当前位置:网站首页>UKFslam
UKFslam
2022-04-23 20:41:00 【xiaoma_bk】
UKF
UKF
- KF 系列求解:
Kalman filter
需要线性模型EKF
通过泰勒展开线性化- 更好的方式线性化 ->
Unscented Transform
->UKF
- 计算一组(所谓的)sigma 点
- 从变换和加权的 sigma 点计算高斯
Unscented Transform
- 计算一系列的 Sigma 点
- 每个Sigma点有一个权重
- 通过非线性函数转换 Sigma 点
- 权重点计算高斯
Sigma and weight
Sigma 点
- 选择 χ [ i ] {\chi^{[i]}} χ[i], w [ i ] {w^{[i]}} w[i] 使得:
- ∑ i w [ i ] = 1 {\sum_i w^{[i]} = 1} ∑iw[i]=1
- μ = ∑ i w [ i ] χ [ i ] { \mu = \sum_i w^{[i]}\chi^{[i]}} μ=∑iw[i]χ[i]
- ∑ = ∑ i w [ i ] ( χ [ i ] − μ ) ( χ [ i ] − μ ) T {\sum = \sum_i w^{[i]}(\chi^{[i]}-\mu)(\chi^{[i]}-\mu)^T} ∑=∑iw[i](χ[i]−μ)(χ[i]−μ)T
- 没有唯一的解决方案
- 如何选择Sigma点
- 第一个Sigma点也是均值 χ [ 0 ] = μ {\chi^[0] = \mu} χ[0]=μ
- χ [ i ] = μ + ( ( n + λ ) ∑ ) i {\chi^[i] = \mu + (\sqrt{(n+\lambda)\sum})_i} χ[i]=μ+((n+λ)∑)i for i=1,…,n
- χ [ i ] = μ − ( ( n + λ ) ∑ ) i − n {\chi^[i] = \mu - (\sqrt{(n+\lambda)\sum})_{i-n}} χ[i]=μ−((n+λ)∑)i−n for i=1+n,…,2n
- 矩阵平方根
- 定义 S S S, ∑ = S S \sum=SS ∑=SS
- 通过对角化计算:
- ∑ = V D V − 1 = ( d 11 ⋯ 0 ⋮ ⋱ ⋮ 0 ⋯ d n n ) = V ( d 11 ⋯ 0 ⋮ ⋱ ⋮ 0 ⋯ d n n ) ( d 11 ⋯ 0 ⋮ ⋱ ⋮ 0 ⋯ d n n ) V − 1 {\sum=VDV^{-1}= \begin{pmatrix} d_{11} & \cdots & 0 \\ \vdots & \ddots & \vdots \\ 0 & \cdots & d_{nn} \end{pmatrix} = V\begin{pmatrix} \sqrt{d_{11}} & \cdots & 0 \\ \vdots & \ddots & \vdots \\ 0 & \cdots & \sqrt{d_{nn}} \end{pmatrix} \begin{pmatrix} \sqrt{d_{11}} & \cdots & 0 \\ \vdots & \ddots & \vdots \\ 0 & \cdots & \sqrt{d_{nn}} \end{pmatrix} V^{-1}} ∑=VDV−1=⎝⎜⎛d11⋮0⋯⋱⋯0⋮dnn⎠⎟⎞=V⎝⎜⎛d11⋮0⋯⋱⋯0⋮dnn⎠⎟⎞⎝⎜⎛d11⋮0⋯⋱⋯0⋮dnn⎠⎟⎞V−1
- 因此可以定义: S = V D 1 / 2 V − 1 {S=VD^{1/2}V^{-1}} S=VD1/2V−1
Cholesky Matrix
平方根法- 矩阵平方根的替代定义: L , ∑ = L L T {L, \sum=LL^T} L,∑=LLT
- L , ∑ {L,\sum} L,∑ 有相同的特征向量
- Sigma 点可以但不必位于 ∑ {\sum} ∑
- 如何设置权重
- w m [ 0 ] = λ n + λ {w_m^{[0]}=\frac{\lambda}{n+\lambda}} wm[0]=n+λλ
- w c [ 0 ] = w m [ 0 ] + ( 1 − α 2 + β ) {w_c^{[0]}=w_m^{[0]}+(1-\alpha^2+\beta)} wc[0]=wm[0]+(1−α2+β)
- w c [ i ] = w m [ i ] + 1 2 ( n + λ ) {w_c^{[i]}=w_m^{[i]}+\frac{1}{2(n+\lambda)}} wc[i]=wm[i]+2(n+λ)1 for i=1,…,2n
- 选择 χ [ i ] {\chi^{[i]}} χ[i], w [ i ] {w^{[i]}} w[i] 使得:
UKF Algorithm
-
Prediction
-
χ t − 1 = ( μ t − 1 , μ t − 1 + ( n + λ ) ∑ t − 1 , μ t − 1 − ( n + λ ) ∑ t − 1 ) {\chi_{t-1}=(\mu_{t-1},\ \ \mu_{t-1}+\sqrt{(n+\lambda)\sum_{t-1}},\ \ \mu_{t-1}-\sqrt{(n+\lambda)\sum_{t-1}})} χt−1=(μt−1, μt−1+(n+λ)∑t−1, μt−1−(n+λ)∑t−1)
-
χ ˉ t ∗ = g ( u t , χ t − 1 ) {\bar{\chi}_t^* = g(u_t,\chi_{t-1})} χˉt∗=g(ut,χt−1)
-
μ t ˉ = ∑ i = 0 2 n w m [ i ] χ ˉ t ∗ [ i ] {\bar{\mu_t}=\sum_{i=0}^{2n}w_m^{[i]}\bar{\chi}_t^{*[i]}} μtˉ=∑i=02nwm[i]χˉt∗[i]
-
Σ ˉ t = ∑ i = 0 2 n w c [ i ] ( χ ˉ t ∗ [ i ] − μ t ˉ ) ( χ ˉ t ∗ [ i ] − μ t ˉ ) T + R t {\bar{\Sigma}_t=\sum_{i=0}^{2n}w_c^{[i]}(\bar{\chi}_t^{*[i]}-\bar{\mu_t})(\bar{\chi}_t^{*[i]}-\bar{\mu_t})^T+R_t} Σˉt=∑i=02nwc[i](χˉt∗[i]−μtˉ)(χˉt∗[i]−μtˉ)T+Rt
