卡尔曼滤波器Kalman Filter公式推导
- 引言
- 预测与校正
- 预测: x ^ n − 1 ∣ n − 1 → x ^ n ∣ n − 1 \hat{\mathrm{x}}_{\mathrm{n}-1 \mid \mathrm{n}-1} \rightarrow \hat{\mathrm{x}}_{\mathrm{n} \mid \mathrm{n}-1} x^n−1∣n−1→x^n∣n−1
- 校正: x ^ n ∣ n − 1 → x ^ n ∣ n \hat{\mathrm{x}}_{\mathrm{n} \mid \mathrm{n}-1} \rightarrow \hat{\mathrm{x}}_{\mathrm{n} \mid \mathrm{n}} x^n∣n−1→x^n∣n
- 卡尔曼增益 K n {K}_{n} Kn
- 预测误差的协方差阵 R ^ n ∣ n − 1 \hat{R}_{n \mid n-1} R^n∣n−1
引言
卡尔曼滤波器(Kalman Filter, KF)是一种状态空间模型,状态为内因素未知,观测为外因素可视为已知量,其模型如图所示:
状态
s
n
{s}_{n}
sn仅与其前一时刻状态
s
n
−
1
{s}_{n-1}
sn−1有关,与其历史状态无关,这里以线性关系为例,其表达式为:
{
X
n
=
F
n
X
n
−
1
+
V
n
Y
n
=
H
n
X
n
−
1
+
W
n
\left\{\begin{array}{c}\mathrm{X}_{n}=\mathrm{F}_{n} \mathrm{X}_{n-1}+\mathrm{V}_{n} \\ \mathrm{Y}_{n}=\mathrm{H}_{n} \mathrm{X}_{n-1}+\mathrm{W}_{n}\end{array}\right.
{Xn=FnXn−1+VnYn=HnXn−1+Wn
其中
X
n
{X}_{n}
Xn表示时刻 n 时的状态,
Y
n
{Y}_{n}
Yn表示时刻 n 时的观测值,
V
n
{V}_{n}
Vn为状态噪声,
W
n
{W}_{n}
Wn为观测噪声,假设
V
n
{V}_{n}
Vn与
W
n
{W}_{n}
Wn为零均值,且相互独立的高斯白噪声,即有:
{
E
(
V
n
)
=
E
(
W
n
)
=
0
E
(
V
n
W
n
)
=
0
E
(
W
n
W
m
)
=
Q
n
δ
n
m
E
(
V
n
V
m
)
=
T
n
δ
n
m
\left\{\begin{array}{c} \mathrm{E}\left(\mathrm{V}_{n}\right)=\mathrm{E}\left(\mathrm{W}_{n}\right)=0 \\ \mathrm{E}\left(\mathrm{V}_{n} \mathrm{W}_{n}\right)=0 \\ \mathrm{E}\left(\mathrm{W}_{n} \mathrm{W}_{\mathrm{m}}\right)=\mathrm{Q}_{n} \delta_{\mathrm{nm}} \\ \mathrm{E}\left(\mathrm{V}_{n} \mathrm{V}_{\mathrm{m}}\right)=\mathrm{T}_{n} \delta_{\mathrm{nm}} \end{array}\right.
⎩⎪⎪⎨⎪⎪⎧E(Vn)=E(Wn)=0E(VnWn)=0E(WnWm)=QnδnmE(VnVm)=Tnδnm
其中
Q
n
{Q}_{n}
Qn与
T
n
{T}_{n}
Tn分别表示
W
n
{W}_{n}
Wn与
V
n
{V}_{n}
Vn的自相关值,
δ
n
m
\delta_{\mathrm{nm}}
δnm表示
δ
n
m
=
{
1
n
=
m
0
otherwise
\delta_{\mathrm{nm}}=\left\{\begin{array}{cc} 1 & n=m \\ 0 & \text { otherwise } \end{array}\right.
