卡尔曼滤波器的参数
首先回顾卡尔曼滤波的概率图:
基于之前内容,卡尔曼滤波器是一种动态图模型,除了遵循齐次马尔科夫假设和观测独立性假设外,卡尔曼滤波器还遵循
Z
t
Z_{t}
Zt和
Z
t
−
1
Z_{t-1}
Zt−1之间,
X
t
X_{t}
Xt和
Z
t
Z_{t}
Zt之间的线性关系:
Z
t
=
A
Z
t
−
1
+
B
+
ε
Z_{t}=AZ_{t-1}+B+\varepsilon
Zt=AZt−1+B+ε
X
t
=
C
Z
t
+
D
+
δ
X_{t}=CZ_{t}+D+\delta
Xt=CZt+D+δ
这里,
(
A
,
B
,
C
,
D
)
(A,B,C,D)
(A,B,C,D)是线性系数,
(
ε
,
δ
)
(\varepsilon,\delta)
(ε,δ)是两个满足均值为0的高斯分布随机变量,记作:
ε
∼
N
(
0
,
Q
)
,
δ
∼
N
(
0
,
R
)
\varepsilon\sim N(0,Q),\delta\sim N(0,R)
ε∼N(0,Q),δ∼N(0,R)
另外,卡尔曼滤波器中,无论是隐变量
Z
Z
Z,还是观测变量
X
X
X,都服从高斯分布:
p
(
z
t
∣
z
t
−
1
)
=
N
(
A
z
t
−
1
+
B
,
Q
)
p(z_{t}|z_{t-1})=N(Az_{t-1}+B,Q)
p(zt∣zt−1)=N(Azt−1+B,Q)
p
(
x
t
∣
z
t
)
=
N
(
C
z
t
+
D
,
R
)
p(x_{t}|z_{t})=N(Cz_{t}+D,R)
p(xt∣zt)=N(Czt+D,R)
对于模型的第一个变量,即隐变量
z
1
z_{1}
z1,我们直接令其服从高斯分布:
p
(
z
1
)
=
N
(
μ
1
,
Σ
1
)
p(z_{1})=N(\mu_{1},\Sigma_{1})
p(z1)=N(μ1,Σ1),这时,已经可以明确卡尔曼滤波器的参数为:
θ
=
(
A
,
B
,
C
,
D
,
Q
,
R
,
μ
1
,
Σ
1
)
\theta=(A,B,C,D,Q,R,\mu_{1},\Sigma_{1})
θ=(A,B,C,D,Q,R,μ1,Σ1)
模型明确后,我们需要关注推理问题中的滤波问题,即计算概率
p
(
z
t
∣
x
1
,
x
2
,
.
.
.
,
x
t
)
p(z_{t}|x_{1},x_{2},...,x_{t})
p(zt∣x1,x2,...,xt)的过程。
滤波问题的解决方案
卡尔曼滤波器按照时间的推演顺序,我们依次观察到观测变量 x 1 , x 2 , . . . , x T x_{1},x_{2},...,x_{T} x1,x2,...,xT,在观察到 x t x_{t} xt前,我们会先观察到 x t − 1 x_{t-1} xt−1,我们应该建立 p ( z t ∣ x 1 , x 2 , . . . , x t ) p(z_{t}|x_{1},x_{2},...,x_{t}) p(zt∣x1,x2,...,xt)和 p ( z t − 1 ∣ x 1 , x 2 , . . . , x t − 1 ) p(z_{t-1}|x_{1},x_{2},...,x_{t-1}) p(zt−1∣x1,x2,...,xt−1)这前后相邻两个时刻的滤波表达式间的递推关系,关系建立后,我们就能从 p ( z 1 ∣ x 1 ) p(z_{1}|x_{1}) p(z1∣x1)依次推导到 p ( z t ∣ x 1 , x 2 , . . . , x t ) p(z_{t}|x_{1},x_{2},...,x_{t}) p(zt∣x1,x2,...,xt)。
首先依据贝叶斯公式有:
p
(
z
t
∣
x
1
,
x
2
,
.
.
.
,
x
t
)
=
p
(
x
1
,
x
2
,
.
.
.
,
x
t
,
z
t
)
p
(
x
1
,
x
2
,
.
.
.
