i(k/k-1)V_i^T(k/k-1)=Y(k/k-1)Y^T(k/k-1)$

Wherein: x _i(k/k-1) be that i target is at kT one-step prediction value constantly, V (k/k-1) D (k/k-1) V ^T(k/k-1) be the variance battle array of the one-step prediction error of correspondence, $Y(k/k-1)=\left[\begin{array}{cc}\mathrm{\Φ}{V}_i(k-1/k-1)D_i^{\frac{1}{2}}(k-1/k-1)& \mathrm{\Λ}{Q}_i^{\frac{1}{2}}(k-1)\end{array}\right],$ V _i(k/k-1) be orthogonal matrix, D _i(k/k-1) be diagonal matrix; Starting condition is x _i(0/0) and $V_i(0/0)D_i(0/0)V_i^T(0/0)=P_i(0/0);$

2, i target observation equation is: z _i(k)=g _i[x _i(k)]+v _i(k)

Wherein: z _i(k) be to the r of i target dimension observation vector, g _i[x _i(k)] be corresponding output, v _i(k) the expression variance is R _i(k) measure noise, obtain through calling singular value decomposition algorithm

${\stackrel{\‾}{V}}_i(k/k){\stackrel{\‾}{D}}_i(k/k){\stackrel{\‾}{V}}_i^T(k/k)=\stackrel{\‾}{Y}(k/k){\stackrel{\‾}{Y}}^T(k/k)$

$G_i\left(k\right)={\stackrel{\‾}{V}}_i(k/k){\stackrel{\‾}{D}}_i^{-1}(k/k){\stackrel{\‾}{V}}_i^T(k/k)R_i^{-1}\left(k\right)$

$x_i(k/k)=x_i(k/k-1)+G_i\left(k\right)\{\underset{j=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\λ}}_{\mathrm{ij}}\left(k\right)z_{\mathrm{ij}}\left(k\right)-g_i[x_i(k/k-1)\left]\right\}$

Wherein: $\stackrel{\‾}{Y}(k/k)=\left[\begin{array}{cc}{V}_i^T(k/k-1)D_i^{-\frac{1}{2}}(k/k-1)& H_i^T\left(k\right)\end{array}\right],$

Be orthogonal matrix, Be diagonal matrix; z _Ij(k) be radar to the j of i target (j=1,2 ..., m) individual echo, x _i(k/k) be i target kT filter value constantly, λ _Ij(k) be weight coefficient, and:

3, i Tracking Estimation method is: obtain through calling singular value decomposition algorithm

$V_i(k/k)D_i(k/k)V_i^T(k/k)=A\left(k\right)A^T\left(k\right)$

Wherein: $A\left(k\right)=\left[\begin{array}{cc}{\stackrel{\‾}{V}}_i^T(k/k-1){\stackrel{\‾}{D}}_i^{-\frac{1}{2}}(k/k-1)& G_i\left(k\right)d^T(I-{\mathrm{\Ω\; Uu}}^T)\end{array}\right],$

$u=\left[\begin{array}{c}1\\ 1\\ .\\ .\\ .\\ 1\end{array}\right],$ $d=\left[\begin{array}{c}{\mathrm{\Δ}}_{i,1}^T\left(k\right)\\ {\mathrm{\Δ}}_{i,2}^T\left(k\right)\\ .\\ .\\ .\\ {\mathrm{\Δ}}_{i,m}^T\left(k\right)\end{array}\right];$

Δ _{I, j}(k) be j candidate's echo information vector,

Δ _i，j(k)＝z _i，j(k)-g _i[x _i(k/k-1)]。

Useful result of the present invention is: through the estimation error variance battle array being carried out three times svd; Set up the multiple target tracking structural model of numerical stability; Avoided that two positive semidefinite matrixs subtract each other in the variance battle array of estimation of error; Numerical value in the disposal system of limited wordlength, can not occur and disperse, thereby guarantee the reliability of multi-object tracking method, avoided the radar tracking enabling objective lose with whole radar property mistake.

Below in conjunction with instance the present invention is elaborated.

