Smith-Cheeseman spatial uncertainty
本文以「近似轉換（approximate transformation, AT）」表示座標框架之間不確定的相對位姿，每個 AT 由平均關係與共變異數矩陣組成。作者定義兩個基本運算：串接（compounding）以一階泰勒展開與 3×6 雅可比矩陣傳遞共變異數，把一連串 AT 合成一個，不確定性隨之變大；合併（merging）以靜態卡爾曼濾波公式加權平均平行的 AT，使不確定性變小，並以電阻串並聯作類比。對惠斯登電橋這類無法以串並聯化簡的網路，作者提出刪除形成迴路的 AT（未用上全部資訊，並非最佳）或以 Delta-Y 轉換改寫網路（可用上全部資訊，但無法化簡所有網路），並表示以共同參考框架做遞迴狀態估計的通用方法仍在研究中。移動機器人範例與蒙地卡羅模擬顯示，平均值與共變異數的相對誤差通常小於 1%，但角度誤差大時分布呈新月形而非高斯分布。
本頁內容
Defines approximate transformations (mean plus covariance), first-order compounding and Kalman-based merging of parallel relations, and a network-reduction procedure validated against Monte Carlo simulation; the joint stochastic-map formulation is only announced here as work in progress.
技術屬性
欄位內容為文獻擷取紀錄的原文用語（英文），以原文為據；「未查證」表示本研究尚未讀到該資訊，不代表該方法不具備此能力。
| 感測輸入 | 未記錄 |
|---|---|
| 原文測試平台 | ["simulation"] |
| 狀態估計 | first-order (linearized) mean and covariance propagation for compounding and reversal of ATs; merging of parallel ATs with static-state Kalman filter equations (optimal for Gaussian variables and linear mappings, optimal-linear otherwise); extended Kalman filter update named for nonlinear coordinate mappings |
| 資料關聯 | no data-association algorithm; the sensing procedure rejects an observation whose probability, given the prior AT estimate and the sensor error, is below a threshold (e.g., the camera viewed the wrong object), which the authors also describe as detecting sensor glitches |
| 時間表示 | discrete moves; each relative motion and each sensing is a static AT in the network |
| 去畸變 | 不適用 |
| 迴圈閉合 | implicit: when the robot observes its start frame from a later pose, the sensed AT is merged with the compounded chain as a parallel relation; no separate loop-closure module |
| 全域最佳化 | none; the network is reduced by repeated compounding and merging; irreducible (Wheatstone-bridge) networks are handled approximately by deleting loop-forming ATs (nonoptimal) or by a Delta-Y transformation, and a general recursive state-estimation method is stated as under investigation |
| 地圖表示 | network of approximate transformations (relational map), each with a mean relation and covariance; the robot keeps the original ATs from motions and sensings and computes composite ATs on demand |
| 先驗資訊 | none required; the robot's starting position is taken as the world frame |
| 可輸出幾何 | mean and covariance of the relative pose (x, y, theta) between any two frames; confidence ellipses derived from the covariance |
| 計算需求 | no runtime or hardware reported; the authors state that the procedures need only simple matrix computations and are computationally simple and fast |
使用設備
尚未收錄此方法的設備紀錄；設備資料仍在分批查證，沒有紀錄不代表原文未使用任何設備。
作者報告的優勢與限制
優勢
- ["First-order estimates agreed with an independent Monte Carlo simulation
- relative errors of the estimated means and covariances were typically below 1% unless angular errors were large (standard deviation above about 6 degrees) (Sec. 6.3, Sec. 7).", "Probabilistic estimates avoid the overly conservative worst-case max/min error bounds of earlier work (Sec. 3).", "Lets a robot decide in advance whether a motion or sensing step will reach the accuracy a task needs, and judge when a sensor has glitched (Sec. 6.1, Sec. 7)."]
限制
- ["First-order approximation needs small errors and can underestimate or overestimate the covariance (Sec. 3.2).", "Errors are assumed independent and zero-mean
- systematic errors are not modelled and must be removed by calibration (Sec. 3.2, Sec. 4.2).", "Series and parallel reduction cannot reduce every network (e.g., a Wheatstone bridge)
- deleting ATs discards information and the Delta-Y method is not general (Sec. 5).", "Only two moments are estimated
- with large angular errors the distribution becomes crescent-shaped rather than Gaussian (Sec. 6.3, Sec. 7, Fig. 5).", "Extension to six degrees of freedom introduces Jacobian singularities that can destroy the estimates (Sec. 7)."]
營建工程相關證據
未在營建場域測試（僅有蒙地卡羅模擬）。其「串接使不確定性增加、平行量測合併使不確定性減少」的共變異數傳遞觀念，以及系統誤差須靠校正排除的前提，可作為討論 SLAM 點雲座標誤差累積、控制點與校正需求的理論背景（推論）。
原文驗證環境：模擬
報告的性能數據
以下是原文作者報告的性能數值（author-reported results），不是本研究重新量測的結果。每張圖只並列同一個比較組（comparison group，同一張表、同一組實驗設定）內的方法；不同比較組之間的數值不可直接比較，也不構成排名。
本方法共出現在 1 個比較組，合計 1 筆紀錄。
Smith & Cheeseman, 1986 · Text Sec.6.3 本方法 1 筆
指標relative error in any component of the estimated means and covariances (compared to the simulated values)
資料集與序列Monte Carlo simulation (authors)
表格設定（擷取紀錄原文）：first-order AT estimates compared with an independent Monte Carlo simulation of a three-degree-of-freedom robot with Gaussian errors in the given relations (Smith & Cheeseman, 1986, Text Sec.6.3)
relative error in any component of the estimated means and covariances (compared to the simulated values)，Monte Carlo simulation (authors)
這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Smith & Cheeseman, 1986 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| compounding of approximate transformations (first-order estimate)本方法原文提出 | 1%僅報告範圍註記（擷取紀錄）：upper bound stated as typically less than 1%; does not hold when angular errors are large (standard deviation greater than 6 degrees) | (Smith & Cheeseman, 1986, Sec. 6.3) |
來源
Smith & Cheeseman, 1986
（1986）On the Representation and Estimation of Spatial UncertaintyThe International Journal of Robotics Research, 5(4):56-68
DOI 10.1177/027836498600500404
同儕審查已出版已讀全文經典