20260625:很多新内容

This commit is contained in:
2026-06-25 14:08:47 +08:00
parent 91fac5b6fc
commit 6021dea160
375 changed files with 19263 additions and 251 deletions

View File

@@ -0,0 +1,35 @@
---
title: "高斯滤波"
created: 2026-06-22
updated: 2026-06-22
type: concept
tags: [state-estimation, filtering, gaussian-approximation]
sources: [nano-filter]
---
# 高斯滤波
Gaussian filtering 是 [[bayesian-filtering|贝叶斯滤波]]中计算效率最高的一族方法。核心假设:每个时间步的状态分布近似为高斯分布:
$$
p(x_t | y_{1:t-1}) \approx N(x_t; \hat{x}_{t|t-1}, P_{t|t-1}), \quad
p(x_t | y_{1:t}) \approx N(x_t; \hat{x}_{t|t}, P_{t|t})
$$
## 两类线性化策略
| 策略 | 原理 | 代表算法 |
|------|------|----------|
| Taylor 展开 | $g(x) \approx g(\bar{x}) + g'(\bar{x})(x - \bar{x})$ | [[extended-kalman-filter|EKF]], IEKF |
| 统计线性回归 | 最小化残差期望 $E\|y - Ax - b\|^2$ | [[unscented-kalman-filter|UKF]], CKF, GHKF, [[posterior-linearization-filter|PLF]] |
## NANO 的超越
[[nano-filter|NANO filter]] 跳出了「先线性化再跑 KF」的使能框架直接从变分优化视角构造 Gaussian 滤波:
- 预测步 → 矩匹配(等价于 UKF/CKF 的做法)
- 更新步 → 在 [[gaussian-manifold|高斯流形]]上用 [[natural-gradient-descent|自然梯度下降]]直接最小化更新代价,**避免线性化误差**
## 参考
- [[bayesian-filtering|Bayesian Filtering]]
- [[kalman-filter|Kalman Filter]]
- [[nano-filter|NANO Filter]]