【论文精读系列】之《Learning Background-Aware Correlation Filters for Visual Tracking》其三
论文地址:《Learning Background-Aware Correlation Filters for Visual Tracking》
注:该篇博客中,为了明确矩阵或者向量的形状,常出现诸如 ( A ) m , n (\boldsymbol A)_{m,n} (A)m,n 或者 ( b ) n , 1 (\boldsymbol b)_{n,1} (b)n,1 的表示,则表明 A \boldsymbol A A 是一个 m m m 行 n n n 列的矩阵,而 b \boldsymbol b b 则是 n n n 行 1 列的列向量。
4 Background-Aware Correlation Filters(背景感知相关滤波器)
首先要对先前博客的内容做一个更正:
在上一篇博客中指出,
x
\boldsymbol x
x 表示拥有
D
D
D 个像素的单帧图像,这是为了便于后续的说明,事实上,
x
\boldsymbol x
x 真正表示的是单帧图像中拥有
D
D
D 个像素的一个图像块(块中包含目标)。
![](https://i-blog.csdnimg.cn/blog_migrate/834b7fa4f805433f8dd55d3b11ec18c6.jpeg)
上图整个是一帧图像,
x
\boldsymbol x
x 表示图中蓝色框内的图像块,它包含了目标(bird),且一共有
D
D
D 个像素(多通道的话每个特征通道都有
D
D
D 个像素)。
如先前博客所述,我们需要更多的来自背景的真实样本,因此论文扩大了选取的图像块的大小。
现在
x
\boldsymbol x
x 表示图中绿色框内的图像块,它包含了目标(bird),且一共有
T
T
T 个像素(多通道的话每个特征通道都有
T
T
T 个像素)。
那么现在有了绿色框内图像块,我们还想把先前小范围精确包含目标(bird)的蓝色框内的图像块提取出来,怎么操作?
论文提出一个裁剪矩阵
(
P
)
D
,
T
(\boldsymbol P)_{D,T}
(P)D,T,其中元素不是 0 就是 1 ,它可以完成这个任务。
补充说明3(开始):
不妨举一个例子(单通道),绿色框内图像块记为
I
m
a
g
e
g
r
e
e
n
Image_{green}
Imagegreen,其为
I
m
a
g
e
g
r
e
e
n
=
[
255
255
255
255
255
255
255
255
255
255
0
255
255
255
255
255
]
Image_{green}=\begin{bmatrix} 255 & 255 & 255 & 255 \\ 255 & 255 & 255 & 255 \\ 255 & 255 & 0 & 255 \\ 255 & 255 & 255 & 255 \\ \end{bmatrix}
Imagegreen=⎣⎢⎢⎡2552552552552552552552552552550255255255255255⎦⎥⎥⎤ 其中灰度值为 0 的像素为目标(目标就一个像素),示意图如下图
![](https://i-blog.csdnimg.cn/blog_migrate/6fce2dc55d16c2106eaa0416546dc978.png)
蓝色框内的图像块记为
I
m
a
g
e
b
l
u
e
Image_{blue}
Imageblue,其为
I
m
a
g
e
b
l
u
e
=
[
255
255
255
0
]
Image_{blue}=\begin{bmatrix} 255 & 255 \\255 & 0 \end{bmatrix}
Imageblue=[2552552550] 向量化之后分别有
x
b
l
u
e
=
[
255
255
255
0
]
=
[
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
]
[
255
255
255
255
255
255
255
255
255
255
0
255
255
255
255
255
]
=
P
x
g
r
e
e
n
\boldsymbol x_{blue}=\begin{bmatrix} 255 \\ 255 \\255\\0 \end{bmatrix}= \begin{bmatrix} 0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0\\0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0\\0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0\end{bmatrix} \begin{bmatrix} 255 \\ 255 \\255\\255\\255\\255\\255\\255\\255\\255\\0\\255\\255\\255\\255\\255 \end{bmatrix}=\boldsymbol P\boldsymbol x_{green}
xblue=⎣⎢⎢⎡2552552550⎦⎥⎥⎤=⎣⎢⎢⎡0000000000000000000010000100000000000010000100000000000000000000⎦⎥⎥⎤⎣⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎡2552552552552552552552552552550255255255255255⎦⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎤=Pxgreen 补充说明3(结束)
有了上述的说明,便可以引出论文核心的公式,公式(3)如下:
E
(
h
)
=
1
2
∑
j
=
1
T
∥
y
(
j
)
−
∑
k
=
1
K
h
k
T
P
x
k
[
Δ
τ
j
]
∥
2
2
+
λ
2
∑
k
=
1
K
∥
h
k
∥
2
2
(3)
E(\boldsymbol h)=\frac{1}{2}{\sum_{j=1}^T\lVert y(j)-\sum_{k=1}^K\boldsymbol h_k^T\boldsymbol P\boldsymbol x_k[\Delta{\boldsymbol\tau}_j]\rVert_2^2}+\frac{\lambda}{2}\sum_{k=1}^K\lVert{\boldsymbol h}_k\rVert_2^2\tag{3}
E(h)=21j=1∑T∥y(j)−k=1∑KhkTPxk[Δτj]∥22+2λk=1∑K∥hk∥22(3) 相比公式(2),仅仅是每一特征通道的图像块
x
k
\boldsymbol x_k
xk 变为了
T
T
T 行 1 列,并使用裁剪矩阵
(
P
)
D
,
T
(\boldsymbol P)_{D,T}
(P)D,T 进行裁剪。对裁剪后的
(
P
x
k
[
Δ
τ
j
]
)
D
,
1
(\boldsymbol P\boldsymbol x_k[\Delta{\boldsymbol\tau}_j])_{D,1}
(Pxk[Δτj])D,1 进行滤波,同样需要滤波器
(
h
)
D
,
1
(\boldsymbol h)_{D,1}
(h)D,1 ,可以看出滤波器的尺寸没有变化,因而需要学习的参数没有增多。
需要指明的是:实际操作时,
T
≫