,
x
t
)
p(z_{t}|x_{1},x_{2},...,x_{t})=\frac{p(x_{1},x_{2},...,x_{t},z_{t})}{p(x_{1},x_{2},...,x_{t})}
p(zt∣x1,x2,...,xt)=p(x1,x2,...,xt)p(x1,x2,...,xt,zt)
由于
p
(
x
1
,
x
2
,
.
.
.
,
x
t
)
p(x_{1},x_{2},...,x_{t})
p(x1,x2,...,xt)描述的是这一组观测变量的概率,因此是一个常数,所以就转换为以下正比关系:
p
(
z
t
∣
x
1
,
x
2
,
.
.
.
,
x
t
)
∝
p
(
x
1
,
x
2
,
.
.
.
,
x
t
,
z
t
)
p(z_{t}|x_{1},x_{2},...,x_{t})\propto p(x_{1},x_{2},...,x_{t},z_{t})
p(zt∣x1,x2,...,xt)∝p(x1,x2,...,xt,zt)
再次依据贝叶斯定理展开:
p
(
x
1
,
x
2
,
.
.
.
,
x
t
,
z
t
)
=
p
(
x
t
∣
x
1
,
x
2
,
.
.
.
,
x
t
−
1
,
z
t
)
p
(
x
1
,
x
2
,
.
.
.
,
x
t
−
1
,
z
t
)
p(x_{1},x_{2},...,x_{t},z_{t})=p(x_{t}|x_{1},x_{2},...,x_{t-1},z_{t})p(x_{1},x_{2},...,x_{t-1},z_{t})
p(x1,x2,...,xt,zt)=p(xt∣x1,x2,...,xt−1,zt)p(x1,x2,...,xt−1,zt)
=
p
(
x
t
∣
z
t
)
p
(
x
1
,
x
2
,
.
.
.
,
x
t
−
1
,
z
t
)
=p(x_{t}|z_{t})p(x_{1},x_{2},...,x_{t-1},z_{t})
=p(xt∣zt)p(x1,x2,...,xt−1,zt)
再展开:
p
(
x
1
,
x
2
,
.
.
.
,
x
t
−
1
,
z
t
)
=
p
(
z
t
∣
x
1
,
x
2
,
.
.
.
,
x
t
−
1
)
p
(
x
1
,
x
2
,
.
.
.
,
x
t
−
1
)
p(x_{1},x_{2},...,x_{t-1},z_{t})=p(z_{t}|x_{1},x_{2},...,x_{t-1})p(x_{1},x_{2},...,x_{t-1})
p(x1,x2,...,xt−1,zt)=p(zt∣x1,x2,...,xt−1)p(x1,x2,...,xt−1)
所以有:
p
(
x
1
,
x
2
,
.
.
.
,
x
t
,
z
t
)
=
p
(
x
t
∣
z
t
)
p
(
z
t
∣
x
1
,
x
2
,
.
.
.
,
x
t
−
1
)
p
(
x
1
,
x
2
,
.
.
.
x
t
−
1
)
p(x_{1},x_{2},...,x_{t},z_{t})=p(x_{t}|z_{t})p(z_{t}|x_{1},x_{2},...,x_{t-1})p(x_{1},x_{2},...x_{t-1})
p(x1,x2,...,xt,zt)=p(xt∣zt)p(zt∣x1,x2,...,xt−1)p(x1,x2,...xt−1)
该式子包含了两个概念:
- 第一: p ( z t ∣ x 1 , x 2 , . . . , x t − 1 ) p(z_{t}|x_{1},x_{2},...,x_{t-1}) p(zt∣x1,x2,...,xt−1)是预测问题;
- 第二: p ( x 1 , x 2 , . . . , x t − 1 ) p(x_{1},x_{2},...,x_{t-1}) p(x1,x2,...,xt−1)是关于观测变量的预测,是一个常数,因此式子可以用正比关系表达;
因此有:
p
(
x
1
,
x
2
,
.
.
.
,
x
t
,
z
t
)
∝
p
(
x
t
∣
z
t
)
p
(
z
t
∣
x
1
,
x
2
,
.
.
.
,
x
t
−
1
)
p(x_{1},x_{2},...,x_{t},z_{t})\propto p(x_{t}|z_{t})p(z_{t}|x_{1},x_{2},...,x_{t-1})
p(x1,x2,...,xt,zt)∝p(xt∣zt)p(zt∣x1,x2,...,xt−1)
⇒
p
(
z
t
∣
x
1
,
x
2
,
.
.
.
,
x
t
)
∝
p
(
x
1
,
x
2
,
.
.
.
