EKF-SLAM convergence
本文以與 Smith 等人相同的估計理論架構，證明線性高斯情形下 EKF-SLAM 的三項性質：相對地圖不確定性單調下降、極限時地標估計完全相關、而絕對誤差下限只由初始車輛不確定性決定。作者強調維持完整地圖共變異數（交互相關項）是收斂與一致性的必要條件，省略它會造成不一致與發散。文中以毫米波雷達與車輛實作，並用測量過的地標位置進行比對，同時指出運算與儲存量隨地標數平方成長。
本頁內容
Proves convergence properties of full-covariance EKF-SLAM in the linear-Gaussian case and demonstrates it with a radar-equipped vehicle against surveyed landmarks, while noting O(N^2) cost.
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 | millimetre-wave radar (77 GHz FMCW, beam scanned 360 deg in azimuth)、drive-shaft encoders (vehicle speed)、LVDT on the steering rack (steering for vehicle heading) |
|---|---|
| 原文測試平台 | conventional utility vehicle, driven manually at speeds up to 10 m/s |
| 狀態估計 | extended Kalman filter over vehicle pose and point landmarks (linear-Gaussian analysis for the proofs) |
| 資料關聯 | point landmarks from thresholded radar returns; Appendix 2 maintains confirmed and tentative landmark lists and associates an observation to a landmark when its Mahalanobis-type distance is below a threshold (d_min); new landmarks are validated by observation counts and a quality measure |
| 時間表示 | discrete poses |
| 去畸變 | 原文未報告 |
| 迴圈閉合 | implicit through full landmark covariance; no separate loop-closure module |
| 全域最佳化 | none (recursive filter) |
| 地圖表示 | point landmark map with full covariance |
| 先驗資訊 | none for estimation; surveyed landmark positions used only for evaluation |
| 可輸出幾何 | landmark coordinates and vehicle trajectory estimates with covariance |
| 計算需求 | 原文未報告 (full text read: radar, encoder and steering data were logged by an on-board computer and processed without reported hardware or timing) |
使用設備
原文使用的感測器、運算硬體與載具（equipment）。型號保留原文寫法，連結到設備頁中同一型號的歸併名稱；角色依原文用途分為方法輸入、資料集感測器、執行運算平台、參考或真值量測（reference or ground truth）與比較對象設備。
| 類別 | 型號（原文寫法） | 角色 | 資料集 | 原文規格 | 出處 |
|---|---|---|---|---|---|
| 雷達 | 77 GHz FMCW millimetre-wave radar (MMWR) | 方法輸入 | 未標示 | beam scanned 360 deg in azimuth at 1 to 3 Hz; amplitude returns at about 1.5 deg angular increments, thresholded to range and bearing; range to 250 m with 10 cm range and 1.5 deg bearing resolution; dual-polarisation receiver | (Dissanayake et al., 2001, Sec. IV-A) |
| 輪式或腿式里程計 | drive-shaft encoders | 方法輸入 | 未標示 | measure vehicle speed | (Dissanayake et al., 2001, Sec. IV-A) |
| 載具平台 | conventional utility vehicle | 方法輸入 | 未標示 | standard road vehicle fitted with the MMWR as primary sensor; driven manually, stationary about 30 s then loops at up to 10 m/s | (Dissanayake et al., 2001, Sec. IV-A, IV-B, Fig. 2) |
| 其他 | Linear Variable Differential Transformer (LVDT) on the steering rack | 方法輸入 | 未標示 | measures steering for vehicle heading | (Dissanayake et al., 2001, Sec. IV-A) |
| 其他 | radar reflectors (10, surveyed) | 參考或真值量測 | 未標示 | omni-directional point landmarks whose locations were accurately surveyed; used to evaluate the map and to compute the reference vehicle path (about 5 cm absolute accuracy) | (Dissanayake et al., 2001, Sec. IV-A, IV-B) |
作者報告的優勢與限制
優勢
- Proof that relative map uncertainty decreases monotonically and absolute accuracy reaches a lower bound set by initial vehicle uncertainty (abstract; Sec. III).
- Experimental results compared with surveyed landmark locations (abstract; Sec. IV).
- Actual vehicle errors stayed inside the filter's 95% confidence bounds, so the estimates were consistent, conservative and non-divergent (Sec. IV-B.1, Fig. 8).
- Landmark standard deviations decreased monotonically and reached a common lower bound matching the initial vehicle uncertainty, as predicted by the theory (Sec. IV-B.2, Figs. 12 and 13).
限制
- Computation and storage grow as N^2 with the number of landmarks (Sec. V).
- Point-landmark framework is stated to be infeasible where geometric features are hard to detect, e.g., an underground mine (Sec. V).
- Implementation is relatively small scale; map management for large areas is left open (Sec. V).
- Durrant-Whyte and Bailey (2006, Part I, footnote 2 in Sec. III.C of the author-prepared copy; Sec. IV.A adds that convergence and consistency can only be guaranteed in the linear case) note that these results are proved only for the linear Gaussian case.
- Only about 30% of radar observations corresponded to identifiable point landmarks; a large freight vehicle and nearby buildings also produced returns, so landmark identification and data association were essential (Sec. IV-A.4, Fig. 5).
- The filter assumed a constant-velocity vehicle model, which caused a jump in x error when the vehicle accelerated; richer models add computation (Sec. IV-B.1).
- Landmark estimates showed some bias, within the roughly 0.1 m accuracy of the surveyed truth (Sec. IV-B.2).
- Using only point landmarks is inefficient indoors because information such as ranges to walls is not used (Sec. V).
營建工程相關證據
未在營建場域測試；實驗場地為布設 10 個雷達反射器的戶外測試場，場內另有大型貨運車輛與鄰近建物（Sec. IV-A.4、IV-B.2）。作者在引言把採礦與營建等任務中的自主全地形車輛列為 SLAM 的潛在應用（Sec. I，作者動機；作者自存稿與版本紀錄第 229 頁皆有此句）。作者並明言在地下礦坑等難以偵測幾何特徵的環境中，點地標策略不可行（Sec. V），與隧道及地下工程的退化問題相關（推論）。
原文驗證環境：受控實驗、獨立參考量測
報告的性能數據
性能數據仍在分批查證，目前尚未收錄此方法的報告值。
來源
Dissanayake et al., 2001
（2001）A solution to the simultaneous localization and map building (SLAM) problemIEEE Transactions on Robotics and Automation, 17(3):229-241
同儕審查已出版已讀全文經典查證後修正