D
T\gg D
T≫D,一般在 4 倍左右。
为了提高计算效率,相关滤波器通常在频域中学习。下面将公式(3)进行频域表达,即对其进行离散傅里叶变换(DFT)。
补充说明4(开始):
任一离散数字信号序列
(
a
)
N
,
1
(\boldsymbol a)_{N,1}
(a)N,1 ,其离散傅里叶变换结果为
(
a
^
)
N
,
1
(\hat \boldsymbol a)_{N,1}
(a^)N,1 ,有:
(
a
^
)
N
,
1
=
N
(
F
)
N
,
N
(
a
)
N
,
1
(\hat \boldsymbol a)_{N,1}=\sqrt{N}(\boldsymbol F)_{N,N}(\boldsymbol a)_{N,1}
(a^)N,1=N(F)N,N(a)N,1 其中矩阵
F
\boldsymbol F
F 如下
F
=
1
N
[
1
1
1
⋯
1
1
W
N
1
W
N
2
⋯
W
N
N
−
1
1
W
N
2
W
N
4
⋯
W
N
2
(
N
−
1
)
⋮
⋮
⋮
⋱
⋮
1
W
N
N
−
1
W
N
2
(
N
−
1
)
⋯
W
N
(
N
−
1
)
(
N
−
1
)
]
N
,
N
\boldsymbol F=\frac{1}{N}\begin{bmatrix} 1&1&1&\cdots&1 \\1&W_N^1&W_N^2&\cdots&W_N^{N-1}\\ 1&W_N^2&W_N^4&\cdots&W_N^{2(N-1)}\\\vdots&\vdots&\vdots&\ddots&\vdots\\1 &W_N^{N-1}&W_N^{2(N-1)}&\cdots&W_N^{(N-1)(N-1)}\end{bmatrix}_{N,N}
F=N1⎣⎢⎢⎢⎢⎢⎢⎡111⋮11WN1WN2⋮WNN−11WN2WN4⋮WN2(N−1)⋯⋯⋯⋱⋯1WNN−1WN2(N−1)⋮WN(N−1)(N−1)⎦⎥⎥⎥⎥⎥⎥⎤N,N 可以看出
N
F
\sqrt{N}\boldsymbol F
NF 是复数基向量的正交矩阵(即酉矩阵),用于将任何
N
N
N 维矢量化信号映射到傅立叶域。为了便于理解,仍需做四点说明:
① 上述矩阵
F
\boldsymbol F
F 对应的傅里叶逆变换矩阵如下
F
−
1
=
[
1
1
1
⋯
1
1
W
N
−
1
W
N
−
2
⋯
W
N
−
(
N
−
1
)
1
W
N
−
2
W
N
−
4
⋯
W
N
−
2
(
N
−
1
)
⋮
⋮
⋮
⋱
⋮
1
W
N
−
(
N
−
1
)
W
N
−
2
(
N
−
1
)
⋯
W
N
−
(
N
−
1
)
(
N
−
1
)
]
N
,
N
\boldsymbol F^{-1}=\begin{bmatrix} 1&1&1&\cdots&1 \\1&W_N^{-1}&W_N^{-2}&\cdots&W_N^{-(N-1)}\\ 1&W_N^{-2}&W_N^{-4}&\cdots&W_N^{-2(N-1)}\\\vdots&\vdots&\vdots&\ddots&\vdots\\1 &W_N^{-(N-1)}&W_N^{-2(N-1)}&\cdots&W_N^{-(N-1)(N-1)}\end{bmatrix}_{N,N}
F−1=⎣⎢⎢⎢⎢⎢⎢⎡111⋮11WN−1WN−2⋮WN−(N−1)1WN−2WN−4⋮WN−2(N−1)⋯⋯⋯⋱⋯1WN−(N−1)WN−2(N−1)⋮WN−(N−1)(N−1)⎦⎥⎥⎥⎥⎥⎥⎤N,N 且有
(
N
F
)
−
1
(
N
F
)
=
(
N
F
)
(
N
F
)
−
1
=
I
(\sqrt{N}\boldsymbol F)^{-1}(\sqrt{N}\boldsymbol F)=(\sqrt{N}\boldsymbol F)(\sqrt{N}\boldsymbol F)^{-1}=\boldsymbol I
(NF)−1(NF)=(NF)(NF)−1=I
② 在很多地方可以看见
F
\boldsymbol F
F 也有别的定义方式,如
F
=
1
N
[
1
1
1
⋯
1
1
W
N
1
W
N
2
⋯
W
N
N
−
1
1
W
N
2
W
N
4
⋯
W
N
2
(
N
−
1
)
⋮
⋮
⋮
⋱
⋮
1
W
N
N
−
1
W
N
2
(
N
−
1
)
⋯
W
N
(
N
−
1
)
(
N
−
1
)
]
N
,
N
\boldsymbol F=\frac{1}{\sqrt{N}}\begin{bmatrix} 1&1&1&\cdots&1 \\1&W_N^1&W_N^2&\cdots&W_N^{N-1}\\ 1&W_N^2&W_N^4&\cdots&W_N^{2(N-1)}\\\vdots&\vdots&\vdots&\ddots&\vdots\\1 &W_N^{N-1}&W_N^{2(N-1)}&\cdots&W_N^{(N-1)(N-1)}\end{bmatrix}_{N,N}
F=N1⎣⎢⎢⎢⎢⎢⎢⎡111⋮11WN1WN2⋮WNN−11WN2WN4⋮WN2(N−1)⋯⋯⋯⋱⋯1WNN−1WN2(N−1)⋮WN(N−1)(N−1)⎦⎥⎥⎥⎥⎥⎥⎤N,N
F
−
1
=
1
N
[
1
1
1
⋯
1
1
W
N
−
1
W
N
−
2
⋯
W
N
−
(
N
−
1
)
1
W
N
−
2
W
N
−
4
⋯
W
N
−
2
(
N
−
1
)
⋮
⋮
⋮
⋱
⋮
1
W
N
−
(
N
−
1
)
W
N
−
2
(
N
−
1
)
⋯
W
N
−
(
N
−
1
)
(
N
−
1
)
]
N
,
N
\boldsymbol F^{-1}=\frac{1}{\sqrt{N}}\begin{bmatrix} 1&1&1&\cdots&1 \\1&W_N^{-1}&W_N^{-2}&\cdots&W_N^{-(N-1)}\\ 1&W_N^{-2}&W_N^{-4}&\cdots&W_N^{-2(N-1)}\\\vdots&\vdots&\vdots&\ddots&\vdots\\1 &W_N^{-(N-1)}&W_N^{-2(N-1)}&\cdots&W_N^{-(N-1)(N-1)}\end{bmatrix}_{N,N}
F−1=N1⎣⎢⎢⎢⎢⎢⎢⎡111⋮11WN−1WN−2⋮WN−(N−1)1WN−2WN−4⋮WN−2(N−1)⋯⋯⋯⋱⋯1WN−(N−1)WN−2(N−1)⋮WN−(N−1)(N−1)⎦⎥⎥⎥⎥⎥⎥⎤N,N对于这个问题,可以参照维基百科词条 DFT matrix 的解释,简而言之就是,矩阵
F
\boldsymbol F
F 和
F
−
1
\boldsymbol F^{-1}
F−1 前的归一化因子可以任意选取,只要保证
F
−
1
F
=
F
F
−
1
=
I
\boldsymbol F^{-1}\boldsymbol F=\boldsymbol F\boldsymbol F^{-1}=\boldsymbol I
F−1F=FF−1=I 即可,但是我们选取
F
=
1
N
[
1
1
1
⋯
1
1
W
N
1
W
N
2
⋯
W
N
N
−
1
1
W
N
2
W
N
4
⋯
W
N
2
(
N
−
1
)
⋮
⋮
⋮
⋱
⋮
1
W
N
N
−
1
W
N
2
(
N
−
1
)
⋯
W
N
(
N
−
1
)
(
N
−
1
)
]
N
,
N
\boldsymbol F=\frac{1}{N}\begin{bmatrix} 1&1&1&\cdots&1 \\1&W_N^1&W_N^2&\cdots&W_N^{N-1}\\ 1&W_N^2&W_N^4&\cdots&W_N^{2(N-1)}\\\vdots&\vdots&\vdots&\ddots&\vdots\\1 &W_N^{N-1}&W_N^{2(N-1)}&\cdots&W_N^{(N-1)(N-1)}\end{bmatrix}_{N,N}
