Cartographer
Cartographer 以背包式平台即時產生竣工平面圖：局部端把連續掃描以非線性最佳化對齊到小型子地圖（submap），誤差隨時間累積；全域端把已完成的子地圖與所有掃描做迴圈候選，以分支定界（branch-and-bound）加速的逐像素掃描匹配產生迴圈約束，再以稀疏位姿調整（SPA）定期最佳化。作者的貢獻在於降低迴圈約束計算成本，使數萬平方公尺樓層也能即時完成最佳化。
本頁內容
Submap-based 2D LiDAR SLAM whose branch-and-bound scan-to-submap matching makes loop-closure constraint search fast enough for real-time pose optimization on a backpack.
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 | horizontally mounted 2D LIDAR on the backpack (model not reported)、IMU on the backpack (model not reported)、Neato Robotics Revo LDS low-cost laser distance sensor (second experiment) |
|---|---|
| 原文測試平台 | backpack、manually pushed trolley carrying a vacuum-cleaner Revo LDS |
| 狀態估計 | nonlinear least squares (Ceres) for scan-to-submap matching and sparse pose adjustment for global optimization |
| 資料關聯 | scan-to-submap matching on probability grids; branch-and-bound pixel-accurate search for loop-closure constraints |
| 時間表示 | discrete poses |
| 去畸變 | 原文未報告 (IMU is used to project scans to the horizontal plane on the unstable backpack) |
| 迴圈閉合 | all finished submaps and scans considered; branch-and-bound scan matching in a search window adds loop constraints in real time |
| 全域最佳化 | sparse pose adjustment over scan and submap poses every few seconds |
| 地圖表示 | 2D probability-grid submaps (5 cm resolution in the paper) |
| 先驗資訊 | none |
| 可輸出幾何 | 2D grid floor plan (5 cm resolution) and optimized poses |
| 計算需求 | Deutsches Museum data (1,913 s, 2,253 m): 1,018 s CPU, up to 2.2 GB and up to 4 background loop-closure threads, 360 s wall clock (5.3x real time) on an Intel Xeon E5-1650 at 3.2 GHz (Sec. VI.A); Radish datasets processed in 10 to 190 s wall clock for 424 to 7,678 s of data on the same workstation, with parameters not tuned for CPU performance (Sec. VI.C, Table V) |
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 | Revo LDS | 方法輸入 | 未標示 | laser distance sensor used in Neato vacuum cleaners, costs under $30; scans taken at approximately 2 Hz over its debug connection | (Hess et al., 2016, Sec. VI.B) |
| LiDAR | horizontally mounted LIDAR on the Cartographer backpack (model not reported)歸入：Livox Horizon | 方法輸入 | 未標示 | provides the laser scans matched to 2D submaps; on the unstable backpack its scans are projected into the 2D world using the IMU gravity estimate; model, range and scan rate not reported | (Hess et al., 2016, Sec. III, IV, VI) |
| 慣性量測單元（IMU） | IMU on the Cartographer backpack (model not reported) | 方法輸入 | 未標示 | used to estimate the orientation of gravity for projecting scans from the horizontally mounted LIDAR; model and rate not reported | (Hess et al., 2016, Sec. IV) |
| 載具平台 | Cartographer backpack | 方法輸入 | 未標示 | sensor-equipped backpack with a horizontally mounted LIDAR and an IMU used to estimate gravity for projecting scans; sensor models not reported; 2D grid maps at 5 cm resolution | (Hess et al., 2016, Sec. III, IV) |
| 載具平台 | vacuum cleaner pushed on a trolley | 方法輸入 | 未標示 | Revo LDS data captured by pushing the vacuum cleaner around on a trolley | (Hess et al., 2016, Sec. VI.B) |
| 運算硬體 | Intel Xeon E5-1650 | 執行運算平台 | 未標示 | 3.2 GHz, in a workstation used for all timing; up to 4 background threads for loop-closure scan matching on the Deutsches Museum data | (Hess et al., 2016, Sec. VI.A, VI.C) |
| 其他 | laser tape measure | 參考或真值量測 | 未標示 | used for five reference straight-line lengths compared with the Revo LDS floor plan | (Hess et al., 2016, Sec. VI.B, Table I) |
作者報告的優勢與限制
優勢
- Real-time loop closure at 5 cm resolution on modest hardware (abstract; Sec. VII).