δnm={10n=m otherwise
F
n
{F}_{n}
Fn,
H
n
{H}_{n}
Hn,
Q
n
{Q}_{n}
Qn,
T
n
{T}_{n}
Tn为系数可认为是确定性的常量(即不含随机性),而
X
n
{X}_{n}
Xn,
Y
n
{Y}_{n}
Yn,
V
n
{V}_{n}
Vn,
W
n
{W}_{n}
Wn为随机变量(注意在随机过程中务必清楚什么量是确定的,什么量是随机的!!!)。卡尔曼滤波器要做的即是从含有噪声的观测值中,尽可能准确地估计出当前时刻的状态。
这里简单介绍下投影的概念,使用前m个时刻的观测值
Y
1
,
Y
2
,
…
,
Y
m
\mathrm{Y}_{1}, \mathrm{Y}_{2}, \ldots, \mathrm{Y}_{\mathrm{m}}
Y1,Y2,…,Ym和第n时刻的状态值
X
n
{X}_{n}
Xn来估计n时刻的真实状态值
x
^
n
∣
m
\hat{\mathrm{x}}_{\mathrm{n} \mid \mathrm{m}}
x^n∣m,可表示为:
x
^
n
∣
m
=
Proj
{
Y
1
,
Y
2
,
…
,
Y
m
}
X
n
(
n
>
=
m
)
\hat{\mathrm{x}}_{\mathrm{n} \mid \mathrm{m}}=\operatorname{Proj}_{\left\{\mathrm{Y}_{1}, \mathrm{Y}_{2}, \ldots, \mathrm{Y}_{\mathrm{m}}\right\}} \mathrm{X}_{n} \quad(\mathrm{n}>=\mathrm{m})
x^n∣m=Proj{Y1,Y2,…,Ym}Xn(n>=m)
即
X
n
{X}_{n}
Xn在m维平面(
Y
1
,
Y
2
,
…
,
Y
m
\mathrm{Y}_{1}, \mathrm{Y}_{2}, \ldots, \mathrm{Y}_{\mathrm{m}}
Y1,Y2,…,Ym)作出的投影。
卡尔曼滤波的推导分为两个步骤,即:
- 预测: x ^ n − 1 ∣ n − 1 → x ^ n ∣ n − 1 \hat{\mathrm{x}}_{\mathrm{n}-1 \mid \mathrm{n}-1} \rightarrow \hat{\mathrm{x}}_{\mathrm{n} \mid \mathrm{n}-1} x^n−1∣n−1→x^n∣n−1,由n-1时刻的状态值预测第n时刻的状态值;
- 校正: x ^ n ∣ n − 1 → x ^ n ∣ n \hat{\mathrm{x}}_{\mathrm{n} \mid \mathrm{n}-1} \rightarrow \hat{\mathrm{x}}_{\mathrm{n} \mid \mathrm{n}} x^n∣n−1→x^n∣n,对之前的结果作出校正
预测与校正
预测: x ^ n − 1 ∣ n − 1 → x ^ n ∣ n − 1 \hat{\mathrm{x}}_{\mathrm{n}-1 \mid \mathrm{n}-1} \rightarrow \hat{\mathrm{x}}_{\mathrm{n} \mid \mathrm{n}-1} x^n−1∣n−1→x^n∣n−1
x ^ n ∣ n − 1 = Proj { Y 1 , Y 2 , … , Y n − 1 } X n = Proj { Y 1 , Y 2 , … , Y n − 1 } ( F n X n − 1 + V n ) = Proj { Y 1 , Y 2 , … , Y n − 1 } ( F n X n − 1 ) + Proj { Y 1 , Y 2 , … , Y n − 1 } V n = Proj { Y 1 , Y 2 , … , Y n − 1 } ( F n X n − 1 ) = F n Proj { Y 1 , Y 2 , … , Y n − 1 } X n − 1 = F n X ^ n − 1 ∣ n − 1 \begin{aligned} \hat{x}_{n \mid n-1} &=\operatorname{Proj}_{\left\{Y_{1}, Y_{2}, \ldots, Y_{n-1}\right\}} X_{n}=\operatorname{Proj}_{\left\{Y_{1}, Y_{2}, \ldots, Y_{n-1}\right\}}\left(F_{n} X_{n-1}+V_{n}\right) \\ &=\operatorname{Proj}_{\left\{Y_{1}, Y_{2}, \ldots, Y_{n-1}\right\}}\left(F_{n} X_{n-1}\right)+\operatorname{Proj}_{\left\{Y_{1}, Y_{2}, \ldots, Y_{n-1}\right\}} V_{n} \\ &=\operatorname{Proj}_{\left\{Y_{1}, Y_{2}, \ldots, Y_{n-1}\right\}}\left(F_{n} X_{n-1}\right) \\ &=F_{n} \operatorname{Proj}_{\left\{Y_{1}, Y_{2}, \ldots, Y_{n-1}\right\}} X_{n-1} \\ &=F_{n} \hat{X}_{n-1 \mid n-1} \end{aligned} x^n∣n−1=Proj{Y1,Y2,…,Yn−1}Xn=Proj{Y1,Y2,…,Yn−1}(FnXn−1+Vn)=Proj{Y1,Y2,…,Yn−1}(FnXn−1)+Proj{Y1,Y2,…,Yn−1}Vn=Proj{Y1,Y2,…,Yn−1}(FnXn−1)=FnProj{Y1,Y2,…,Yn−1}Xn−1=FnX^n−1∣n−1
注意V与Y是正交的,因此 Proj { Y 1 , Y 2 , … , Y n − 1 } V n = 0 \operatorname{Proj}_{\left\{Y_{1}, Y_{2}, \ldots, Y_{n-1}\right\}} V_{n}=0 Proj{Y1,Y2,…,Yn−1}Vn=0。
校正: x ^ n ∣ n − 1 → x ^ n ∣ n \hat{\mathrm{x}}_{\mathrm{n} \mid \mathrm{n}-1} \rightarrow \hat{\mathrm{x}}_{\mathrm{n} \mid \mathrm{n}} x^n∣n−1→x^n∣n
我们将
Y
n
{Y}_{n}
Yn进行拆分,
Y
n
=
Z
n
+
O
n
{Y}_{n}={Z}_{n}+{O}_{n}
Yn=Zn+On,其中
O
n
=
Proj
{
Y
1
,
Y
2
,
…
,
Y
n
−
1
}
Y
n
=
Proj
{
Y
1
,
Y
2
,
…
,
Y
n
−
1
}
(
H
n
X
n
+
W
n
)
=
H
n
Proj
{
Y
1
,
Y
2
,
…
,
Y
n
−
1
}
X
n
=
H
n
X
^
n
∣
n
−
1
\begin{aligned} O_{n} &=\operatorname{Proj}_{\left\{Y_{1}, Y_{2}, \ldots, Y_{n-1}\right\}} Y_{n} \\ &=\operatorname{Proj}_{\left\{Y_{1}, Y_{2}, \ldots, Y_{n-1}\right\}}\left(H_{n} X_{n}+W_{n}\right) \\ &=H_{n} \operatorname{Proj}_{\left\{Y_{1}, Y_{2}, \ldots, Y_{n-1}\right\}} X_{n} \\ &=H_{n} \hat{X}_{n \mid n-1} \end{aligned}
On=Proj{Y1,Y2,…,Yn−1}Yn=Proj{Y1,Y2,…,Yn−1}(HnXn+Wn)=HnProj{Y1,Y2,…,Yn−1}Xn=HnX^n∣n−1
Z
n
=
Y
n
−
O
n
=
Y
n
−
H
n
x
^
n
∣
n
−
1
{Z}_{n}={Y}_{n}-{O}_{n}={Y}_{n}-H_{n} \hat{x}_{n \mid n-1}
Zn=Yn−On=Yn−Hnx^n∣n−1表示投影误差(又称为新息,即新产生的信息)。
X
^
n
∣
n
=
Proj
{
Y
1
,
Y
2
,
…
,
Y
n
}
X
n
=
Proj
{
Y
1
,
Y
2
,
…
,
Y
n
−
1
}
X
n
+
Proj
Z
n
X
n
=
X
^
n
∣
n
−
1
+
Proj
z
n
X
n
\begin{aligned} \hat{X}_{n \mid n} &=\operatorname{Proj}_{\left\{Y_{1}, Y_{2}, \ldots, Y_{n}\right\}} X_{n} \\ &=\operatorname{Proj}_{\left\{Y_{1}, Y_{2}, \ldots, Y_{n-1}\right\}} X_{n}+\operatorname{Proj}_{Z_{n}} X_{n} \\ &=\hat{X}_{n \mid n-1}+\operatorname{Proj}_{z_{n}} X_{n} \end{aligned}