F=N1⎣⎢⎢⎢⎢⎢⎢⎡111⋮11WN1WN2⋮WNN−11WN2WN4⋮WN2(N−1)⋯⋯⋯⋱⋯1WNN−1WN2(N−1)⋮WN(N−1)(N−1)⎦⎥⎥⎥⎥⎥⎥⎤N,N 时,可以保证
N
F
\sqrt{N}\boldsymbol F
NF 是酉矩阵(在论文中
N
F
\sqrt{N}\boldsymbol F
NF 才是离散傅里叶变换矩阵,而非
F
\boldsymbol F
F)。若
N
F
\sqrt{N}\boldsymbol F
NF 是酉矩阵,则有
(
N
F
)
H
(
N
F
)
=
(
N
F
)
(
N
F
)
H
=
I
(\sqrt{N}\boldsymbol F)^H(\sqrt{N}\boldsymbol F)=(\sqrt{N}\boldsymbol F)(\sqrt{N}\boldsymbol F)^H=\boldsymbol I
(NF)H(NF)=(NF)(NF)H=I,这会给傅里叶域计算带来便捷。
③ 矩阵元素中的
W
N
=
e
−
i
2
π
N
W_N=e^{-i\frac{2\pi}{N}}
WN=e−iN2π,因此当
N
=
4
N=4
N=4 时,有
W
4
=
e
−
i
2
π
4
=
−
i
W_4=e^{-i\frac{2\pi}{4}}=-i
W4=e−i42π=−i
4
F
=
1
4
[
1
1
1
1
1
−
i
−
1
i
1
−
1
1
−
1
1
i
−
1
−
i
]
4
,
4
\sqrt{4}\boldsymbol F=\frac{1}{\sqrt{4}}\begin{bmatrix} 1&1&1&1 \\1&-i&-1&i\\ 1&-1&1&-1\\1 &i&-1&-i\end{bmatrix}_{4,4}
4F=41⎣⎢⎢⎡11111−i−1i1−11−11i−1−i⎦⎥⎥⎤4,4 易验证上面是一个酉矩阵。
W
N
=
e
−
i
2
π
N
W_N=e^{-i\frac{2\pi}{N}}
WN=e−iN2π(称之为旋转因子) 还有计算性质如下
• 周期性
W
N
n
=
W
N
n
+
r
N
W_N^n=W_N^{n+rN}
WNn=WNn+rN
• 共轭对称性
W
N
n
=
(
W
N
−
n
)
∗
W_N^n=(W_N^{-n})^*
WNn=(WN−n)∗
• 可约性
W
r
N
r
n
=
W
N
n
W_{rN}^{rn}=W_N^n
WrNrn=WNn
• 正交性
1
N
∑
n
=
0
N
−
1
W
N
k
n
(
W
N
m
n
)
∗
=
1
N
∑
n
=
0
N
−
1
W
N
(
k
−
m
)
n
=
{
1
,
k
=
m
0
,
k
≠
m
\frac{1}{N}\sum_{n=0}^{N-1}W_{N}^{kn}(W_N^{mn})^*=\frac{1}{N}\sum_{n=0}^{N-1}W_{N}^{(k-m)n}=\begin{cases} 1,&k=m \\ 0,&k \neq m \end{cases}
N1∑n=0N−1WNkn(WNmn)∗=N1∑n=0N−1WN(k−m)n={1,0,k=mk=m
④
(
a
^
)
N
,
1
=
N
(
F
)
N
,
N
(
a
)
N
,
1
(\hat \boldsymbol a)_{N,1}=\sqrt{N}(\boldsymbol F)_{N,N}(\boldsymbol a)_{N,1}
(a^)N,1=N(F)N,N(a)N,1 是矩阵形式的表达,DFT 及 IDFT 常见写法如下
a
^
(
k
)
=
1
N
∑
n
=
0
N
−
1
a
(
n
)
W
N
n
k
,
k
=
0
,
1
,
⋯
,
N
−
1
\hat a(k)=\frac{1}{\sqrt{N}}\sum_{n=0}^{N-1} a(n)W_N^{nk},k=0,1,\cdots,N-1
a^(k)=N1n=0∑N−1a(n)WNnk,k=0,1,⋯,N−1
a
(
n
)
=
1
N
∑
k
=
0
N
−
1
a
^
(
k
)
W
N
−
n
k
,
n
=
0
,
1
,
⋯
,
N
−
1
a(n)=\frac{1}{\sqrt{N}}\sum_{k=0}^{N-1}\hat a(k)W_N^{-nk},n=0,1,\cdots,N-1
a(n)=N1k=0∑N−1a^(k)WN−nk,n=0,1,⋯,N−1 补充说明4(结束)
下面我们考虑将公式(3)变换到频域(傅氏域):
有了上述的补充说明,便可以进行下述推导:
重点先看公式(3)的第一项(省略系数
1
2
\frac{1}{2}
21)
∑
j
=
1
T
∥
y
(
j
)
−
∑
k
=
1
K
h
k
T
P
x
k
[
Δ
τ
j
]
∥
2
2
{\sum_{j=1}^T\lVert y(j)-\sum_{k=1}^K\boldsymbol h_k^T\boldsymbol P\boldsymbol x_k[\Delta{\boldsymbol\tau}_j]\rVert_2^2}
j=1∑T∥y(j)−k=1∑KhkTPxk[Δτj]∥22 将其展开写可写成:
[
y
(
1
)
−
(
⟨
P
T
h
1
,
x
1
[
Δ
τ
1
]
⟩
+
⟨
P
T
h
2
,
x
2
[
Δ
τ
1
]
⟩
+
⋯
+
⟨
P
T
h
K
,
x
K
[
Δ
τ
1
]
⟩
)
]
2
+
\begin{bmatrix} y(1)-\begin{pmatrix}\langle\boldsymbol P^T\boldsymbol h_1,\boldsymbol x_1[\Delta{\boldsymbol\tau}_1]\rangle+\langle\boldsymbol P^T\boldsymbol h_2,\boldsymbol x_2[\Delta{\boldsymbol\tau}_1]\rangle+\cdots+\langle\boldsymbol P^T\boldsymbol h_K,\boldsymbol x_K[\Delta{\boldsymbol\tau}_1]\rangle\end{pmatrix} \end{bmatrix}^2+
[y(1)−(⟨PTh1,x1[Δτ1]⟩+⟨PTh2,x2[Δτ1]⟩+⋯+⟨PThK,xK[Δτ1]⟩)]2+
[
y
(
2
)
−
(
⟨
P
T
h
1
,
x
1
[
Δ
τ
2
]
⟩
+
⟨
P
T
h
2
,
x
2
[
Δ
τ
2
]
⟩
+
⋯
+
⟨
P
T
h
K
,
x
K
[
Δ
τ
2
]
⟩
)
]
2
\begin{bmatrix} y(2)-\begin{pmatrix}\langle\boldsymbol P^T\boldsymbol h_1,\boldsymbol x_1[\Delta{\boldsymbol\tau}_2]\rangle+\langle\boldsymbol P^T\boldsymbol h_2,\boldsymbol x_2[\Delta{\boldsymbol\tau}_2]\rangle+\cdots+\langle\boldsymbol P^T\boldsymbol h_K,\boldsymbol x_K[\Delta{\boldsymbol\tau}_2]\rangle\end{pmatrix} \end{bmatrix}^2