- Tuning reported as needed only for the sensor configuration, not for the specific surroundings (Sec. VI.C).
- Loop-closure precision of 93.4% to 99.8% on five of the six Radish datasets, with enough constraints in all cases and Huber loss in SPA helping robustness to false positives (Sec. VI.C, Table IV).
- Branch-and-bound search is exact: it returns the same match as the naive exhaustive search when inner-node scores are upper bounds (Sec. V.B).
限制
- Local scan matching accumulates error that is removed only by the global loop-closure optimization (Sec. III; Sec. IV).
- Loop-closure matching must keep pace with incoming scans (soft real-time constraint) (Sec. III).
- (inference) Geometric validation in the paper is five line lengths on one floor plan and relative-pose benchmarks, not 3D point-cloud accuracy.
- On the Radish benchmarks, the authors report considerably worse results than Graph Mapping on MIT CSAIL and state they cannot be sure that parameters were not fitted to the specific locations of the public datasets (Sec. VI.C, Table II).
- Scan-to-submap matching produces false-positive loop constraints; on Freiburg hospital, low resolution and a low minimum score gave 77.3% precision, and raising the score lowers some ground-truth metrics (Sec. VI.C, Table IV).
- The submap-scan loop constraints have no ground truth; precision is defined by constraints not violated by more than 20 cm or 1 deg after SPA (Sec. VI.C).
- Incorrect constraints can arise in locally symmetric environments such as office cubicles (Sec. V.A).
營建工程相關證據
作者明確以竣工平面圖（as-built floor plans）與建築管理為應用動機，並以背包系統在德意志博物館（既有建築）實測（Sec. VI.A，無獨立幾何參考）。另以人推手推車上的低價 Revo LDS 建成 5 cm 平面圖，與雷射捲尺量得的 5 條直線長度比對（Sec. VI.B, Table I）；文中未說明此實驗地點。均為既有建築 2D 平面圖，不是施工中工地，也不是 3D 點雲品質驗證。
原文驗證環境：已完工建築、公開基準、獨立參考量測
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 21 個比較組，合計 199 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 17 組列在最後，並連到性能比較頁。
Zou et al., 2022 · Table VII 本方法 33 筆
指標Error (m)
表格設定（擷取紀錄原文）：Exp. I data from Secs. IV-B.2 and IV-B.3 rerun with Cartographer in three sensor settings; end-point error (m) as listed (signed); GT end points (18.007, 0), (11.985, 5.003), (0, 0), (0, 0) (Zou et al., 2022, Table VII)
Error (m)，authors' own indoor recordings (rosbag) · line, run 1
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。
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 Zou et al., 2022 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(Zou et al., 2022, Table VII)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Cartographer, LiDAR本方法 | 0.003 m | (Zou et al., 2022, Table VII) |
| Cartographer, LiDAR (+IMU)本方法 | -0.016 m | (Zou et al., 2022, Table VII) |
| Cartographer, LiDAR (+IMU+Wheel)本方法 | 0.002 m | (Zou et al., 2022, Table VII) |
Hess et al., 2016 · Table II 本方法 28 筆
資料集與序列Radish · Aces
表格設定（擷取紀錄原文）：Radish benchmarks, relative-pose error against manually verified relations (metric of Kuemmerle et al. [21]); mean with std; GM values quoted from [21] (Hess et al., 2016, Table II)
Absolute translational, mean ± std 0.0375 ± 0.0426，Radish · Aces
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 Hess et al., 2016 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 |
|---|---|---|
| Cartographer本方法原文提出 | 0.0375 m | (Hess et al., 2016, Table II) |
Rogers et al., 2020 · Table I 本方法 21 筆
表格設定（擷取紀錄原文）：Artifact-based absolute mapping score on SubT-Tunnel: SLAM map aligned to the surveyed darpa frame by Umeyama on >=3 surveyed AprilTags (stereo depth); hand-coded artifact sightings scored as a point if within 5 m of the surveyed position; RMSE and max error over artifact reports. '*' is not defined in the table (per Sec. IV ORB-SLAM2 runs were stopped once the robot left the lit area); '-' = not run. Score converted to percent from the reported fraction. (Rogers et al., 2020, Table I)