Embodiment

1, the discretization model of i target does in N target following of definition

x _i(k+1)＝Φ(k+1，k)x _i(k)+Λω _i(k)，

Wherein:

Be state vector, (x, y z) are the position coordinates of target under the ground reference rectangular coordinate system, ω _i(k) the expression variance is Q _i(k) process noise vector, and Φ (k+1, k)=Φ=diag [Φ ₁, Φ ₁, Φ ₁] be state-transition matrix, $\mathrm{\Λ}={\∫}_{\mathrm{KT}}^{(k+1)T}\mathrm{\Φ}(k+1,\mathrm{\τ})\mathrm{\Γ}\left(\mathrm{\τ}\right)\mathrm{D\τ}=\left[\begin{array}{ccc}{\mathrm{\Λ}}_i& 0& 0\\ 0& {\mathrm{\Λ}}_i& 0\\ 0& 0& {\mathrm{\Λ}}_i\end{array}\right],$ Γ (t) is a matrix of coefficients, $\mathrm{\Γ}=\left[\begin{array}{ccc}{\mathrm{\Γ}}_1& 0& 0\\ 0& {\mathrm{\Γ}}_1& 0\\ 0& 0& {\mathrm{\Γ}}_1\end{array}\right],$ Γ ₁=[0 0 1] ^T, ${\mathrm{\Φ}}_1=\left[\begin{array}{ccc}1& T& \frac{1}{2}{T}^2\\ 0& 1& T\\ 0& 0& 1\\ & & \end{array}\right],$ ${\mathrm{\Λ}}_1={\left[\begin{array}{ccc}\frac{1}{6}{T}^3& \frac{1}{2}{T}^3& T\end{array}\right]}^T,$ T is the sampling period;

The time of i target is updated to:

x _i(k/k-1)＝Φx _i(k-1/k-1)

Obtain through calling singular value decomposition algorithm

$V_i(k/k-1)D_i(k/k-1)V_i^T(k/k-1)=Y(k/k-1)Y^T(k/k-1)$

Wherein: x _i(k/k-1) be that i target is at kT one-step prediction value constantly, V (k/k-1) D (k/k-1) V ^T(k/k-1) be the variance battle array of the one-step prediction error of correspondence, $Y(k/k-1)=\left[\begin{array}{cc}\mathrm{\Φ}{V}_i(k-1/k-1)D_i^{\frac{1}{2}}(k-1/k-1)& \mathrm{\Λ}{Q}_i^{\frac{1}{2}}(k-1)\end{array}\right],$ V _i(k/k-1) be orthogonal matrix, D _i(k/k-1) be diagonal matrix; Starting condition is x _i(0/0) and $V_i(0/0)D_i(0/0)V_i^T(0/0)=P_i(0/0);$

2, i target observation equation is: z _i(k)=g _i[x _i(k)]+v _i(k)

Wherein: z _i(k) be to the r of i target dimension observation vector, g _i[x _i(k)] be corresponding output, v _i(k) the expression variance is R _i(k) measure noise, obtain through calling singular value decomposition algorithm

${\stackrel{\‾}{V}}_i(k/k){\stackrel{\‾}{D}}_i(k/k){\stackrel{\‾}{V}}_i^T(k/k)=\stackrel{\‾}{Y}(k/k){\stackrel{\‾}{Y}}^T(k/k)$

$G_i\left(k\right)={\stackrel{\‾}{V}}_i(k/k){\stackrel{\‾}{D}}_i^{-1}(k/k){\stackrel{\‾}{V}}_i^T(k/k)R_i^{-1}\left(k\right)$

$x_i(k/k)=x_i(k/k-1)+G_i\left(k\right)\{\underset{j=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\λ}}_{\mathrm{ij}}\left(k\right)z_{\mathrm{ij}}\left(k\right)-g_i[x_i(k/k-1)\left]\right\}$

Wherein: $\stackrel{\‾}{Y}(k/k)=\left[\begin{array}{cc}{V}_i^T(k/k-1)D_i^{-\frac{1}{2}}(k/k-1)& H_i^T\left(k\right)\end{array}\right],$ Be orthogonal matrix,

Be diagonal matrix; z _Ij(k) be radar to the j of i target (j=1,2 ..., m) individual echo, x _i(k/k) be i target kT filter value constantly, λ _Ij(k) be weight coefficient, and: For example get g _i[x _i(k)]=[r _i(k) α _i(k) β _i(k)] ^T, r _iBe radar energy measurement oblique distance, α _iBe angular altitude, β _iThe position angle, and