[y(2)−(⟨PTh1,x1[Δτ2]⟩+⟨PTh2,x2[Δτ2]⟩+⋯+⟨PThK,xK[Δτ2]⟩)]2
+
⋯
+
+\cdots +
+⋯+
[
y
(
T
)
−
(
⟨
P
T
h
1
,
x
1
[
Δ
τ
T
]
⟩
+
⟨
P
T
h
2
,
x
2
[
Δ
τ
T
]
⟩
+
⋯
+
⟨
P
T
h
K
,
x
K
[
Δ
τ
T
]
⟩
)
]
2
\begin{bmatrix} y(T)-\begin{pmatrix}\langle\boldsymbol P^T\boldsymbol h_1,\boldsymbol x_1[\Delta{\boldsymbol\tau}_T]\rangle+\langle\boldsymbol P^T\boldsymbol h_2,\boldsymbol x_2[\Delta{\boldsymbol\tau}_T]\rangle+\cdots+\langle\boldsymbol P^T\boldsymbol h_K,\boldsymbol x_K[\Delta{\boldsymbol\tau}_T]\rangle\end{pmatrix} \end{bmatrix}^2
[y(T)−(⟨PTh1,x1[ΔτT]⟩+⟨PTh2,x2[ΔτT]⟩+⋯+⟨PThK,xK[ΔτT]⟩)]2 将上述展开式看成一个 “大” 的 2-范数的平方,即
∥
(
y
)
T
,
1
−
(
z
)
T
,
1
∥
2
2
\lVert(\boldsymbol y)_{T,1}-(\boldsymbol z)_{T,1}\rVert_2^2
∥(y)T,1−(z)T,1∥22 ,其中
z
(
j
)
=
∑
k
=
1
K
⟨
P
T
h
k
,
x
k
[
Δ
τ
j
+
1
]
⟩
,
j
=
0
,
1
,
⋯
,
T
−
1
z(j)=\sum_{k=1}^K\langle\boldsymbol P^T\boldsymbol h_k,\boldsymbol x_k[\Delta{\boldsymbol\tau}_{j+1}]\rangle,j=0,1,\cdots,T-1
z(j)=k=1∑K⟨PThk,xk[Δτj+1]⟩,j=0,1,⋯,T−1 显然有对
y
\boldsymbol y
y 的 DFT:
y
^
=
T
F
y
\hat\boldsymbol y=\sqrt{T}\boldsymbol F\boldsymbol y
y^=TFy
下面考虑对
z
\boldsymbol z
z 的 DFT,
z
^
\hat\boldsymbol z
z^ ,显然有:
z
^
=
T
F
z
=
\hat\boldsymbol z=\sqrt{T}\boldsymbol F\boldsymbol z=
z^=TFz=
1
T
[
1
1
1
⋯
1
1
W
T
1
W
T
2
⋯
W
T
T
−
1
1
W
T
2
W
T
4
⋯
W
T
2
(
T
−
1
)
⋮
⋮
⋮
⋱
⋮
1
W
T
T
−
1
W
T
2
(
T
−
1
)
⋯
W
T
(
T
−
1
)
(
T
−
1
)
]
T
,
T
[
⟨
P
T
h
1
,
x
1
[
Δ
τ
1
]
⟩
+
⟨
P
T
h
2
,
x
2
[
Δ
τ
1
]
⟩
+
⋯
+
⟨
P
T
h
K
,
x
K
[
Δ
τ
1
]
⟩
⟨
P
T
h
1
,
x
1
[
Δ
τ
2
]
⟩
+
⟨
P
T
h
2
,
x
2
[
Δ
τ
2
]
⟩
+
⋯
+
⟨
P
T
h
K
,
x
K
[
Δ
τ
2
]
⟩
⋮
⟨
P
T
h
1
,
x
1
[
Δ
τ
T
]
⟩
+
⟨
P
T
h
2
,
x
2
[
Δ
τ
T
]
⟩
+
⋯
+
⟨
P
T
h
K
,
x
K
[
Δ
τ
T
]
⟩
]
T
,
1
\frac{1}{\sqrt{T}}\begin{bmatrix} 1&1&1&\cdots&1 \\1&W_T^1&W_T^2&\cdots&W_T^{T-1}\\ 1&W_T^2&W_T^4&\cdots&W_T^{2(T-1)}\\\vdots&\vdots&\vdots&\ddots&\vdots\\1 &W_T^{T-1}&W_T^{2(T-1)}&\cdots&W_T^{(T-1)(T-1)}\end{bmatrix}_{T,T}\begin{bmatrix}\langle\boldsymbol P^T\boldsymbol h_1,\boldsymbol x_1[\Delta{\boldsymbol\tau}_1]\rangle+\langle\boldsymbol P^T\boldsymbol h_2,\boldsymbol x_2[\Delta{\boldsymbol\tau}_1]\rangle+\cdots+\langle\boldsymbol P^T\boldsymbol h_K,\boldsymbol x_K[\Delta{\boldsymbol\tau}_1]\rangle\\\langle\boldsymbol P^T\boldsymbol h_1,\boldsymbol x_1[\Delta{\boldsymbol\tau}_2]\rangle+\langle\boldsymbol P^T\boldsymbol h_2,\boldsymbol x_2[\Delta{\boldsymbol\tau}_2]\rangle+\cdots+\langle\boldsymbol P^T\boldsymbol h_K,\boldsymbol x_K[\Delta{\boldsymbol\tau}_2]\rangle\\\vdots\\\langle\boldsymbol P^T\boldsymbol h_1,\boldsymbol x_1[\Delta{\boldsymbol\tau}_T]\rangle+\langle\boldsymbol P^T\boldsymbol h_2,\boldsymbol x_2[\Delta{\boldsymbol\tau}_T]\rangle+\cdots+\langle\boldsymbol P^T\boldsymbol h_K,\boldsymbol x_K[\Delta{\boldsymbol\tau}_T]\rangle\end{bmatrix}_{T,1}
T1⎣⎢⎢⎢⎢⎢⎢⎡111⋮11WT1WT2⋮WTT−11WT2WT4⋮WT2(T−1)⋯⋯⋯⋱⋯1WTT−1WT2(T−1)⋮WT(T−1)(T−1)⎦⎥⎥⎥⎥⎥⎥⎤T,T⎣⎢⎢⎢⎡⟨PTh1,x1[Δτ1]⟩+⟨PTh2,x2[Δτ1]⟩+⋯+⟨PThK,xK[Δτ1]⟩⟨PTh1,x1[Δτ2]⟩+⟨PTh2,x2[Δτ2]⟩+⋯+⟨PThK,xK[Δτ2]⟩⋮⟨PTh1,x1[ΔτT]⟩+⟨PTh2,x2[ΔτT]⟩+⋯+⟨PThK,xK[ΔτT]⟩⎦⎥⎥⎥⎤T,1 写成另外一种形式为:
z
^
(
i
)
=
1
T
∑
j
=
0
T
−
1
z
(
j
)
W
N
j
i
=
1
T
∑
j
=
0
T
−
1
∑
k
=
1
K
⟨
P
T
h
k
,
x
k
[
Δ
τ
j
+
1
]
⟩
W
N
j
i
,
i
=
0
,
1
,
⋯
,
T
−
1
\hat z(i)=\frac{1}{\sqrt{T}}\sum_{j=0}^{T-1} z(j)W_N^{ji}=\frac{1}{\sqrt{T}}\sum_{j=0}^{T-1}\sum_{k=1}^K\langle\boldsymbol P^T\boldsymbol h_k,\boldsymbol x_k[\Delta{\boldsymbol\tau}_{j+1}]\rangle W_N^{ji},i=0,1,\cdots,T-1