X^n∣n=Proj{Y1,Y2,…,Yn}Xn=Proj{Y1,Y2,…,Yn−1}Xn+ProjZnXn=X^n∣n−1+ProjznXn
Proj
z
n
X
n
\operatorname{Proj}_{z_{n}} X_{n}
ProjznXn即为校正量,即预测误差,有:
Proj
z
n
X
n
=
K
n
(
Y
n
−
H
n
X
^
n
∣
n
−
1
)
\operatorname{Proj}_{z_{n}} X_{n}=K_{n}\left(Y_{n}-H_{n} \hat{X}_{n \mid n-1}\right)
ProjznXn=Kn(Yn−HnX^n∣n−1)
其中
K
n
{K}_{n}
Kn称为卡尔曼增益,将在下节介绍。最终得到校正结果:
X
^
n
∣
n
=
X
^
n
∣
n
−
1
+
K
n
(
Y
n
−
H
n
X
^
n
∣
n
−
1
)
\hat{X}_{n \mid n}=\hat{X}_{n \mid n-1}+K_{n}\left(Y_{n}-H_{n} \hat{X}_{n \mid n-1}\right)
X^n∣n=X^n∣n−1+Kn(Yn−HnX^n∣n−1)
卡尔曼增益 K n {K}_{n} Kn
注意到上式得到的校正结果分为两部分,第一部分为预测模型,第二部分为观察模型,由卡尔曼增益
K
n
{K}_{n}
Kn调节两者的比例,即卡尔曼增益高越大表示更相信观察模型,反之则更相信预测模型,下面我们对该值进行求解。
K
n
=
E
(
X
n
Z
n
T
)
(
E
(
Z
n
Z
n
T
)
)
−
1
K_{n}=\mathrm{E}\left(X_{n} Z_{n}^{T}\right)\left(\mathrm{E}\left(Z_{n} Z_{n}^{T}\right)\right)^{-1}
Kn=E(XnZnT)(E(ZnZnT))−1
其中E表示求期望(均值),上角标T表示对矩阵求转置。
第一项
E
(
X
n
Z
n
T
)
=
E
(
X
n
(
Y
n
−
H
n
X
^
n
∣
n
−
1
)
T
)
=
E
(
X
n
(
H
n
X
n
+
W
n
−
H
n
X
^
n
∣
n
−
1
)
T
)
=
E
(
X
n
(
X
n
−
X
^
n
∣
n
−
1
)
T
H
n
T
)
+
E
(
X
n
W
n
T
)
=
E
(
X
n
(
X
n
−
X
^
n
∣
n
−
1
)
T
H
n
T
)
=
E
(
X
n
(
X
n
−
X
^
n
∣
n
−
1
)
)
H
n
T
\begin{aligned} \mathrm{E}\left(X_{n} Z_{n}^{T}\right) &=\mathrm{E}\left(X_{n}\left(Y_{n}-H_{n} \hat{X}_{n \mid n-1}\right)^{T}\right) \\ &=\mathrm{E}\left(X_{n}\left(H_{n} X_{n}+W_{n}-H_{n} \hat{X}_{n \mid n-1}\right)^{T}\right) \\ &=\mathrm{E}\left(X_{n}\left(X_{n}-\hat{X}_{n \mid n-1}\right)^{T} H_{n}^{T}\right)+\mathrm{E}\left(X_{n} W_{n}^{T}\right) \\ &=\mathrm{E}\left(X_{n}\left(X_{n}-\hat{X}_{n \mid n-1}\right)^{T} H_{n}^{T}\right) \\ &=\mathrm{E}\left(X_{n}\left(X_{n}-\hat{X}_{n \mid n-1}\right)\right) H_{n}^{T} \end{aligned}
E(XnZnT)=E(Xn(Yn−HnX^n∣n−1)T)=E(Xn(HnXn+Wn−HnX^n∣n−1)T)=E(Xn(Xn−X^n∣n−1)THnT)+E(XnWnT)=E(Xn(Xn−X^n∣n−1)THnT)=E(Xn(Xn−X^n∣n−1))HnT
注意到
X
^
n
∣
n
−
1
=
Proj
{
Y
1
,
Y
2
,
…
,
Y
n
−
1
}
X
n
\hat{X}_{n \mid n-1}=\operatorname{Proj}_{\left\{Y_{1}, Y_{2}, \ldots, Y_{n-1}\right\}} X_{n}
X^n∣n−1=Proj{Y1,Y2,…,Yn−1}Xn,即
X
^
n
∣
n
−
1
\hat{X}_{n \mid n-1}
X^n∣n−1是
X
n
X_{n}
Xn在超平面
Y
1
,
Y
2
