剔除低对比度特征点中的矩阵求导公式推导

剔除低对比度特征点中的矩阵求导公式推导

假定
$$\mathbf{x}=[x_1,x_2{]}^T,\Delta \mathbf{x}=[\Delta x_1,\Delta x_2{]}^T$$
泰勒展开式：
$$D(\mathbf{x}+\Delta \mathbf{x})=D(\mathbf{x})+(\frac{\partial D}{\partial \mathbf{x}}{)}^T\Delta \mathbf{x}+\frac{1}{2}\Delta {\mathbf{x}}^T\frac{{\partial }^{2}D}{\partial \mathbf{x}\partial {\mathbf{x}}^T}\Delta \mathbf{x}$$
其中

$$\begin{gathered} \frac{\partial D}{\partial \mathbf{x}}=[\frac{\partial D}{\partial x_1},\frac{\partial D}{\partial x_2}{]}^T \\ \frac{{\partial }^{2}D}{\partial \mathbf{x}\partial {\mathbf{x}}^T}=\left[\begin{array}{cc}\frac{{\partial }^{2}D}{\partial x_1^2}& \frac{{\partial }^{2}D}{\partial x_1\partial x_2}\\ \frac{{\partial }^{2}D}{\partial x_2\partial x_1}& \frac{{\partial }^{2}D}{\partial x_2^2}\end{array}\right]=\left[\begin{array}{cc}\frac{{\partial }^{2}D}{\partial x_1^2}& \frac{{\partial }^{2}D}{\partial x_1\partial x_2}\\ \frac{{\partial }^{2}D}{\partial x_1\partial x_2}& \frac{{\partial }^{2}D}{\partial x_2^2}\end{array}\right] \end{gathered}$$

对$\Delta \mathbf{x}$求导

$$\frac{\partial D(\mathbf{x}+\Delta \mathbf{x})}{\partial \Delta \mathbf{x}}=\frac{\partial D(\mathbf{x})}{\partial \Delta \mathbf{x}}+\frac{\partial }{\partial \Delta \mathbf{x}}[(\frac{\partial D}{\partial \mathbf{x}}{)}^T\Delta \mathbf{x}]+\frac{\partial }{\partial \Delta \mathbf{x}}[\frac{1}{2}\Delta {\mathbf{x}}^T\frac{{\partial }^{2}D}{\partial \mathbf{x}\partial {\mathbf{x}}^T}\Delta \mathbf{x}]$$

其中

$$\begin{gathered} \frac{\partial D(\mathbf{x})}{\partial \Delta \mathbf{x}}=0 \\ \frac{\partial }{\partial \Delta \mathbf{x}}[(\frac{\partial D}{\partial \mathbf{x}}{)}^T\Delta \mathbf{x}]=\frac{\partial D}{\partial \mathbf{x}} \\ \frac{\partial }{\partial \Delta \mathbf{x}}[\frac{1}{2}\Delta {\mathbf{x}}^T\frac{{\partial }^{2}D}{\partial \mathbf{x}\partial {\mathbf{x}}^T}\Delta \mathbf{x}]=\frac{1}{2}[\frac{{\partial }^{2}D}{\partial \mathbf{x}\partial {\mathbf{x}}^T}+(\frac{{\partial }^{2}D}{\partial \mathbf{x}\partial {\mathbf{x}}^T}{)}^T]\Delta \mathbf{x}=\frac{{\partial }^{2}D}{\partial \mathbf{x}\partial {\mathbf{x}}^T}\Delta \mathbf{x} \end{gathered}$$

用到的知识点：1. $\frac{\partial (AB)}{\partial B}=A^T$；2. $\frac{\partial ({\mathbf{x}}^{T}A\mathbf{x})}{\partial \mathbf{x}}=\frac{1}{2}(A+A^T)\mathbf{x}$；3. $\frac{{\partial }^{2}D}{\partial \mathbf{x}\partial {\mathbf{x}}^T}$是对称矩阵，所以$\frac{{\partial }^{2}D}{{\partial }^2\mathbf{x}\partial {\mathbf{x}}^T}=(\frac{{\partial }^{2}D}{\partial \mathbf{x}\partial {\mathbf{x}}^T}{)}^T$

那么
$$\frac{\partial D(\mathbf{x}+\Delta \mathbf{x})}{\partial \Delta \mathbf{x}}=\frac{\partial D}{\partial \mathbf{x}}+\frac{{\partial }^{2}D}{\partial \mathbf{x}\partial {\mathbf{x}}^T}\Delta x$$
令上面求导等于0，求出$\Delta \mathbf{x}$为
$$\Delta \mathbf{x}=-(\frac{{\partial }^{2}D}{\partial \mathbf{x}\partial {\mathbf{x}}^T}{)}^{-1}\frac{\partial D}{\partial \mathbf{x}}$$
把求得的$\Delta \mathbf{x}$代入泰勒展开式
$$\begin{array}{rl}D(\hat{\mathbf{x}})=D(\mathbf{x}+\Delta \mathbf{x})& =D(\mathbf{x})+(\frac{\partial D}{\partial \mathbf{x}}{)}^T\Delta \mathbf{x}-\frac{1}{2}(\frac{\partial D}{\partial \mathbf{x}}{)}^T[(\frac{{\partial }^{2}D}{\partial \mathbf{x}\partial {\mathbf{x}}^T}{)}^{-1}{]}^T\frac{{\partial }^{2}D}{\partial \mathbf{x}\partial {\mathbf{x}}^T}\Delta \mathbf{x}\\ & =D(\mathbf{x})+(\frac{\partial D}{\partial \mathbf{x}}{)}^T\Delta \mathbf{x}-\frac{1}{2}(\frac{\partial D}{\partial \mathbf{x}}{)}^T[(\frac{{\partial }^{2}D}{\partial \mathbf{x}\partial {\mathbf{x}}^T}{)}^T{]}^{-1}\frac{{\partial }^{2}D}{\partial \mathbf{x}\partial {\mathbf{x}}^T}\Delta \mathbf{x}\\ & =D(\mathbf{x})+(\frac{\partial D}{\partial \mathbf{x}}{)}^T\Delta \mathbf{x}-\frac{1}{2}(\frac{\partial D}{\partial \mathbf{x}}{)}^T(\frac{{\partial }^{2}D}{\partial \mathbf{x}\partial {\mathbf{x}}^T}{)}^{-1}\frac{{\partial }^{2}D}{\partial \mathbf{x}\partial {\mathbf{x}}^T}\Delta \mathbf{x}\\ & =D(\mathbf{x})+(\frac{\partial D}{\partial \mathbf{x}}{)}^T\Delta \mathbf{x}-\frac{1}{2}(\frac{\partial D}{\partial \mathbf{x}}{)}^T\Delta \mathbf{x}\\ & =D(\mathbf{x})+\frac{1}{2}(\frac{\partial D}{\partial \mathbf{x}}{)}^T\Delta \mathbf{x}\\ & =D(\mathbf{x})-\frac{1}{2}(\frac{\partial D}{\partial \mathbf{x}}{)}^T(\frac{{\partial }^{2}D}{\partial \mathbf{x}\partial {\mathbf{x}}^T}{)}^{-1}\frac{\partial D}{\partial \mathbf{x}}\end{array}$$