z^(i)=T1j=0∑T−1z(j)WNji=T1j=0∑T−1k=1∑K⟨PThk,xk[Δτj+1]⟩WNji,i=0,1,⋯,T−1 利用公式
λ
⟨
a
,
b
⟩
=
⟨
a
,
λ
b
⟩
\lambda\langle \boldsymbol a,\boldsymbol b\rangle=\langle \boldsymbol a,\lambda \boldsymbol b\rangle
λ⟨a,b⟩=⟨a,λb⟩ 则有:
z
^
(
i
)
=
1
T
∑
j
=
0
T
−
1
∑
k
=
1
K
⟨
P
T
h
k
,
x
k
[
Δ
τ
j
+
1
]
W
N
j
i
⟩
\hat z(i)=\frac{1}{\sqrt{T}}\sum_{j=0}^{T-1}\sum_{k=1}^K\langle\boldsymbol P^T\boldsymbol h_k,\boldsymbol x_k[\Delta{\boldsymbol\tau}_{j+1}]W_N^{ji}\rangle
z^(i)=T1j=0∑T−1k=1∑K⟨PThk,xk[Δτj+1]WNji⟩ 交换求和顺序(互换符号
∑
\sum
∑):
z
^
(
i
)
=
1
T
∑
k
=
1
K
∑
j
=
0
T
−
1
⟨
P
T
h
k
,
x
k
[
Δ
τ
j
+
1
]
W
N
j
i
⟩
\hat z(i)=\frac{1}{\sqrt{T}}\sum_{k=1}^K\sum_{j=0}^{T-1}\langle\boldsymbol P^T\boldsymbol h_k,\boldsymbol x_k[\Delta{\boldsymbol\tau}_{j+1}]W_N^{ji}\rangle
z^(i)=T1k=1∑Kj=0∑T−1⟨PThk,xk[Δτj+1]WNji⟩ 利用公式
∑
⟨
a
,
b
j
⟩
=
⟨
a
,
∑
b
j
⟩
\sum\langle \boldsymbol a,\boldsymbol b_j\rangle=\langle \boldsymbol a,\sum \boldsymbol b_j\rangle
∑⟨a,bj⟩=⟨a,∑bj⟩ 则有:
z
^
(
i
)
=
1
T
∑
k
=
1
K
⟨
P
T
h
k
,
∑
j
=
0
T
−
1
x
k
[
Δ
τ
j
+
1
]
W
N
j
i
⟩
=
∑
k
=
1
K
⟨
P
T
h
k
,
1
T
∑
j
=
0
T
−
1
x
k
[
Δ
τ
j
+
1
]
W
N
j
i
⟩
\hat z(i)=\frac{1}{\sqrt{T}}\sum_{k=1}^K\langle\boldsymbol P^T\boldsymbol h_k,\sum_{j=0}^{T-1}\boldsymbol x_k[\Delta{\boldsymbol\tau}_{j+1}]W_N^{ji}\rangle=\sum_{k=1}^K\langle\boldsymbol P^T\boldsymbol h_k,\frac{1}{\sqrt{T}}\sum_{j=0}^{T-1}\boldsymbol x_k[\Delta{\boldsymbol\tau}_{j+1}]W_N^{ji}\rangle
z^(i)=T1k=1∑K⟨PThk,j=0∑T−1xk[Δτj+1]WNji⟩=k=1∑K⟨PThk,T1j=0∑T−1xk[Δτj+1]WNji⟩ 上式中
(
x
k
[
Δ
τ
j
+
1
]
)
T
,
1
(\boldsymbol x_k[\Delta{\boldsymbol\tau}_{j+1}])_{T,1}
(xk[Δτj+1])T,1 为,写开有如下:
z
^
(
i
)
=
∑
k
=
1
K
⟨
P
T
h
k
,
1
T
[
x
k
[
Δ
τ
1
]
x
k
[
Δ
τ
2
]
⋯
x
k
[
Δ
τ
T
]
]
[
W
T
0
i
W
T
1
i
⋮
W
T
(
T
−
1
)
i
]
⟩
\hat z(i)=\sum_{k=1}^K\langle\boldsymbol P^T\boldsymbol h_k,\frac{1}{\sqrt{T}}\begin{bmatrix}\boldsymbol x_k[\Delta{\boldsymbol\tau}_1] &\boldsymbol x_k[\Delta{\boldsymbol\tau}_2] &\cdots &\boldsymbol x_k[\Delta{\boldsymbol\tau}_{T}] \end{bmatrix}\begin{bmatrix} W_T^{0i}\\W_T^{1i}\\\vdots\\W_T^{(T-1)i}\end{bmatrix}\rangle
z^(i)=k=1∑K⟨PThk,T1[xk[Δτ1]xk[Δτ2]⋯xk[ΔτT]]⎣⎢⎢⎢⎡WT0iWT1i⋮WT(T−1)i⎦⎥⎥⎥⎤⟩
=
∑
k
=
1
K
(
h
k
T
P
1
T
[
x
k
[
Δ
τ
1
]
x
k
[
Δ
τ
2
]
⋯
x
k
[
Δ
τ
T
]
]
[
W
T
0
i
W
T
1
i
⋮
W
T
(
T
−
1
)
i
]
)
=\sum_{k=1}^K(\boldsymbol h_k^T\boldsymbol P\frac{1}{\sqrt{T}}\begin{bmatrix}\boldsymbol x_k[\Delta{\boldsymbol\tau}_1] &\boldsymbol x_k[\Delta{\boldsymbol\tau}_2] &\cdots &\boldsymbol x_k[\Delta{\boldsymbol\tau}_{T}] \end{bmatrix}\begin{bmatrix} W_T^{0i}\\W_T^{1i}\\\vdots\\W_T^{(T-1)i}\end{bmatrix})
=k=1∑K(hkTPT1[xk[Δτ1]xk[Δτ2]⋯xk[ΔτT]]⎣⎢⎢⎢⎡WT0iWT1i⋮WT(T−1)i⎦⎥⎥⎥⎤)
=
∑
k
=
1
K
⟨
[
x
k
[
Δ
τ
1
]
T
x
k
[
Δ
τ
2
]
T
⋮
x
k
[
Δ
τ
T
]
T
]
P
T
h
k
,
1
T
[
W
T
0
i
W
T
1
i
⋮
W
T
(
T
−
1
)
i
]
⟩
=
⟨
∑
k
=
1
K
[
x
k
[
Δ
τ
1
]
T
x
k
[
Δ
τ
2
]
T
⋮
x
k
[
Δ
τ
T
]
T
]
P
T
h
k
,
1
T
[
W
T
0
i
W
T
1
i
⋮
W
T
(
T
−
1
)
i
]
⟩
=\sum_{k=1}^K\langle{\begin{bmatrix}\boldsymbol x_k[\Delta{\boldsymbol\tau}_1]^T\\\boldsymbol x_k[\Delta{\boldsymbol\tau}_2]^T\\\vdots\\\boldsymbol x_k[\Delta{\boldsymbol\tau}_T]^T\end{bmatrix}}\boldsymbol P^T\boldsymbol h_k,\frac{1}{\sqrt{T}}\begin{bmatrix} W_T^{0i}\\W_T^{1i}\\\vdots\\W_T^{(T-1)i}\end{bmatrix}\rangle=\langle\sum_{k=1}^K{\begin{bmatrix}\boldsymbol x_k[\Delta{\boldsymbol\tau}_1]^T\\\boldsymbol x_k[\Delta{\boldsymbol\tau}_2]^T\\\vdots\\\boldsymbol x_k[\Delta{\boldsymbol\tau}_T]^T\end{bmatrix}}\boldsymbol P^T\boldsymbol h_k,\frac{1}{\sqrt{T}}\begin{bmatrix} W_T^{0i}\\W_T^{1i}\\\vdots\\W_T^{(T-1)i}\end{bmatrix}\rangle