,
…
,
Y
n
−
\mathrm{Y}_{1}, \mathrm{Y}_{2}, \ldots, \mathrm{Y}_{\mathrm{n-}}
Y1,Y2,…,Yn−上的投影,如图所示:
投影误差
X
n
−
X
^
n
∣
n
−
1
X_{n}-\hat{X}_{n \mid n-1}
Xn−X^n∣n−1与投影量
X
^
n
∣
n
−
1
\hat{X}_{n \mid n-1}
X^n∣n−1正交(此即正交性原理,即最优估计的误差正交于估计原材料),即
X
^
n
∣
n
−
1
(
X
n
−
X
^
n
∣
n
−
1
)
T
=
0
\hat{X}_{n \mid n-1}\left(X_{n}-\hat{X}_{n \mid n-1}\right)^{T}=0
X^n∣n−1(Xn−X^n∣n−1)T=0,因此上式可进一步改写为:
E
(
X
n
Z
n
T
)
=
E
(
(
X
n
−
X
^
n
∣
n
−
1
)
(
X
n
−
X
^
n
∣
n
−
1
)
T
)
H
n
T
\mathrm{E}\left(X_{n} Z_{n}^{T}\right)=\mathrm{E}\left(\left(X_{n}-\hat{X}_{n \mid n-1}\right)\left(X_{n}-\hat{X}_{n \mid n-1}\right)^{T}\right) H_{n}^{T}
E(XnZnT)=E((Xn−X^n∣n−1)(Xn−X^n∣n−1)T)HnT
其中
R
^
n
∣
n
−
1
=
E
(
(
X
n
−
X
^
n
∣
n
−
1
)
(
X
n
−
X
^
n
∣
n
−
1
)
T
)
\hat{R}_{n \mid n-1}=\mathrm{E}\left(\left(X_{n}-\hat{X}_{n \mid n-1}\right)\left(X_{n}-\hat{X}_{n \mid n-1}\right)^{T}\right)
R^n∣n−1=E((Xn−X^n∣n−1)(Xn−X^n∣n−1)T)为预测误差的协方差矩阵,因此:
E
(
X
n
Z
n
T
)
=
R
^
n
∣
n
−
1
H
n
T
\mathrm{E}\left(X_{n} Z_{n}^{T}\right)=\widehat{R}_{n \mid n-1} H_{n}^{T}
E(XnZnT)=R
n∣n−1HnT
第二项
E
(
Z
n
Z
n
T
)
=
E
(
(
Y
n
−
H
n
X
^
n
∣
n
−
1
)
(
Y
n
−
H
n
X
^
n
∣
n
−
1
)
T
)
=
E
(
(
H
n
X
n
+
W
n
−
H
n
X
^
n
∣
n
−
1
)
(
H
n
X
n
+
W
n
−
H
n
X
^
n
∣
n
−
1
)
T
)
=
H
n
E
(
(
X
n
−
X
^
n
∣
n
−
1
)
(
X
n
−
X
^
n
∣
n
−
1
)
)
H
n
T
+
E
(
W
n
W
n
T
)
=
H
n
R
^
n
∣
n
−
1
H
n
T
+
Q
n
\begin{aligned} \mathrm{E}\left(Z_{n} Z_{n}^{T}\right) &=E\left(\left(Y_{n}-H_{n} \hat{X}_{n \mid n-1}\right)\left(Y_{n}-H_{n} \hat{X}_{n \mid n-1}\right)^{T}\right) \\ &=E\left(\left(H_{n} X_{n}+W_{n}-H_{n} \hat{X}_{n \mid n-1}\right)\left(H_{n} X_{n}+W_{n}-H_{n} \hat{X}_{n \mid n-1}\right)^{T}\right) \\ &=H_{n} E\left(\left(X_{n}-\hat{X}_{n \mid n-1}\right)\left(X_{n}-\hat{X}_{n \mid n-1}\right)\right) H_{n}^{T}+E\left(W_{n} W_{n}^{T}\right) \\ &=H_{n} \hat{R}_{n \mid n-1} H_{n}^{T}+Q_{n} \end{aligned}
E(ZnZnT)=E((Yn−HnX^n∣n−1)(Yn−HnX^n∣n−1)T)=E((HnXn+Wn−HnX^n∣n−1)(HnXn+Wn−HnX^n∣n−1)T)=HnE((Xn−X^n∣n−1)(Xn−X^n∣n−1))HnT+E(WnWnT)=HnR^n∣n−1HnT+Qn
综上可以得到卡尔曼增益的表达式:
K
n
=
R
^
n
∣
n
−
1
H
n