,
x
t
,
z
t
)
∝
p
(
x
t
∣
z
t
)
p
(
z
t
∣
x
1
,
x
2
,
.
.
.
,
x
t
−
1
)
\Rightarrow p(z_{t}|x_{1},x_{2},...,x_{t})\propto p(x_{1},x_{2},...,x_{t},z_{t})\propto p(x_{t}|z_{t})p(z_{t}|x_{1},x_{2},...,x_{t-1})
⇒p(zt∣x1,x2,...,xt)∝p(x1,x2,...,xt,zt)∝p(xt∣zt)p(zt∣x1,x2,...,xt−1)
然后思考预测问题的概率化简:
p
(
z
t
∣
x
1
,
x
2
,
.
.
.
,
x
t
−
1
)
=
∫
z
t
−
1
p
(
z
t
,
z
t
−
1
∣
x
1
,
x
2
,
.
.
.
,
x
t
−
1
)
d
z
t
−
1
p(z_{t}|x_{1},x_{2},...,x_{t-1})=\int_{z_{t-1}}p(z_{t},z_{t-1}|x_{1},x_{2},...,x_{t-1})dz_{t-1}
p(zt∣x1,x2,...,xt−1)=∫zt−1p(zt,zt−1∣x1,x2,...,xt−1)dzt−1
=
∫
z
t
−
1
p
(
z
t
∣
z
t
−
1
,
x
1
,
x
2
,
.
.
.
,
x
t
−
1
)
p
(
z
t
−
1
∣
x
1
,
x
2
,
.
.
.
,
x
t
−
1
)
d
z
t
−
1
=\int_{z_{t-1}}p(z_{t}|z_{t-1},x_{1},x_{2},...,x_{t-1})p(z_{t-1}|x_{1},x_{2},...,x_{t-1})dz_{t-1}
=∫zt−1p(zt∣zt−1,x1,x2,...,xt−1)p(zt−1∣x1,x2,...,xt−1)dzt−1
=
∫
z
t
−
1
p
(
z
t
∣
z
t
−
1
)
p
(
z
t
−
1
∣
x
1
,
x
2
,
.
.
.
,
x
t
−
1
)
d
z
t
−
1
=\int_{z_{t-1}}p(z_{t}|z_{t-1})p(z_{t-1}|x_{1},x_{2},...,x_{t-1})dz_{t-1}
=∫zt−1p(zt∣zt−1)p(zt−1∣x1,x2,...,xt−1)dzt−1
此时,得到两个具有明显含义的式子:
- 基于马尔科夫假设化简得到的 p ( z t ∣ z t − 1 ) p(z_{t}|z_{t-1}) p(zt∣zt−1);
- t − 1 t-1 t−1时刻的滤波问题表达式 p ( z t − 1 ∣ x 1 , x 2 , . . . , x t − 1 ) p(z_{t-1}|x_{1},x_{2},...,x_{t-1}) p(zt−1∣x1,x2,...,xt−1);
卡尔曼滤波的predict和update
在递推的过程中,
t
−
1
t-1
t−1时刻,总是先利用
t
−
1
t-1
t−1时刻的滤波结果,在
t
−
1
t-1
t−1时刻的观测序列
x
1
,
x
2
,
.
.
.
,
x
t
−
1
x_{1},x_{2},...,x_{t-1}
x1,x2,...,xt−1的基础上,去预测下一时刻
t
t
t的状态
z
t
z_{t}
zt的概率,这一步,称为predict:
p
(
z
t
∣
x
1
,
x
2
,
.
.
.
,
x
t
−
1
)
=
∫
z
t
−
1
p
(
z
t
∣
z
t
−
1
)
p
(
z
t
−
1
∣
x
1
,
x
2
,
.
.
.
,
x
t
−
1
)
d
z
t
−
1
p(z_{t}|x_{1},x_{2},...,x_{t-1})=\int_{z_{t-1}}p(z_{t}|z_{t-1})p(z_{t-1}|x_{1},x_{2},...,x_{t-1})dz_{t-1}
p(zt∣x1,x2,...,xt−1)=∫zt−1p(zt∣zt−1)p(zt−1∣x1,x2,...,xt−1)dzt−1
得到
t
−
1
t-1
t−1时刻对
t
t
t时刻状态的预测后,进入
t
t
t时刻,此时,新增
t
t
t时刻的观测值
x
t
x_{t}
xt,我们通过它来对
t
−
1
t-1
t−1时刻的预测进行修正,也就是修正隐变量
z
t
z_{t}
zt的概率,这一步称为update:
p
(
z
t
∣
x
1
,
x
2
,
.