=k=1∑K⟨⎣⎢⎢⎢⎡xk[Δτ1]Txk[Δτ2]T⋮xk[ΔτT]T⎦⎥⎥⎥⎤PThk,T1⎣⎢⎢⎢⎡WT0iWT1i⋮WT(T−1)i⎦⎥⎥⎥⎤⟩=⟨k=1∑K⎣⎢⎢⎢⎡xk[Δτ1]Txk[Δτ2]T⋮xk[ΔτT]T⎦⎥⎥⎥⎤PThk,T1⎣⎢⎢⎢⎡WT0iWT1i⋮WT(T−1)i⎦⎥⎥⎥⎤⟩
=
(
∑
k
=
1
K
h
k
T
P
[
x
k
[
Δ
τ
1
]
x
k
[
Δ
τ
2
]
⋯
x
k
[
Δ
τ
T
]
]
)
1
,
T
(
1
T
[
W
T
0
i
W
T
1
i
⋮
W
T
(
T
−
1
)
i
]
)
T
,
1
=\left(\sum_{k=1}^K\boldsymbol h_k^T\boldsymbol P\begin{bmatrix}\boldsymbol x_k[\Delta{\boldsymbol\tau}_1] &\boldsymbol x_k[\Delta{\boldsymbol\tau}_2] &\cdots&\boldsymbol x_k[\Delta{\boldsymbol\tau}_{T}] \end{bmatrix}\right)_{1,T}\left(\frac{1}{\sqrt{T}}\begin{bmatrix} W_T^{0i}\\W_T^{1i}\\\vdots\\W_T^{(T-1)i}\end{bmatrix}\right)_{T,1}
=(k=1∑KhkTP[xk[Δτ1]xk[Δτ2]⋯xk[ΔτT]])1,T⎝⎜⎜⎜⎛T1⎣⎢⎢⎢⎡WT0iWT1i⋮WT(T−1)i⎦⎥⎥⎥⎤⎠⎟⎟⎟⎞T,1
因此
z
^
T
=
[
z
^
(
0
)
,
z
^
(
1
)
,
⋯
,
z
^
(
T
−
1
)
]
\hat\boldsymbol z^T=[\hat z(0),\hat z(1),\cdots,\hat z(T-1)]
z^T=[z^(0),z^(1),⋯,z^(T−1)] 为:
z
^
T
=
(
∑
k
=
1
K
h
k
T
P
[
x
k
[
Δ
τ
1
]
x
k
[
Δ
τ
2
]
⋯
x
k
[
Δ
τ
T
]
]
)
1
,
T
(
1
T
[
W
T
0
W
T
0
⋯
W
T
0
W
T
0
W
T
1
⋯
W
T
T
−
1
⋮
⋮
⋱
⋮
W
T
0
W
T
1
(
T
−
1
)
⋯
W
T
(
T
−
1
)
(
T
−
1
)
]
)
T
,
T
\hat\boldsymbol z^T=\left(\sum_{k=1}^K\boldsymbol h_k^T\boldsymbol P\begin{bmatrix}\boldsymbol x_k[\Delta{\boldsymbol\tau}_1] &\boldsymbol x_k[\Delta{\boldsymbol\tau}_2] &\cdots&\boldsymbol x_k[\Delta{\boldsymbol\tau}_{T}] \end{bmatrix}\right)_{1,T}\left(\frac{1}{\sqrt{T}} \begin{bmatrix} W_T^{0}&W_T^{0}&\cdots&W_T^{0}\\W_T^{0}&W_T^{1}&\cdots&W_T^{T-1}\\\vdots&\vdots&\ddots&\vdots\\W_T^{0}&W_T^{1(T-1)}&\cdots&W_T^{(T-1)(T-1)}\end{bmatrix}\right)_{T,T}
z^T=(k=1∑KhkTP[xk[Δτ1]xk[Δτ2]⋯xk[ΔτT]])1,T⎝⎜⎜⎜⎛T1⎣⎢⎢⎢⎡WT0WT0⋮WT0WT0WT1⋮WT1(T−1)⋯⋯⋱⋯WT0WTT−1⋮WT(T−1)(T−1)⎦⎥⎥⎥⎤⎠⎟⎟⎟⎞T,T
=
(
∑
k
=
1
K
h
k
T
P
[
x
k
[
Δ
τ
1
]
x
k
[
Δ
τ
2
]
⋯
x
k
[
Δ
τ
T
]
]
)
1
,
T
(
T
F
)
T
,
T
=\left(\sum_{k=1}^K\boldsymbol h_k^T\boldsymbol P\begin{bmatrix}\boldsymbol x_k[\Delta{\boldsymbol\tau}_1] &\boldsymbol x_k[\Delta{\boldsymbol\tau}_2] &\cdots&\boldsymbol x_k[\Delta{\boldsymbol\tau}_{T}] \end{bmatrix}\right)_{1,T}\left(\sqrt{T} \boldsymbol F\right)_{T,T}
=(k=1∑KhkTP[xk[Δτ1]xk[Δτ2]⋯xk[ΔτT]])1,T(TF)T,T
记矩阵如下:
X
k
=
[
x
k
[
Δ
τ
1
]
T
x
k
[
Δ
τ
2
]
T
⋮
x
k
[
Δ
τ
T
]
T
]
=
[
x
k
(
1
)
x
k
(
2
)
⋯
x
k
(
T
)
x
k
(
T
)
x
k
(
1
)
⋯
x
k
(
T
−
1
)
⋮
⋮
⋱
⋮
x
k
(
2
)
x
k
(
3
)
⋯
x
k
(
1
)
]
\boldsymbol X_k={\begin{bmatrix}\boldsymbol x_k[\Delta{\boldsymbol\tau}_1]^T\\\boldsymbol x_k[\Delta{\boldsymbol\tau}_2]^T\\\vdots\\\boldsymbol x_k[\Delta{\boldsymbol\tau}_T]^T\end{bmatrix}}={\begin{bmatrix}x_k(1)&x_k(2)&\cdots&x_k(T)\\x_k(T)&x_k(1)&\cdots&x_k(T-1)\\\vdots&\vdots&\ddots&\vdots\\x_k(2)&x_k(3)&\cdots&x_k(1)\end{bmatrix}}
Xk=⎣⎢⎢⎢⎡xk[Δτ1]Txk[Δτ2]T⋮xk[ΔτT]T⎦⎥⎥⎥⎤=⎣⎢⎢⎢⎡xk(1)xk(T)⋮xk(2)xk(2)xk(1)⋮xk(3)⋯⋯⋱⋯xk(T)xk(T−1)⋮xk(1)⎦⎥⎥⎥⎤ 很明显每一个
X
k
(
k
=
1
,
2
,
⋯
,
K
)
\boldsymbol X_k(k=1,2,\cdots,K)
Xk(k=1,2,⋯,K) 是一个循环矩阵,我们可以很方便地对循环矩阵进行 DFT 。
这样,可以写出:
z
^
T
=
∑
k
=
1
K
h
k
T
P
X
k
T
T
F
\hat\boldsymbol z^T=\sum_{k=1}^K\boldsymbol h_k^T\boldsymbol P\boldsymbol X_k^T\sqrt{T}\boldsymbol F
z^T=k=1∑KhkTPXkTTF 写成列向量的形式,需要进行转置(
F
=
F
T
\boldsymbol F=\boldsymbol F^T
F=FT):
