LRGC (Local Registration and Global Correlation)
LRGC 以 Lu 與 Milios 的一致位姿估計為核心，分兩種方式使用：每加入一筆新掃描，只與最近 K 個位姿做局部配準，所以每步計算量固定；偵測到迴圈後，才對整個迴圈做一致位姿估計。迴圈偵測不用單一掃描，而是把最新 m 筆掃描組成地圖片段，以相關運算在較舊的地圖中搜尋，並以高匹配分數、低歧異與低變異三個條件拒絕誤判；搜尋範圍與片段大小都隨位置不確定度增加。作者稱這是第一個不需操作者輸入即可在大型循環環境即時產生稠密度量地圖的系統。
本頁內容
Incremental 2D laser mapping with constant-time local Lu-Milios registration over the last K poses and correlation of multi-scan map patches for loop detection, followed by consistent pose estimation over the closed loop.
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 (180 deg SICK)、wheel odometry |
|---|---|
| 原文測試平台 | wheeled UGV (B21, Pioneer II, Pioneer I) |
| 狀態估計 | Lu-Milios consistent pose estimation (least-squares pose network) applied to the last K poses for each new scan and, after a loop is found, to the poses along the loop with sparse linear algebra and strong-link reduction (maximum 200 poses) (Sec. 2.2) |
| 資料關聯 | pairwise scan matching combining Cox's point-to-line method with Lu-Milios matching (Gutmann-Schlegel method); loop detection by correlating a patch of the newest m scans with the older map on a grid, accepted only if match score is high, ambiguity low and variance low (Sec. 2.1, 2.3) |
| 時間表示 | discrete poses (new scan added after about 20 to 50 cm of motion) |
| 去畸變 | 原文未報告 |
| 迴圈閉合 | patch-to-map correlation run every few scans inside a Mahalanobis-gated search area; patch size grows with position uncertainty; an accepted topological link cannot be undone (Sec. 2.3-2.4) |
| 全域最佳化 | consistent pose estimation over the closed loop, rerun with new scan matches between newly linked poses; optional final optimization over all poses at the end of a run (Sec. 2.2, 2.4) |
| 地圖表示 | undirected graph of robot poses with attached scans; links from dead reckoning, scan matching or correlation (Sec. 2.4) |
| 先驗資訊 | none |
| 可輸出幾何 | 2D dense metric scan-point maps |
| 計算需求 | incremental registration under 100 ms in typical circumstances; closing even large loops under 10 s in almost all cases (Sec. 2.2); constant time except when closing loops |
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 laser range finder | 方法輸入 | 未標示 | 180 deg field of view | (Gutmann & Konolige, 1999, Sec. 3) |
| 載具平台 | B21 | 方法輸入 | 未標示 | 原文未報告 | (Gutmann & Konolige, 1999, Sec. 3) |
| 載具平台 | Pioneer II | 方法輸入 | 未標示 | 原文未報告 | (Gutmann & Konolige, 1999, Sec. 3) |
| 載具平台 | Pioneer I | 方法輸入 | 未標示 | 原文未報告 | (Gutmann & Konolige, 1999, Sec. 3) |
作者報告的優勢與限制
優勢
- For K >= 7, local registration differed from full registration by less than 1 mm average error per pose on a 150-pose map (Sec. 2.2; Fig. 3).
- Closed a cycle of about 200 m in Wean Hall (80 m x 25 m, two cycles) and corrected strong directional-carpet odometry drift in an 85 m x 15 m SRI environment (Sec. 3; Figs. 7-8).
- Constant-time incremental updates; large loops closed in less than 10 s in almost all cases (Sec. 2.2).
限制
- A topological connection, once made, cannot be undone, so false positives are critical (Sec. 2.3).
- Requires good scan-matching results; the authors report that stereo-camera range data was much less accurate, but the rest of that passage is missing from the VoR scan (Sec. 4).
- If the filters reject too many good matches loops cannot be closed, and if they accept a false match the map becomes inconsistent; multiple loop hypotheses are left for future work (Sec. 4, author copy).
- Local registration can sometimes differ radically from global registration, which the authors did not examine closely (Sec. 2.2).
營建工程相關證據
未在營建場域驗證；資料來自 CMU Wean Hall、SRI 人工智慧中心、卡內基自然史博物館與 Freiburg 人工智慧實驗室等既有建築。作者以多筆掃描組成的地圖片段做迴圈偵測，理由是單筆掃描在沿走廊方向變化少時難以排除誤判（Sec. 1.4），這與大型建築長走廊與重複樓層的建圖風險直接相關（推論）。
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 1 個比較組，合計 3 筆紀錄。
Gutmann & Konolige, 1999 · Text Sec. 2.2 本方法 3 筆
資料集與序列150-pose map (source log not identified in the paper; about 0.3 m between poses) · 150-pose map
表格設定（擷取紀錄原文）：Local registration over the last K poses versus update of all poses on a 150-pose map (about 0.3 m between poses); pose error from incremental differences between consecutive poses (Gutmann & Konolige, 1999, Text Sec. 2.2)
average error per pose for K >= 7 (smaller than a millimeter)，150-pose map (source log not identified in the paper; about 0.3 m between poses) · 150-pose map
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 Gutmann & Konolige, 1999 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 |
|---|---|---|
| LRGC local registration (K >= 7)本方法原文提出 | 1 mm僅報告範圍註記（擷取紀錄）：upper bound: below 1 mm | (Gutmann & Konolige, 1999, Sec. 2.2; Fig. 3; footnote 1) |
來源
Gutmann & Konolige, 1999
（1999）Incremental mapping of large cyclic environmentsProceedings 1999 IEEE International Symposium on Computational Intelligence in Robotics and Automation (CIRA'99), Monterey, CA, pp. 318-325
同儕審查已出版已讀全文經典查證後修正
相關版本
- author copy：Author pre-print on Konolige's SRI page (conclusion wording differs slightly from the VoR) http://www.ai.sri.com/~konolige/papers/mapping.pdf