.
.
,
x
t
)
∝
p
(
x
t
∣
z
t
)
p
(
z
t
∣
x
1
,
x
2
,
.
.
.
,
x
t
−
1
)
p(z_{t}|x_{1},x_{2},...,x_{t})\propto p(x_{t}|z_{t})p(z_{t}|x_{1},x_{2},...,x_{t-1})
p(zt∣x1,x2,...,xt)∝p(xt∣zt)p(zt∣x1,x2,...,xt−1)
直观地,从
t
=
1
t=1
t=1演示以上过程:
t
=
1
t=1
t=1时刻:
update:计算
p
(
z
1
∣
x
1
)
p(z_{1}|x_{1})
p(z1∣x1);
predict:计算
p
(
z
2
∣
x
1
)
p(z_{2}|x_{1})
p(z2∣x1);
t
=
2
t=2
t=2时刻:
update:计算
p
(
z
2
∣
x
2
,
x
1
)
p(z_{2}|x_{2},x_{1})
p(z2∣x2,x1);
predict:计算
p
(
z
3
∣
x
2
,
x
1
)
p(z_{3}|x_{2},x_{1})
p(z3∣x2,x1);
然后不断按照 t = 3 , 4 , . . . , n − 1 , n t=3,4,...,n-1,n t=3,4,...,n−1,n迭代;对于 t = 1 t=1 t=1的启动步update, p ( z 1 ∣ x 1 ) ∝ p ( x 1 ∣ z 1 ) p ( z 1 ) p(z_{1}|x_{1})\propto p(x_{1}|z_{1})p(z_{1}) p(z1∣x1)∝p(x1∣z1)p(z1),隐状态 z 1 z_{1} z1会是一个提前给定的正态分布。
基于正态分布得到解析结果
证明解析解存在
在上面看到,predict和update都是比较复杂的运算,实际处理时,我们将基于正态分布的性质得到解析结果。
卡尔曼滤波器中有重要假设,即服从带高斯分布噪声的线性关系:
p
(
z
t
∣
z
t
−
1
)
=
N
(
A
z
t
−
1
+
B
,
Q
)
p(z_{t}|z_{t-1})=N(Az_{t-1}+B,Q)
p(zt∣zt−1)=N(Azt−1+B,Q)
p
(
x
t
∣
z
t
)
=
N
(
C
z
t
+
D
,
R
)
p(x_{t}|z_{t})=N(Cz_{t}+D,R)
p(xt∣zt)=N(Czt+D,R)
p
(
z
1
)
=
N
(
μ
1
,
Σ
1
)
p(z_{1})=N(\mu_{1},\Sigma_{1})
p(z1)=N(μ1,Σ1)
另外已知,高斯分布的联合概率,边缘概率,条件概率依然是高斯分布。
从
t
=
1
t=1
t=1时刻看:
update:
p
(
z
1
∣
x
1
)
∝
p
(
x
1
∣
z
1
)
p
(
z
1
)
p(z_{1}|x_{1})\propto p(x_{1}|z_{1})p(z_{1})
p(z1∣x1)∝p(x1∣z1)p(z1)
其中,
p
(
x
1
∣
z
1
)
p(x_{1}|z_{1})
p(x1∣z1)和
p
(
z
1
)
p(z_{1})
p(z1)都服从正态分布,它们联合概率的乘积
p
(
z
1
∣
x
1
)
p(z_{1}|x_{1})
p(z1∣x1)也服从正态分布。
predict:
p
(
z
2
∣
x
1
)
=
∫
z
1
p
(
z
2
∣
z
1
)
p
(
z
1
∣
x
1
)
d
z
1
p(z_{2}|x_{1})=\int_{z_{1}}p(z_{2}|z_{1})p(z_{1}|x_{1})dz_{1}
p(z2∣x1)=∫z1p(z2∣z1)p(z1∣x1)dz1
这里
p
(
z
1
∣
x
1
)
p(z_{1}|x_{1})
p(z1∣x1)和
p
(
z
2
∣
z
1
)
p(z_{2}|z_{1})
p(z2∣z1)都服从正态分布,所以联合概率
p
(
z
2
∣
z
1
)
p
(
z
1
∣
x
1
)
p(z_{2}|z_{1})p(z_{1}|x_{1})
p(z2∣z1)p(z1∣x1)也服从正态分布,边缘概率
p
(
z
2
∣
x
1
)
=
∫
z
1
p
(
z
2
∣
z
1
)
p
(
z
1
∣
x
1
)
d
z
1
p(z_{2}|x_{1})=\int_{z_{1}}p(z_{2}|z_{1})p(z_{1}|x_{1})dz_{1}
p(z2∣x1)=∫z1p(z2∣z1)p(z1∣x1)dz1也服从正态分布。
过度到
t
−
1
t-1
t−1时刻的predict步:
p
(
z
t
∣
x
1
,
.