z
^
=
∑
k
=
1
K
F
X
k
T
P
T
h
k
\hat\boldsymbol z=\sum_{k=1}^K\boldsymbol F\boldsymbol X_k\sqrt{T}\boldsymbol P^T\boldsymbol h_k
z^=k=1∑KFXkTPThk
补充说明5(开始):
循环矩阵的重要性质:
任一循环矩阵均可以被离散傅里叶变换矩阵对角化
证明略,详见维基百科词条 Circulant matrix
及由博客【目标跟踪: 相关滤波器 三】循环矩阵 可知:
若使用论文定义的
F
\boldsymbol F
F,则有:
(
X
^
k
)
T
,
T
=
(
T
F
)
T
,
T
⋅
(
d
i
a
g
(
x
k
)
)
T
,
T
⋅
(
T
F
H
)
T
,
T
(\hat\boldsymbol X_k)_{T,T}=(\sqrt{T}\boldsymbol F)_{T,T}\cdot (diag({\boldsymbol x}_k))_{T,T}\cdot(\sqrt{T}\boldsymbol F^H)_{T,T}
(X^k)T,T=(TF)T,T⋅(diag(xk))T,T⋅(TFH)T,T 又
(
T
F
)
H
(
T
F
)
=
(
T
F
)
(
T
F
)
H
=
I
(\sqrt{T}\boldsymbol F)^H(\sqrt{T}\boldsymbol F)=(\sqrt{T}\boldsymbol F)(\sqrt{T}\boldsymbol F)^H=\boldsymbol I
(TF)H(TF)=(TF)(TF)H=I 且由于
(
T
F
)
T
=
(
T
F
)
(\sqrt{T}\boldsymbol F)^T=(\sqrt{T}\boldsymbol F)
(TF)T=(TF) ,故可得:
(
d
i
a
g
(
x
k
)
)
T
,
T
=
(
T
F
)
T
,
T
⋅
(
X
^
k
)
T
,
T
⋅
(
T
F
)
T
,
T
H
(diag({\boldsymbol x}_k))_{T,T}=(\sqrt{T}\boldsymbol F)_{T,T}\cdot (\hat\boldsymbol X_k)_{T,T}\cdot(\sqrt{T}\boldsymbol F)_{T,T}^H
(diag(xk))T,T=(TF)T,T⋅(X^k)T,T⋅(TF)T,TH 补充说明5(结束)
利用上述说明,有:
z
^
=
∑
k
=
1
K
F
X
k
F
H
F
T
P
T
h
k
\hat\boldsymbol z=\sum_{k=1}^K\boldsymbol F\boldsymbol X_k\boldsymbol F^H\boldsymbol F\sqrt{T}\boldsymbol P^T\boldsymbol h_k
z^=k=1∑KFXkFHFTPThk 记
X
k
\boldsymbol X_k
Xk 的生成向量的傅里叶变换为
x
^
k
\hat\boldsymbol x_k
x^k,则:
z
^
=
∑
k
=
1
K
d
i
a
g
(
x
^
k
)
T
(
F
P
T
)
h
k
(*)
\hat\boldsymbol z=\sum_{k=1}^K diag(\hat\boldsymbol x_k)\sqrt{T}(\boldsymbol F\boldsymbol P^T)\boldsymbol h_k\tag{*}
z^=k=1∑Kdiag(x^k)T(FPT)hk(*) 不妨记
g
k
\boldsymbol g_k
gk ,有
(
g
^
k
)
T
,
1
=
T
(
F
)
T
,
T
(
P
T
h
k
)
T
,
1
=
T
(
F
)
T
,
T
(
P
T
)
T
,
D
(
h
k
)
D
,
1
(\hat\boldsymbol g_k)_{T,1}=\sqrt{T}(\boldsymbol F)_{T,T}(\boldsymbol P^T{\boldsymbol h}_k )_{T,1}=\sqrt{T}(\boldsymbol F)_{T,T}(\boldsymbol P^T)_{T,D}({\boldsymbol h}_k )_{D,1}
(g^k)T,1=T(F)T,T(PThk)T,1=T(F)T,T(PT)T,D(hk)D,1 再进行如下设计:
将滤波器的
k
k
k 个分量并成一个特别长的列向量
(
h
)
(
D
×
K
)
,
1
(\boldsymbol h)_{(D\times K),1}
(h)(D×K),1
h
=
[
h
1
h
2
⋮
h
K
]
=
[
[
h
1
(
1
)
h
1
(
2
)
⋯
h
1
(
D
)
]
T
[
h
2
(
1
)
h
2
(
2
)
⋯
h
2
(
D
)
]
T
⋮
[
h
K
(
1
)
h
K
(
2
)
⋯
h
K
(
D
)
]
T
]
\boldsymbol h=\begin{bmatrix} \boldsymbol h_1\\\boldsymbol h_2\\ \vdots\\\boldsymbol h_K\end{bmatrix}=\begin{bmatrix} {\begin{bmatrix}h_1(1)&h_1(2)&\cdots&h_1(D)\end{bmatrix}}^T\\{\begin{bmatrix}h_2(1)&h_2(2)&\cdots&h_2(D)\end{bmatrix}}^T\\ \vdots\\{\begin{bmatrix}h_K(1)&h_K(2)&\cdots&h_K(D)\end{bmatrix}}^T\end{bmatrix}
h=⎣⎢⎢⎢⎡h1h2⋮hK⎦⎥⎥⎥⎤=⎣⎢⎢⎢⎢⎡[h1(1)h1(2)⋯h1(D)]T[h2(1)h2(2)⋯h2(D)]T⋮[hK(1)hK(2)⋯hK(D)]T⎦⎥⎥⎥⎥⎤ 现在的
h
\boldsymbol h
h 是
D
×
K
D\times K
D×K 行 1 列的列向量,裁剪矩阵
(
P
)
D
,
T
(\boldsymbol P)_{D,T}
(P)D,T 无法直接与
(
h
)
(
D
×
K
)
,
1
(\boldsymbol h)_{(D\times K),1}
(h)(D×K),1 进行计算(尺寸不匹配),因此我们需要引入克罗内克积(Kronecker Product),记其运算符为
⊗
\otimes
⊗ 。
补充说明6(开始):
矩阵
(
A
)
m
,
n
(\boldsymbol A)_{m,n}
(A)m,n 和
(
B
)
p
,
q
(\boldsymbol B)_{p,q}
(B)p,q 进行克罗内克积运算,则有
(
A
⊗
B
)
(
m
×
p
)
,
(
n
×
q
)
(\boldsymbol A\otimes\boldsymbol B)_{(m\times p),(n\times q)}
(A⊗B)(m×p),(n×q) ,具体如下
A
⊗
B
=
[
a
11
B
a
12
B
⋯
a
1
n
B
a
21
B
a
22
B
⋯
a
2
n
B
⋮
⋮
⋱
⋮
a
m
1
B
a
m
2
B
⋯
a
m
n
B
]
\boldsymbol A\otimes\boldsymbol B={\begin{bmatrix}a_{11}\boldsymbol B&a_{12}\boldsymbol B&\cdots&a_{1n}\boldsymbol B\\ a_{21}\boldsymbol B&a_{22}\boldsymbol B&\cdots&a_{2n}\boldsymbol B \\ \vdots&\vdots&\ddots&\vdots\\a_{m1}\boldsymbol B&a_{m2}\boldsymbol B&\cdots&a_{mn}\boldsymbol B \end{bmatrix}}