$\left\{\begin{array}{c}{r}_i=\sqrt{x_i^2+y_i^2+z_i^2}\\ {\mathrm{\α}}_i={\tan }^{-1}\frac{z_i}{\sqrt{x_i^2+y_i^2}}\\ {\mathrm{\β}}_i={\tan }^{-1}\frac{x_i}{y_i}\end{array}\right.$

$H_i\left(k\right)=\frac{\∂g_i\left[x_i\left(k\right)\right]}{\∂x_i\left(k\right)}{|}_{x_i\left(k\right)=x_i(k/k-1)}$

$={\left[\begin{array}{ccccccccc}\frac{x_i}{\sqrt{x_i^2+y_i^2+z_i^2}}& 0& 0& \frac{y_i}{\sqrt{x_i^2+y_i^2+z_i^2}}& 0& 0& \frac{z_i}{\sqrt{x_i^2+y_i^2+z_i^2}}& 0& 0\\ \frac{-x_i{z}_i}{(x_i^2+y_i^2+z_i^2)\sqrt{x_i^2+y_i^2}}& 0& 0& \frac{-y_i{z}_i}{(x_i^2+y_i^2+z_i^2)\sqrt{x_i^2+y_i^2}}& 0& 0& \frac{\sqrt{x_i^2+y_i^2}}{(x_i^2+y_i^2+z_i^2)}& 0& 0\\ \frac{y_i}{x_i^2+y_i^2}& 0& 0& \frac{-x_i}{x_i^2+y_i^2}& 0& 0& 0& 0& 0\end{array}\right]}_{x_i\left(k\right)=x_i(k/k-1)};$

3, i Tracking Estimation method is: obtain through calling singular value decomposition algorithm

$V_i(k/k)D_i(k/k)V_i^T(k/k)=A\left(k\right)A^T\left(k\right)$

Wherein: $A\left(k\right)=\left[\begin{array}{cc}{\stackrel{\‾}{V}}_i^T(k/k-1){\stackrel{\‾}{D}}_i^{-\frac{1}{2}}(k/k-1)& G_i\left(k\right)d^T(I-{\mathrm{\Ω\; Uu}}^T)\end{array}\right],$

$u=\left[\begin{array}{c}1\\ 1\\ .\\ .\\ .\\ 1\end{array}\right],$ $d=\left[\begin{array}{c}{\mathrm{\Δ}}_{i,1}^T\left(k\right)\\ {\mathrm{\Δ}}_{i,2}^T\left(k\right)\\ .\\ .\\ .\\ {\mathrm{\Δ}}_{i,m}^T\left(k\right)\end{array}\right];$

Δ _{I, j}(k) be j candidate's echo information vector, Δ _{I, j}(k)=z _{I, j}(k)-g _i[x _i(k/k-1)].

Claims (1)

1. the multi-object tracking method of a svd is characterized in that may further comprise the steps:

(1), the discretization model of i target does in N target following of definition

x _i(k+1)＝Φ(k+1，k)x _i(k)+Λω _i(k)，

Wherein:

Be state vector, (x, y z) are the position coordinates of target under the ground reference rectangular coordinate system, ω _i(k) the expression variance is Q _i(k) process noise vector, and Φ (k+1, k)=Φ=diag [Φ ₁, Φ ₁, Φ ₁] be state-transition matrix, $\mathrm{\Λ}={\∫}_{\mathrm{KT}}^{(k+1)T}\mathrm{\Φ}(k+1,\mathrm{\τ})\mathrm{\Γ}\left(\mathrm{\τ}\right)\mathrm{D\τ}=\left[\begin{array}{ccc}{\mathrm{\Λ}}_i& 0& 0\\ 0& {\mathrm{\Λ}}_i& 0\\ 0& 0& {\mathrm{\Λ}}_i\end{array}\right],$ Γ (t) is a matrix of coefficients, $\mathrm{\Γ}=\left[\begin{array}{ccc}{\mathrm{\Γ}}_1& 0& 0\\ 0& {\mathrm{\Γ}}_1& 0\\ 0& 0& {\mathrm{\Γ}}_1\end{array}\right],$ Γ ₁=[0 0 1] ^T, ${\mathrm{\Φ}}_1=\left[\begin{array}{ccc}1& T& \frac{1}{2}{T}^2\\ 0& 1& T\\ 0& 0& 1\\ & & \end{array}\right],$ ${\mathrm{\Λ}}_1={\left[\begin{array}{ccc}\frac{1}{6}{T}^3& \frac{1}{2}{T}^3& T\end{array}\right]}^T,$ T is the sampling period;

