Thrun-Burgard-Fox real-time 2D and 3D laser mapping
本文把增量式雷射掃描匹配與以樣本表示的位姿後驗結合：後驗的計算方式與蒙地卡羅定位相同，每次掃描以多個樣本作為爬山搜尋的起點，找到最可能的位姿後把掃描加入地圖。當後驗推得的位姿與單純增量估計不一致時，系統判定發生迴圈閉合，先把位姿差按比例分配到迴圈內各位姿，再以梯度下降反覆修正，於兩次量測之間完成反向校正。同一後驗機制讓第二台機器人先在領隊機器人的地圖中全域定位，再共同建圖；另以向前雷射做 2D 定位、向上雷射擷取 3D 資料，產生經多邊形簡化的精簡建物 3D 模型。
本頁內容
Real-time incremental 2D laser mapping that combines maximum-likelihood scan alignment with a sample-based pose posterior for loop detection and backward correction, extended to multi-robot mapping and to compact 3D building models from an upward-pointing laser.
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 finders: forward-looking for 2D mapping and localization, upward-pointed for 3D (model not reported)、wheel odometry (optional; also run with odometry removed) |
|---|---|
| 原文測試平台 | wheeled UGV (Pioneer, RWI B21, Nomad Scout)、tracked UGV (Urban Robot) |
| 狀態估計 | incremental maximum-likelihood scan alignment by gradient ascent, combined with a sample-based posterior over the current pose computed like Monte Carlo localization; every sample seeds a hill-climbing search (Sec. 2.2-2.3) |
| 資料關聯 | no explicit correspondences: perceptual likelihood penalizes obstacles in space previously seen as free (ray-traced), with a map made of past scans and poses (Sec. 2.1) |
| 時間表示 | discrete poses (scans appended every 2 m; all scans used for localization) |
| 去畸變 | 原文未報告 |
| 迴圈閉合 | detected when the posterior-based pose differs from the incremental maximum-likelihood pose; the loop is identified from the scan that caused the adjustment (Sec. 2.4) |
| 全域最佳化 | backward correction over the loop only: the pose difference is distributed proportionally over the loop poses, then gradient descent is iterated over them (Sec. 2.4); no full batch optimization |
| 地圖表示 | collection of 2D scans with poses; 3D model as polygons from the upward laser, outlier-filtered and simplified with a computer-graphics polygon simplification (Sec. 2.6) |
| 先驗資訊 | none (for multi-robot mapping each robot must start inside the team leader's map) |
| 可輸出幾何 | 2D scan map and a simplified 3D polygonal model viewed in a standard virtual reality tool (VRweb) (Sec. 3.5) |
| 計算需求 | real time on a low-end PC; more than 1,000 gradient computations per second (Sec. 2.1; Sec. 3.1) |
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 | 2D laser range finders, forward-looking and upward-pointed (model not reported) | 方法輸入 | 未標示 | two lasers on the 3D-mapping Pioneer: the forward-looking one for 2D mapping and localization, the upward-pointed one for 3D data | (Thrun et al., 2000, Sec. 2.6; Fig. 1b caption) |
| 載具平台 | Pioneer | 方法輸入 | 未標示 | robots used for multi-robot mapping; one Pioneer carries two laser range finders for 3D mapping | (Thrun et al., 2000, Fig. 1 caption) |
| 載具平台 | Urban Robot | 方法輸入 | 未標示 | tracked skid-steering robot for indoor and outdoor exploration with extremely poor odometry | (Thrun et al., 2000, Fig. 1 caption; Sec. 3.3) |
| 載具平台 | RWI B21 | 方法輸入 | 未標示 | 原文未報告 | (Thrun et al., 2000, Sec. 1) |
| 載具平台 | Nomad Scout | 方法輸入 | 未標示 | 原文未報告 | (Thrun et al., 2000, Sec. 1) |
作者報告的優勢與限制
優勢
- Maps of cyclic environments built in real time on a low-end PC, with occasional injected odometry errors of 30 degrees or 1 m (Sec. 3.1).
- The same cycle was mapped with odometry removed, with a final result optically equivalent to the odometry-based map (Sec. 3.2; Fig. 8).
- The authors report never obtaining a single wrong result with backward correction, which runs between two sensor measurements (Sec. 2.4).
- A roughly 60 m cyclic 3D map was reduced from 82,899 to 8,289 polygons with similar appearance and about an order of magnitude faster rendering (Sec. 3.5).
限制
- Odometry-free mapping works only with sufficient environmental variation and would fail in a long featureless corridor (Sec. 3.2).
- With the Urban Robot's very poor odometry some walls were rotated by about 2 degrees (Sec. 3.3).
- Real-time operation comes at the price of increased brittleness compared with EM (Sec. 4).
- Multi-robot mapping assumes every robot starts within the team leader's map (Sec. 2.5).
- (inference) Map accuracy is judged visually; no metric error against an independent reference is reported.
營建工程相關證據
未在營建場域驗證。以水平雷射定位、向上雷射累積 3D 資料並簡化成多邊形模型，是早期以移動平台建立建物室內 3D 模型的做法，作者說明使用者可在 VR 工具中飛越模型遠端檢視建物（Sec. 3.5）。3D 點位精度完全取決於 2D 位姿估計，對工程級竣工量測的適用性未評估（推論）。Surmann 等（Surmann et al., 2003）把本文列為以水平加垂直雙雷射取得 3D 資料的代表；Hähnel 等（Hähnel et al., 2003b）的 2D 掃描對齊延伸自本文，並指出本文模型的多邊形數與原始掃描數相近（該文 Sec. 1、2.1）。
原文驗證環境：已完工建築
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 2 個比較組，合計 3 筆紀錄。
Thrun et al., 2000 · Text Sec. 3.3, 3.5 本方法 2 筆
資料集與序列authors' dual-laser Pioneer data · cyclic indoor map (about 60 m)
表格設定（擷取紀錄原文）：Same map after polygon simplification (Garland-Heckbert fusion); appearance similar and rendering about an order of magnitude faster (Thrun et al., 2000, Text Sec. 3.3, 3.5)
number of polygons of the simplified model，authors' dual-laser Pioneer data · cyclic indoor map (about 60 m)
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 Thrun et al., 2000 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 |
|---|---|---|
| simplified polygonal model本方法原文提出 | 8289 polygons | (Thrun et al., 2000, Sec. 3.5; Fig. 11) |
Hähnel et al., 2003a · Text Sec. IV 本方法 1 筆
指標loop closure outcome
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 |
|---|---|---|
| particle filter strategy of Thrun et al. [20], [19] (single map)本方法 | 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) |
| proposed RBPF with scan-matching-corrected odometry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) |
來源
Thrun et al., 2000
（2000）A real-time algorithm for mobile robot mapping with applications to multi-robot and 3D mappingProceedings 2000 IEEE International Conference on Robotics and Automation (ICRA 2000), San Francisco, CA, vol. 1, pp. 321-328
同儕審查已出版已讀全文經典查證後修正