RMSE of artifact position error against surveyed ground truth，SubT-Tunnel · sr_B_route1.bag, run length 909 m
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。
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 Rogers et al., 2020 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(Rogers et al., 2020, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| OmniMapper | 1.12 m | (Rogers et al., 2020, Table I) |
| Cartographer本方法 | 3.3 m | (Rogers et al., 2020, Table I) |
| Cartographer 2D (artifacts and detections projected to X-Y plane)本方法 | 1.5 m | (Rogers et al., 2020, Table I) |
| ORB SLAM2+ (modified recovery: continue from last pose when tracking is lost) | 13.7 m有附註註記（擷取紀錄）：other: score and RMSE marked with * in the table | (Rogers et al., 2020, Table I) |
| ORB SLAM2 | 1.2 m有附註註記（擷取紀錄）：other: score marked with * in Table I; the symbol is not defined in the table (Sec. IV only says each ORB-SLAM2 run was terminated once the robot left the lit area) | (Rogers et al., 2020, Table I) |
| Odometry (wheel odometry with IMU orientation, no mapping) | 13.25 m | (Rogers et al., 2020, Table I) |
Palieri et al., 2021 · Table II 本方法 14 筆
表格設定（擷取紀錄原文）：Husky field datasets from the SubT Urban (Alpha, Beta courses at the Satsop power plant) and Tunnel (Safety Research course, Bruceton mine) circuits; APE via evo against a reference from LOCUS scan matching on the DARPA ground-truth map; ME = RMSE of cloud-to-cloud error after ICP alignment of the map to the DARPA ground-truth map; loop closures disabled; FLOAM and LIO-Mapping ran with one LiDAR in Urban Alpha, LIO-SAM with one LiDAR; CPU loads from Urban Beta (LIO-SAM from Tunnel) (Palieri et al., 2021, Table II)
APE max，DARPA SubT Husky datasets (CoSTAR) · Urban Alpha course
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
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 Palieri et al., 2021 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(Palieri et al., 2021, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LOCUS原文提出 | 1.69 m | (Palieri et al., 2021, Table II) |
| LOCUS FGA原文提出 | 0.63 m | (Palieri et al., 2021, Table II) |
| BLAM | 3.44 m | (Palieri et al., 2021, Table II) |
| ALOAM | 4.33 m | (Palieri et al., 2021, Table II) |
| FLOAM | 29.49 m | (Palieri et al., 2021, Table II) |
| Cartographer本方法 | 5.84 m | (Palieri et al., 2021, Table II) |
| LIO-Mapping | 2.12 m | (Palieri et al., 2021, Table II) |
| LIO-SAM | 無數值失敗註記（擷取紀錄）：failed (authors could not get LIO-SAM working on the Urban datasets, likely because the 50 Hz IMU rate is below the recommended 200 Hz) | (Palieri et al., 2021, Table II) |
其他比較組
列出其餘 17 個比較組
- Hess et al., 2016 · Table IV
- Torres et al., 2023 · Table 1
- Hess et al., 2016 · Table I
- Chen et al., 2022a · Table III
- Hess et al., 2016 · Table III
- Labbé & Michaud, 2019 · Table 9
- Zou et al., 2022 · Table IX
- Hess et al., 2016 · Table V
- Hess et al., 2016 · Text Sec. VI.A
- Wang et al., 2023b · Table IV
- Zou et al., 2022 · Table III
- Zou et al., 2022 · Table V
- Zou et al., 2022 · Table VI
- Palieri et al., 2021 · Table III
- Zou et al., 2022 · Table IV
- Droeschel & Behnke, 2018 · Table I
- Ghadimzadeh Alamdari et al., 2025 · Table 3
來源
Hess et al., 2016
（2016）Real-time loop closure in 2D LIDAR SLAM2016 IEEE International Conference on Robotics and Automation (ICRA), pp. 1271-1278
DOI 10.1109/icra.2016.7487258程式碼
同儕審查已出版已讀全文經典查證後修正
相關版本
- 程式碼釋出：Cartographer (cartographer-project); the open-source code also contains 3D SLAM, which this 2D paper does not describe https://github.com/cartographer-project/cartographer
程式碼：https://github.com/cartographer-project/cartographer（授權：Apache-2.0）。有公開程式碼不等於已被重現，也不代表目前版本與論文版本相同。