Grid-based FastSLAM with scan matching
本文把 Rao-Blackwellized 粒子濾波與雷射掃描匹配結合：每 k 步先以前 k-1 筆掃描與最近的里程計讀值做掃描匹配，得到修正後的里程量測並用於粒子取樣，再以第 k 筆掃描計算粒子權重，使每筆資料只使用一次。掃描匹配殘差以三參數誤差模型描述，參數由 Intel Research Lab 資料學得，因此取樣分布比原始里程計集中得多，所需粒子數與重取樣次數下降，也減輕粒子耗盡，使機器人能閉合大迴圈。每個粒子各有一張佔據網格地圖，但只用與其可視區域相交的有限掃描更新。
本頁內容
Grid-based FastSLAM that converts scan-matching results into corrected odometry with a learned error model before Rao-Blackwellized particle sampling, reducing particle count and depletion so that large loops close with about 100 particles.
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 | 2D laser range finder (SICK LMS)、wheel odometry |
|---|---|
| 原文測試平台 | wheeled UGV (Pioneer 2)、simulation (B21r simulator) |
| 狀態估計 | Rao-Blackwellized particle filter over robot paths with one grid map per particle; every k steps a scan-matching-corrected odometry measurement is computed from the k-1 previous scans and the k most recent odometry readings and used for sampling with a learned three-parameter error model, and the k-th scan weights the particles (Sec. III) |
| 資料關聯 | grid-based 2D scan matching of a scan against an occupancy grid built from previous measurements, using a beam-endpoint likelihood (for max-range readings the cell 20 cm before the end is assumed free) (Sec. III) |
| 時間表示 | discrete poses |
| 去畸變 | 原文未報告 |
| 迴圈閉合 | implicit through the particle filter; the scan-matching correction reduces resampling operations and particle depletion so that large loops can be closed (Sec. I, III) |
| 全域最佳化 | none |
| 地圖表示 | 2D occupancy grid per particle, updated from a limited number of scans that intersect the particle's visible area (constant-time approximation); 10 cm grid in the Sieg Hall run (Sec. III; Sec. IV.A) |
| 先驗資訊 | none |
| 可輸出幾何 | 2D occupancy grid map |
| 計算需求 | real time with 100 samples; for the standard RBPF, 200 samples was the real-time limit and 1000 samples the memory limit on a 1.8 GHz Pentium IV PC with 768 MB (Sec. IV) |
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 |
|---|---|---|---|---|---|
| LiDAR | SICK LMS | 方法輸入 | 未標示 | 原文未報告 | (Hähnel et al., 2003a, Sec. IV.A) |
| 載具平台 | Pioneer 2 | 方法輸入 | 未標示 | 原文未報告 | (Hähnel et al., 2003a, Sec. IV.A) |
| 運算硬體 | 1.8GHz Pentium IV PC | 執行運算平台 | 未標示 | 768 MB main memory | (Hähnel et al., 2003a, Sec. IV.B) |
| 其他 | B21r simulator | 方法輸入 | 未標示 | simulator of a B21r robot used to generate the Wean Hall data (32 m x 10 m, 251 m, noise added to the ground truth) | (Hähnel et al., 2003a, Sec. IV.B) |
作者報告的優勢與限制
優勢
- Globally consistent map of the Intel Research Lab (28 m x 28 m, 491 m traveled) built in real time with 100 samples (Sec. IV.A; Fig. 6).
- Consistent real-time map of Sieg Hall (50 m x 12 m) despite repeated loops, with 100 samples and a 10 cm grid (Sec. IV.A; Fig. 8).
- In simulation, a single-map posterior approach that keeps only the best particle at loop closure produced inconsistencies, while the proposed method produced a consistent map (Sec. IV.B; Fig. 10).
- A standard RBPF without scan-matching correction did not converge on the Intel data with up to 1000 samples (Sec. IV.B).
限制
- The real-time 100-sample map is less sharp than the scan-matching-only map; a crisper 500-particle map took several hours (Sec. IV.A).
- Each particle's map is updated from a limited set of scans, an approximation (Sec. III).
- The scan-matching error model was learned by treating a map produced by the system itself as ground truth (Sec. III).
- (inference) Map quality is judged visually; no metric error against an independent reference is given, and the window k is not stated.
營建工程相關證據
未在營建場域驗證；實驗在 Intel Research Lab、University of Washington Sieg Hall 等既有建築與模擬環境中進行，只以目視判斷地圖一致性。以掃描匹配先修正里程再進入粒子濾波的設計，是 FastSLAM （Montemerlo et al., 2002）與 GMapping （Grisetti et al., 2007）之間的方法銜接，對大型建築室內 2D 建圖的迴圈閉合穩健性有參考價值（推論）。
原文驗證環境：已完工建築、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 4 筆紀錄。
Hähnel et al., 2003a · Text Sec. IV 本方法 4 筆
資料集與序列B21r simulator, Wean Hall · simulated run
表格設定（擷取紀錄原文，這些數值分屬表中不同部分）：(Hähnel et al., 2003a, Text Sec. IV)
- Simulated Wean Hall (32 m x 10 m, 251 m, noise added); single-map posterior approach keeping only the best particle at loop closure
- Simulated Wean Hall, proposed method
loop closure outcome，B21r simulator, Wean Hall · simulated run
只並列這張表在相同設定下報告的方法；以「本方法：」開頭者為本頁方法。失敗、未執行與未報告以標記呈現，不是 0。
- 失敗
- 未報告（沒有數值，不是 0）
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 Hähnel et al., 2003a 在此表設定下報告的數值（author-reported results），只能在同一個比較組內對照，不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Hähnel et al., 2003a, Text Sec. IV)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| proposed RBPF with scan-matching-corrected odometry本方法原文提出 | 無數值未報告註記（擷取紀錄）：consistent map | (Hähnel et al., 2003a, Sec. IV.B; Fig. 10) |
| particle filter strategy of Thrun et al. [20], [19] (single map) | 無數值失敗註記（擷取紀錄）：failed: inconsistencies after closing the loop | (Hähnel et al., 2003a, Sec. IV.B; Fig. 10) |
來源
Hähnel et al., 2003a
（2003）An efficient FastSLAM algorithm for generating maps of large-scale cyclic environments from raw laser range measurementsProceedings 2003 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2003), Las Vegas, NV, vol. 1, pp. 206-211
同儕審查已出版已讀全文經典查證後修正