T
(
H
n
R
^
n
∣
n
−
1
H
n
T
+
Q
n
)
−
1
K_{n}=\hat{R}_{n \mid n-1} H_{n}^{T}\left(H_{n} \hat{R}_{n \mid n-1} H_{n}^{T}+Q_{n}\right)^{-1}
Kn=R^n∣n−1HnT(HnR^n∣n−1HnT+Qn)−1
预测误差的协方差阵 R ^ n ∣ n − 1 \hat{R}_{n \mid n-1} R^n∣n−1
同求解状态过程步骤类似,对协方差阵 R ^ n ∣ n − 1 \hat{R}_{n \mid n-1} R^n∣n−1的求解同样分为预测、校正两部分,即:
- R ^ n ∣ n − 1 → R ^ n ∣ n \hat{\mathrm{R}}_{\mathrm{n} \mid \mathrm{n}-1} \rightarrow \hat{\mathrm{R}}_{\mathrm{n} \mid \mathrm{n}} R^n∣n−1→R^n∣n;
- R ^ n ∣ n → R ^ n + 1 ∣ n \hat{\mathrm{R}}_{\mathrm{n} \mid \mathrm{n}} \rightarrow \hat{\mathrm{R}}_{\mathrm{n+1} \mid \mathrm{n}} R^n∣n→R^n+1∣n
R ^ n + 1 ∣ n = E ( ( X n + 1 − X ^ n + 1 ∣ n ) ( X n + 1 − X ^ n + 1 ∣ n ) T ) = E ( ( F n + 1 X n + V n + 1 − F n + 1 X ^ n ∣ n ) ( F n + 1 X n + V n + 1 − F n + 1 X ^ n ∣ n ) T ) = E ( ( F n + 1 ( X n − X ^ n ∣ n ) + V n + 1 ) ( F n + 1 ( X n − X ^ n ∣ n ) + V n + 1 ) T ) = F n + 1 E ( ( X n − X ^ n ∣ n ) ( X n − X ^ n ∣ n ) T F n + 1 T + E ( V n + 1 V n + 1 T ) = F n + 1 R ^ n ∣ n F n + 1 T + T n + 1 \begin{aligned} \hat{R}_{n+1 \mid n} &=\mathrm{E}\left(\left(X_{n+1}-\hat{X}_{n+1 \mid n}\right)\left(X_{n+1}-\hat{X}_{n+1 \mid n}\right)^{T}\right) \\ &=\mathrm{E}\left(\left(F_{n+1} X_{n}+V_{n+1}-F_{n+1} \hat{X}_{n \mid n}\right)\left(F_{n+1} X_{n}+V_{n+1}-F_{n+1} \hat{X}_{n \mid n}\right)^{T}\right) \\ &=\mathrm{E}\left(\left(F_{n+1}\left(X_{n}-\hat{X}_{n \mid n}\right)+V_{n+1}\right)\left(F_{n+1}\left(X_{n}-\hat{X}_{n \mid n}\right)+V_{n+1}\right)^{T}\right) \\ &=F_{n+1} \mathrm{E}\left(\left(X_{n}-\hat{X}_{n \mid n}\right)\left(X_{n}-\hat{X}_{n \mid n}\right)^{T} F_{n+1}^{T}+\mathrm{E}\left(V_{n+1} V_{n+1}^{T}\right)\right.\\ &=F_{n+1} \hat{R}_{n \mid n} F_{n+1}^{T}+T_{n+1} \end{aligned} R^n+1∣n=E((Xn+1−X^n+1∣n)(Xn+1−X^n+1∣n)T)=E((Fn+1Xn+Vn+1−Fn+1X^n∣n)(Fn+1Xn+Vn+1−Fn+1X^n∣n)T)=E((Fn+1(Xn−X^n∣n)+Vn+1)(Fn+1(Xn−X^n∣n)+Vn+1)T)=Fn+1E((Xn−X^n∣n)(Xn−X^n∣n)TFn+1T+E(Vn+1Vn+1T)=Fn+1R^n∣nFn+1T+Tn+1