.
.
,
x
t
−
1
)
=
∫
z
t
−
1
p
(
z
t
∣
z
t
−
1
)
p
(
z
t
−
1
∣
x
1
,
x
2
,
.
.
.
,
x
t
−
1
)
d
z
t
−
1
p(z_{t}|x_{1},...,x_{t-1})=\int_{z_{t-1}}p(z_{t}|z_{t-1})p(z_{t-1}|x_{1},x_{2},...,x_{t-1})dz_{t-1}
p(zt∣x1,...,xt−1)=∫zt−1p(zt∣zt−1)p(zt−1∣x1,x2,...,xt−1)dzt−1
p
(
z
t
∣
z
t
−
1
)
p(z_{t}|z_{t-1})
p(zt∣zt−1)和
p
(
z
t
−
1
∣
x
1
,
.
.
.
,
x
t
−
1
)
p(z_{t-1}|x_{1},...,x_{t-1})
p(zt−1∣x1,...,xt−1)都服从正态分布,所以其联合概率,边缘概率也服从。
再进入
t
t
t时刻的update步:
p
(
z
t
∣
x
1
,
.
.
.
,
x
t
)
∝
p
(
x
t
∣
z
t
)
p
(
z
t
∣
x
1
,
.
.
.
,
x
t
−
1
)
p(z_{t}|x_{1},...,x_{t})\propto p(x_{t}|z_{t})p(z_{t}|x_{1},...,x_{t-1})
p(zt∣x1,...,xt)∝p(xt∣zt)p(zt∣x1,...,xt−1)
同样服从正态分布,所以卡尔曼滤波器的每一步都可以得到解析解。
解析解计算
假设
p
(
x
)
p(x)
p(x)和
p
(
y
∣
x
)
p(y|x)
p(y∣x)都服从正态分布,且随机变量
y
y
y和
x
x
x满足带噪声的线性关系,所以有:
p
(
x
)
=
N
(
X
∣
μ
,
Λ
−
1
)
p(x)=N(X|\mu,\Lambda^{-1})
p(x)=N(X∣μ,Λ−1)
p
(
y
∣
x
)
=
N
(
y
∣
A
x
+
b
,
L
−
1
)
p(y|x)=N(y|Ax+b,L^{-1})
p(y∣x)=N(y∣Ax+b,L−1)
所以有:
p
(
y
)
=
∫
x
p
(
x
)
p
(
y
∣
x
)
d
x
=
N
(
y
∣
A
μ
+
b
,
L
−
1
+
A
Λ
−
1
A
T
)
p(y)=\int_{x}p(x)p(y|x)dx=N(y|A\mu+b,L^{-1}+A\Lambda^{-1}A^{T})
p(y)=∫xp(x)p(y∣x)dx=N(y∣Aμ+b,L−1+AΛ−1AT)
这个过程于predict对应,所以有:
p
(
x
)
⇒
p
(
z
t
−
1
∣
x
1
,
x
2
,
.
.
.
,
x
t
−
1
)
=
N
(
μ
t
−
1
,
Σ
t
−
1
)
p(x)\Rightarrow p(z_{t-1}|x_{1},x_{2},...,x_{t-1})=N(\mu_{t-1},\Sigma_{t-1})
p(x)⇒p(zt−1∣x1,x2,...,xt−1)=N(μt−1,Σt−1)
p
(
y
∣
x
)
⇒
p
(
z
t
∣
z
t
−
1
)
=
N
(
A
z
t
−
1
+
B
,
Q
)
p(y|x)\Rightarrow p(z_{t}|z_{t-1})=N(Az_{t-1}+B,Q)
p(y∣x)⇒p(zt∣zt−1)=N(Azt−1+B,Q)
p
(
y
)
⇒
p
(
z
t
∣
x
1
,
.