A⊗B=⎣⎢⎢⎢⎡a11Ba21B⋮am1Ba12Ba22B⋮am2B⋯⋯⋱⋯a1nBa2nB⋮amnB⎦⎥⎥⎥⎤ 例如
[
a
11
a
12
a
21
a
22
a
31
a
32
]
⊗
[
b
11
b
12
b
13
b
21
b
22
b
23
]
=
[
a
11
b
11
a
11
b
12
a
11
b
13
a
12
b
11
a
12
b
12
a
12
b
13
a
11
b
21
a
11
b
22
a
11
b
23
a
12
b
21
a
12
b
22
a
12
b
23
a
21
b
11
a
21
b
12
a
21
b
13
a
22
b
11
a
22
b
12
a
22
b
13
a
21
b
21
a
21
b
22
a
21
b
23
a
22
b
21
a
22
b
22
a
22
b
23
a
31
b
11
a
31
b
12
a
31
b
13
a
32
b
11
a
32
b
12
a
32
b
13
a
31
b
21
a
31
b
22
a
31
b
23
a
32
b
21
a
32
b
22
a
32
b
23
]
\begin{bmatrix}a_{11}&a_{12}\\ a_{21}&a_{22}\\a_{31}&a_{32}\end{bmatrix}\otimes\begin{bmatrix}b_{11}&b_{12}&b_{13}\\ b_{21}&b_{22}&b_{23}\end{bmatrix}= \begin{bmatrix}a_{11}b_{11}&a_{11}b_{12}&a_{11}b_{13}&a_{12}b_{11}&a_{12}b_{12}&a_{12}b_{13}\\ a_{11}b_{21}&a_{11}b_{22}&a_{11}b_{23}&a_{12}b_{21}&a_{12}b_{22}&a_{12}b_{23}\\ a_{21}b_{11}&a_{21}b_{12}&a_{21}b_{13}&a_{22}b_{11}&a_{22}b_{12}&a_{22}b_{13}\\ a_{21}b_{21}&a_{21}b_{22}&a_{21}b_{23}&a_{22}b_{21}&a_{22}b_{22}&a_{22}b_{23}\\ a_{31}b_{11}&a_{31}b_{12}&a_{31}b_{13}&a_{32}b_{11}&a_{32}b_{12}&a_{32}b_{13}\\ a_{31}b_{21}&a_{31}b_{22}&a_{31}b_{23}&a_{32}b_{21}&a_{32}b_{22}&a_{32}b_{23}\end{bmatrix}
⎣⎡a11a21a31a12a22a32⎦⎤⊗[b11b21b12b22b13b23]=⎣⎢⎢⎢⎢⎢⎢⎡a11b11a11b21a21b11a21b21a31b11a31b21a11b12a11b22a21b12a21b22a31b12a31b22a11b13a11b23a21b13a21b23a31b13a31b23a12b11a12b21a22b11a22b21a32b11a32b21a12b12a12b22a22b12a22b22a32b12a32b22a12b13a12b23a22b13a22b23a32b13a32b23⎦⎥⎥⎥⎥⎥⎥⎤ 补充说明6(结束)
由上述说明,可知:若定义单位阵
(
I
K
)
K
,
K
(\boldsymbol I_K)_{K,K}
(IK)K,K,则有
(
(
F
)
T
,
T
(
P
T
)
T
,
D
)
T
,
D
⊗
(
I
K
)
K
,
K
=
(
F
P
T
⊗
I
K
)
(
T
×
K
)
,
(
D
×
K
)
((\boldsymbol F)_{T,T}(\boldsymbol P^T)_{T,D})_{T,D}\otimes (\boldsymbol I_K)_{K,K}=(\boldsymbol F\boldsymbol P^T\otimes\boldsymbol I_K)_{(T\times K),(D\times K)}
((F)T,T(PT)T,D)T,D⊗(IK)K,K=(FPT⊗IK)(T×K),(D×K) 这样便可以与设计的特别长的列向量
(
h
)
(
D
×
K
)
,
1
(\boldsymbol h)_{(D\times K),1}
(h)(D×K),1 进行运算,如下:
(
g
^
)
(
T
×
K
)
,
1
=
T
(
F
P
T
⊗
I
K
)
(
T
×
K
)
,
(
D
×
K
)
(
h
)
(
D
×
K
)
,
1
(\hat\boldsymbol g)_{(T\times K),1}=\sqrt{T}(\boldsymbol F\boldsymbol P^T\otimes\boldsymbol I_K)_{(T\times K),(D\times K)}(\boldsymbol h)_{(D\times K),1}
(g^)(T×K),1=T(FPT⊗IK)(T×K),(D×K)(h)(D×K),1 若记
X
^
\hat\boldsymbol X
X^ 为:
X
^
=
[
d
i
a
g
(
x
^
1
)
T
,
d
i
a
g
(
x
^
2
)
T
,
⋯
,
d
i
a
g
(
x
^
K
)
T
]
\hat\boldsymbol X=\begin{bmatrix}diag(\hat\boldsymbol x_1)^T,diag(\hat\boldsymbol x_2)^T,\cdots,diag(\hat\boldsymbol x_K)^T\end{bmatrix}
X^=[diag(x^1)T,diag(x^2)T,⋯,diag(x^K)T] 这样
(
∗
)
(*)
(∗) 式可以写成:
X
^
T
(
F
P
T
⊗
I
K
)
h
\hat\boldsymbol X\sqrt{T}(\boldsymbol F\boldsymbol P^T\otimes\boldsymbol I_K)\boldsymbol h
X^T(FPT⊗IK)h 引入等式表达(用于后续 ADMM 求解):
g
^
=
T
(
F
P
T
⊗
I
K
)
h
\hat\boldsymbol g=\sqrt{T}(\boldsymbol F\boldsymbol P^T\otimes\boldsymbol I_K)\boldsymbol h
g^=T(FPT⊗IK)h 最后得到频域表达式,公式(4):
E
(
h
,
g
^
)
=
1
2
∥
y
^
−
X
^
g
^
∥
2
2
+
λ
2
∥
h
∥
2
2
(3)
E(\boldsymbol h,\hat\boldsymbol g)=\frac{1}{2}\lVert\hat\boldsymbol y-\hat\boldsymbol X\hat\boldsymbol g\rVert_2^2+\frac{\lambda}{2}\lVert\boldsymbol h\rVert_2^2\tag{3}
E(h,g^)=21∥y^−X^g^∥22+2λ∥h∥22(3)
参考链接:
辨析matmul product(一般矩阵乘积),hadamard product(哈达玛积)、kronecker product(克罗内克积)
【目标跟踪: 相关滤波器 三】循环矩阵