The time of i target is updated to:

x _i(k/k-1)＝Φx _i(k-1/k-1)

Obtain through calling singular value decomposition algorithm

$V_i(k/k-1)D_i(k/k-1)V_i^T(k/k-1)=Y(k/k-1)Y^T(k/k-1)$

Wherein: x _i(k/k-1) be that i target is at kT one-step prediction value constantly, V (k/k-1) D (k/k-1) V ^T(k/k-1) be the variance battle array of the one-step prediction error of correspondence, $Y(k/k-1)=\left[\begin{array}{cc}\mathrm{\Φ}{V}_i(k-1/k-1)D_i^{\frac{1}{2}}(k-1/k-1)& \mathrm{\Λ}{Q}_i^{\frac{1}{2}}(k-1)\end{array}\right],$ V _i(k/k-1) be orthogonal matrix, D _i(k/k-1) be diagonal matrix; Starting condition is x _i(0/0) and $V_i(0/0)D_i(0/0)V_i^T(0/0)=P_i(0/0);$

(2), i target observation equation is: z _i(k)=g _i[x _i(k)]+v _i(k)

Wherein: z _i(k) be to the r of i target dimension observation vector, g _i[x _i(k)] be corresponding output, v _i(k) the expression variance is R _i(k) measure noise, obtain through calling singular value decomposition algorithm

${\stackrel{\‾}{V}}_i(k/k){\stackrel{\‾}{D}}_i(k/k){\stackrel{\‾}{V}}_i^T(k/k)=\stackrel{\‾}{Y}(k/k){\stackrel{\‾}{Y}}^T(k/k)$

$G_i\left(k\right)={\stackrel{\‾}{V}}_i(k/k){\stackrel{\‾}{D}}_i^{-1}(k/k){\stackrel{\‾}{V}}_i^T(k/k)R_i^{-1}\left(k\right)$

$x_i(k/k)=x_i(k/k-1)+G_i\left(k\right)\{\underset{j=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\λ}}_{\mathrm{ij}}\left(k\right)z_{\mathrm{ij}}\left(k\right)-g_i[x_i(k/k-1)\left]\right\}$

Wherein: $\stackrel{\‾}{Y}(k/k)=\left[\begin{array}{cc}{V}_i^T(k/k-1)D_i^{-\frac{1}{2}}(k/k-1)& H_i^T\left(k\right)\end{array}\right],$

Be orthogonal matrix,

Be diagonal matrix; z _Ij(k) be radar to the j of i target (j=1,2 ..., m) individual echo, x _i(k/k) be i target kT filter value constantly, λ _Ij(k) be weight coefficient, and: $\underset{j=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\λ}}_{i,j}(k+1)=1,$ $H_i\left(k\right)=\frac{{\∂g}_i\left[x_i\left(k\right)\right]}{{\∂x}_i\left(k\right)}{|}_{x_i\left(k\right)=x_i(k/k-1)};$

(3), i Tracking Estimation method is: obtain through calling singular value decomposition algorithm

$V_i(k/k)D_i(k/k)V_i^T(k/k)=A\left(k\right)A^T\left(k\right)$

Wherein: $A\left(k\right)=\left[\begin{array}{cc}{\stackrel{\‾}{V}}_i^T(k/k-1){\stackrel{\‾}{D}}_i^{-\frac{1}{2}}(k/k-1)& G_i\left(k\right)d^T(I-{\mathrm{\Ω\; Uu}}^T)\end{array}\right],$

$u=\left[\begin{array}{c}1\\ 1\\ .\\ .\\ .\\ 1\end{array}\right],$ $d=\left[\begin{array}{c}{\mathrm{\Δ}}_{i,1}^T\left(k\right)\\ {\mathrm{\Δ}}_{i,2}^T\left(k\right)\\ .\\ .\\ .\\ {\mathrm{\Δ}}_{i,m}^T\left(k\right)\end{array}\right];$

Δ _{I, j}(k) be j candidate's echo information vector,

Δ _i，j(k)＝z _i，j(k)-g _i[x _i(k/k-1)]。