-
-
Correction
- χ t ˉ = ( μ t ˉ , μ t ˉ + ( n + λ ) ∑ t − 1 , μ t ˉ − ( n + λ ) ∑ t − 1 ) {\bar{\chi_{t} }=(\bar{\mu_{t}},\ \ \bar{\mu_{t}}+\sqrt{(n+\lambda)\sum_{t-1}},\ \ \bar{\mu_{t}}-\sqrt{(n+\lambda)\sum_{t-1}})} χtˉ=(μtˉ, μtˉ+(n+λ)∑t−1, μtˉ−(n+λ)∑t−1)
- z t ˉ = h ( χ t ˉ ) {\bar{z_t}=h(\bar{\chi_t})} ztˉ=h(χtˉ)
- z t ^ = ∑ i = 0 2 n w m [ i ] z ˉ t [ i ] {\hat{z_t}=\sum_{i=0}^{2n}w_m^{[i]}\bar{z}_t^{[i]}} zt^=∑i=02nwm[i]zˉt[i]
- S t = ∑ i = 0 2 n w c [ i ] ( χ ˉ t [ i ] − μ ˉ t ) ( χ ˉ t [ i ] − μ ˉ t ) T {S_t=\sum_{i=0}^{2n}w_c^{[i]}(\bar{\chi}_t^{[i]}-\bar{\mu}_t)(\bar{\chi}_t^{[i]}-\bar{\mu}_t)^T} St=∑i=02nwc[i](χˉt[i]−μˉt)(χˉt[i]−μˉt)T
- K t = ∑ ˉ t x , z S t − 1 {K_t = \bar{\sum}_t^{x,z}S_t^{-1}} Kt=∑ˉtx,zSt−1
- ∑ t = ∑ ˉ t − K t S t K t T {\sum_t =\bar{\sum}_t - K_tS_tK_t^T} ∑t=∑ˉt−KtStKtT
UT/UKF/EKF Summary
-
UT/UKF
- 无迹卡尔曼作为线性化的替代方案
- UT 是比泰勒展开更好的近似值
- UT 使用 sigma 点传播
- UT中的自由参数
- UKF 在预测和校正步骤中使用 UT
-
UKF VS EKF
- 线性模型的结果与 EKF 相同
- 非线性模型比 EKF 更好的近似
- 差异通常“有点小”
- UKF 不需要雅可比行列式
- 相同的复杂度类
- 比 EKF 稍慢
- 仍然受限于高斯分布
版权声明
本文为[xiaoma_bk]所创,转载请带上原文链接,感谢
https://blog.csdn.net/xiaoma_bk/article/details/124243965
边栏推荐
- A useless confession artifact
- 一些接地气的话儿
- Some basic configurations in interlij idea
- Learn to C language fourth day
- MySQL进阶之表的增删改查
- Vscode download speed up
- 居家第二十三天的午饭
- LeetCode 116. Populate the next right node pointer for each node
- How to do after winning the new debt? Is it safe to open an account online
- Flex layout
猜你喜欢
Mathematical modeling column | Part 5: MATLAB optimization model solving method (Part I): Standard Model
【SQL】字符串系列2:将一个字符串根据特定字符分拆成多行
DOS command of Intranet penetration
A useless confession artifact
【PTA】整除光棍
GO語言開發天天生鮮項目第三天 案例-新聞發布系統二
Devexpress 14.1 installation record
Leetcode 994, rotten orange
Resolve the eslint warning -- ignore the warning that there is no space between the method name and ()
How can matlab obtain the truncated image in trainingimagelabeler
随机推荐
LeetCode 994、腐烂的橘子
JS arrow function user and processing method of converting arrow function into ordinary function
Communication between RING3 and ring0
Introduction to standardization, regularization and normalization
Analysis of the relationship between generalized Bim and CAD under the current background
star
Scripy tutorial - (2) write a simple crawler
上海回应“面粉官网是非法网站”:疏于运维被“黑”,警方已立案
LeetCode-279-完全平方数
Plato Farm元宇宙IEO上线四大,链上交易颇高
Case of the third day of go language development fresh every day project - news release system II
GO语言开发天天生鲜项目第三天 案例-新闻发布系统二
[PTA] get rid of singles
BMP JPEG 图片转换为矢量图像 ContourTrace
On IRP from the perspective of source code
LeetCode 542、01 矩阵
On BIM data redundancy theory
C migration project record: modify namespace and folder name
vulnhub DC:1渗透笔记
【PTA】整除光棍