.
.
,
x
t
−
1
)
=
N
(
μ
t
∗
,
Σ
t
∗
)
p(y)\Rightarrow p(z_{t}|x_{1},...,x_{t-1})=N(\mu_{t}^{*},\Sigma_{t}^{*})
p(y)⇒p(zt∣x1,...,xt−1)=N(μt∗,Σt∗)
其中,参数为:
μ
t
∗
=
A
μ
t
−
1
+
B
\mu_{t}^{*}=A\mu_{t-1}+B
μt∗=Aμt−1+B
Σ
t
∗
=
Q
+
A
Σ
t
−
1
A
T
\Sigma_{t}^{*}=Q+A\Sigma_{t-1}A^{T}
Σt∗=Q+AΣt−1AT
所以predict的预测结果可以得到,由于服从正态分布,得到参数就是得到分布,即得到解析解。
另一个
p
(
x
∣
y
)
p(x|y)
p(x∣y)的分布解析解如下:
p
(
x
∣
y
)
=
N
(
x
∣
M
(
A
T
L
(
y
−
b
)
+
Λ
μ
)
,
M
)
p(x|y)=N(x|M(A^{T}L(y-b)+\Lambda\mu),M)
p(x∣y)=N(x∣M(ATL(y−b)+Λμ),M)其中有:
M
=
(
Λ
+
A
T
L
A
)
−
1
M=(\Lambda+A^{T}LA)^{-1}
M=(Λ+ATLA)−1
注意到
p
(
x
∣
y
)
∝
p
(
x
)
p
(
y
∣
x
)
p(x|y)\propto p(x)p(y|x)
p(x∣y)∝p(x)p(y∣x)与update对应:
p
(
x
)
⇒
p
(
z
t
∣
x
1
,
.
.
.
,
x
t
−
1
)
=
N
(
μ
t
∗
,
Σ
t
∗
)
p(x)\Rightarrow p(z_{t}|x_{1},...,x_{t-1})=N(\mu_{t}^{*},\Sigma_{t}^{*})
p(x)⇒p(zt∣x1,...,xt−1)=N(μt∗,Σt∗)
p
(
y
∣
x
)
⇒
p
(
x
t
∣
z
t
)
=
N
(
C
z
t
+
D
,
R
)
p(y|x)\Rightarrow p(x_{t}|z_{t})=N(Cz_{t}+D,R)
p(y∣x)⇒p(xt∣zt)=N(Czt+D,R)
p
(
x
∣
y
)
⇒
p
(
z
t
∣
x
1
,
.
.
.
,
x
t
)
=
N
(
μ
t
,
Σ
t
)
p(x|y)\Rightarrow p(z_{t}|x_{1},...,x_{t})=N(\mu_{t},\Sigma_{t})
p(x∣y)⇒p(zt∣x1,...,xt)=N(μt,Σt)
对应着参数为:
μ
t
=
Σ
t
(
C
T
R
−
1
(
x
t
−
D
)
+
(
Σ
t
∗
)
−
1
μ
t
∗
)
\mu_{t}=\Sigma_{t}(C^{T}R^{-1}(x_{t}-D)+(\Sigma_{t}^{*})^{-1}\mu_{t}^{*})
μt=Σt(CTR−1(xt−D)+(Σt∗)−1μt∗)其中
Σ
t
=
(
(
Σ
t
∗
)
−
1
+
C
T
R
−
1
C
)
−
1
\Sigma_{t}=((\Sigma_{t}^{*})^{-1}+C^{T}R^{-1}C)^{-1}
Σt=((Σt∗)−1+CTR−1C)−1
其中,
μ
t
∗
\mu_{t}^{*}
μt∗和
Σ
t
∗
\Sigma_{t}^{*}
Σt∗是
t
−
1
t-1
t−1时刻的预测概率,即
p
(
z
t
∣
x
1
,
.
.
.
,
x
t
−
1
)
=
N
(
μ
t
∗
,
Σ
t
∗
)
p(z_{t}|x_{1},...,x_{t-1})=N(\mu_{t}^{*},\Sigma_{t}^{*})
p(zt∣x1,...,xt−1)=N(μt∗,Σt∗)的参数;
所以,我们按照以上解析解,一轮一轮进行滤波,即卡尔曼滤波。对于 t t t轮的 update 步,从分布中取期望作为该时刻的滤波结果。卡尔曼滤波器的参数通常根据具体控制系统的建模确定。