R ^ n ∣ n = E ( ( X n − X ^ n ∣ n ) ( X n − X ^ n ∣ n ) T ) = E ( ( X n − X ^ n ∣ n − 1 − K n ( Y n − H n X ^ n ∣ n − 1 ) ) ( X n − X ^ n ∣ n − 1 − K n ( Y n − H n X ^ n ∣ n − 1 ) ) T ) = E ( ( X n − X ^ n ∣ n − 1 − K n ( H n X n + W n ) + K n H n X ^ n ∣ n − 1 ) ( X n − X ^ n ∣ n − 1 − K n ( H n X n + W n ) + K n H n X ^ n ∣ n − 1 ) T ) = E ( ( I − K n H n ) ( X n − X ^ n ∣ n − 1 ) − K n W n ) ( ( I − K n H n ) ( X n − X ^ n ∣ n − 1 ) − K n W n ) T ) = ( I − K n H n ) E ( ( X n − X ^ n ∣ n − 1 ) ( X n − X ^ n ∣ n − 1 ) T ) ( I − K n H n ) T + K n Q n K n T = ( I − K n H n ) R ^ n ∣ n − 1 ( I − K n H n ) T + K n Q n K n T = ( I − K n H n ) R ^ n ∣ n − 1 \begin{aligned} \hat{R}_{n \mid n} &=\mathrm{E}\left(\left(X_{n}-\hat{X}_{n \mid n}\right)\left(X_{n}-\hat{X}_{n \mid n}\right)^{T}\right) \\ &=\mathrm{E}\left(\left(X_{n}-\hat{X}_{n \mid n-1}-K_{n}\left(Y_{n}-H_{n} \hat{X}_{n \mid n-1}\right)\right)\left(X_{n}-\hat{X}_{n \mid n-1}-K_{n}\left(Y_{n}-H_{n} \hat{X}_{n \mid n-1}\right)\right)^{T}\right) \\ &=E\left(\left(X_{n}-\hat{X}_{n \mid n-1}-K_{n}\left(H_{n} X_{n}+W_{n}\right)+K_{n} H_{n} \hat{X}_{n \mid n-1}\right)\left(X_{n}-\hat{X}_{n \mid n-1}-K_{n}\left(H_{n} X_{n}+W_{n}\right)+K_{n} H_{n} \hat{X}_{n \mid n-1}\right)^{T}\right) \\ &\left.=E\left(\left(I-K_{n} H_{n}\right)\left(X_{n}-\hat{X}_{n \mid n-1}\right)-K_{n} W_{n}\right)\left(\left(I-K_{n} H_{n}\right)\left(X_{n}-\hat{X}_{n \mid n-1}\right)-K_{n} W_{n}\right)^{T}\right) \\ &=\left(I-K_{n} H_{n}\right) E\left(\left(X_{n}-\hat{X}_{n \mid n-1}\right)\left(X_{n}-\hat{X}_{n \mid n-1}\right)^{T}\right)\left(I-K_{n} H_{n}\right)^{T}+K_{n} Q_{n} K_{n}^{T} \\ &=\left(I-K_{n} H_{n}\right) \hat{R}_{n \mid n-1}\left(I-K_{n} H_{n}\right)^{T}+K_{n} Q_{n} K_{n}^{T} \\ &=\left(I-K_{n} H_{n}\right) \hat{R}_{n \mid n-1} \end{aligned} R^n∣n=E((Xn−X^n∣n)(Xn−X^n∣n)T)=E((Xn−X^n∣n−1−Kn(Yn−HnX^n∣n−1))(Xn−X^n∣n−1−Kn(Yn−HnX^n∣n−1))T)=E((Xn−X^n∣n−1−Kn(HnXn+Wn)+KnHnX^n∣n−1)(Xn−X^n∣n−1−Kn(HnXn+Wn)+KnHnX^n∣n−1)T)=E((I−KnHn)(Xn−X^n∣n−1)−KnWn)((I−KnHn)(Xn−X^n∣n−1)−KnWn)T)=(I−KnHn)E((Xn−X^n∣n−1)(Xn−X^n∣n−1)T)(I−KnHn)T+KnQnKnT=(I−KnHn)R^n∣n−1(I−KnHn)T+KnQnKnT=(I−KnHn)R^n∣n−1