SLAM 點雲圖鑑Point clouds for construction
M-01 · 2021

R-LOAM

R-LOAM 延伸 LOAM（A-LOAM 實作）的建圖模組：假設環境中有一個幾何與全域位姿皆已知的參考物件，先以物件包圍盒裁切掃描點，再透過 AABB 樹找出每個掃描點在三角網格上的最近虛擬點，形成點到網格（point-to-mesh）殘差，與 LOAM 的角點、面點殘差經正規化後共同最佳化，網格權重隨迭代次數以對數方式增加。驗證在 Gazebo 模擬的機庫 B737、廂型車與艾菲爾鐵塔三種情境，分別模擬 Velodyne VLP-16 與 Ouster OS1-128；相對 LOAM，三種情境的中位數 APE 平均降幅分別超過 92%、69% 與 94%，但需要 15 至 35 次迭代，且完全依賴網格與物件位姿的正確性。

既有地圖或 BIM 上定位近十年已讀全文同儕審查已出版LOAM 系列與特徵式 LiDAR 建圖

閱讀版本
version of record: IEEE Xplore HTML full text, IEEE RA-L 6(2):2068-2075 (published 18 February 2021); Tables I-IV read from the publisher's table images
全文取得途徑
國立臺灣大學機構訂閱
本頁內容

R-LOAM adds normalized point-to-mesh residuals, found via bounding-box scan isolation and AABB-tree closest-point search on a known, globally placed CAD mesh, to A-LOAM's mapping optimization; in Gazebo simulations with VLP-16 and OS1-128 it reduced median APE versus LOAM by over 92%, 69% and 94% on average in the three scenarios, at the cost of 15-35 optimization iterations and a perfect-model assumption.

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Simulated Velodyne VLP-16 (16 lines, up to 30,000 points per scan, 10 Hz, Gaussian noise sigma 0.03) and simulated Ouster OS1-128 (128 lines, up to 262,144 points per scan, 10 Hz, sigma 0.05), one at a time, on a 1-DoF gimbal tilting between -0.6 and +0.6 rad on a quadcopter UAV in Gazebo (Sec. IV-A, Fig. 4)
原文測試平台quadcopter UAV with 1-DoF LiDAR gimbal in Gazebo simulation: manual and autonomous flights around a B737 in a hangar and around the lower Eiffel Tower (Sec. IV-A, Table I)
狀態估計LOAM (A-LOAM) mapping optimization extended to a joint cost of normalized corner, surface and point-to-mesh residuals with Huber loss; mesh weight lambda raised logarithmically from 0.1 to 40 over up to 35 correspondence and optimization iterations; Ceres Levenberg-Marquardt trust region; A-LOAM's real-time abort was disabled so every scan is map-optimized (Sec. III-B3, IV-B)
資料關聯LOAM corner and surface point correspondences plus point-to-mesh correspondences: scans are cropped to the object's bounding box plus a buffer (2 m, 200 m for the Eiffel Tower, ground removed below 0.5 or 1 m), then for each remaining point the closest virtual point on the triangular mesh is found via an AABB tree and the IGL library (Sec. III-B1, III-B2, IV-B)
時間表示discrete per-scan poses as in A-LOAM; no additional time model described
去畸變原文未報告 (the paper does not describe motion or gimbal-actuation compensation beyond LOAM defaults)
迴圈閉合none; the paper notes LOAM performs no loop closure and R-LOAM does not add one (Sec. II-A, III)
全域最佳化none; drift is reduced by aligning each scan to the mesh within the per-scan map optimization (Sec. III-B3)
地圖表示LOAM feature map of corner and surface points stored in a cube structure; map building unchanged (Sec. III-A, III-B3)
先驗資訊exact 3D triangular mesh (CAD) of the reference object, up to hundreds of thousands of faces (B737, van or Eiffel Tower), with its exact pose in the map frame assumed known before takeoff (Sec. I, II-B, VI)
可輸出幾何per-scan 6-DoF poses and the LOAM feature map; improved 3D reconstruction is claimed but no map accuracy metric is reported (Sec. V, VII)
計算需求原文未報告; high iteration counts (15-35) are said to need a remote multi-core SLAM setup or post-processing to be real-time (Sec. 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
LiDARVelodyne VLP-16 (simulated in Gazebo)歸入：Velodyne VLP-16方法輸入R-LOAM simulated datasets 1, 3, 516 scan lines, up to 30,000 points per scan, 10 Hz, zero-mean Gaussian noise sigma 0.03(Oelsch et al., 2021, Sec. IV-A, Table I)
LiDAROuster OS1-128 (simulated in Gazebo)歸入：Ouster OS1-128方法輸入R-LOAM simulated datasets 2, 4, 6128 scan lines, up to 262,144 points per scan, 10 Hz, zero-mean Gaussian noise sigma 0.05(Oelsch et al., 2021, Sec. IV-A, Table I)
載具平台Quadcopter UAV (simulated in Gazebo)方法輸入R-LOAM simulated datasetsmanual flight in dataset 1; datasets 2-4 follow this trajectory with an autonomous flight controller, max 0.5 m/s (Sec. IV-A); control mode for the Eiffel Tower datasets 5-6 not stated(Oelsch et al., 2021, Sec. IV-A, Table I)
其他1-DoF gimbal (simulated)方法輸入R-LOAM simulated datasetstilts the LiDAR back and forth between -0.6 and +0.6 rad throughout the flight(Oelsch et al., 2021, Sec. IV-A, Fig. 4)
其他Gazebo simulator ground-truth pose參考或真值量測R-LOAM simulated datasets6-DoF ground-truth pose of the LiDAR in the global frame(Oelsch et al., 2021, Sec. IV-A)

作者報告的優勢與限制

優勢

  • Average reduction in median APE over LOAM of over 92% (scenario 1), 69% (scenario 2) and 94% (scenario 3) (Sec. V)
  • scenario 1 VLP-16: median APE 2.0 cm at 35 iterations vs best LOAM 30.9 cm, median RE 0.09 vs 1.59 deg
  • OS1-128: 1.2 cm vs 13.2 cm (Table II)
  • with the VLP-16 at the Eiffel Tower LOAM fails while R-LOAM reaches 0.3 cm median APE (Table IV)
  • gains remain with a small reference object (van), where only 22% (VLP-16) or 45% (OS1-128) of scans have more than 100 mesh features (Sec. V, Table III)

限制

  • Assumes a perfect mesh and exact object pose
  • mesh imperfections or object pose errors directly degrade R-LOAM and a large error could make it worse than LOAM (Sec. VI)
  • validated only in Gazebo simulation with added Gaussian noise (Sec. IV-A)
  • scan isolation assumes no other objects inside the object's bounding box (Sec. III-B1)
  • best results need 15-35 map optimization iterations, beyond LOAM's default 2, which the authors say requires remote computing or post-processing (Sec. V)
  • in scenario 2 (van) the VLP-16 error rose again after 15 iterations (Table III)

營建工程相關證據

作者在相關研究回顧營建工地的 In situ Fabricator 與以 CAD 建築模型定位末端執行器的機器人（Sec. II-B），並在限制一節指出 BIM 相關設備可產生次毫米精度的三維模型（Sec. VI）。本文驗證僅在 Gazebo 模擬的機庫飛機與艾菲爾鐵塔目視檢測情境，未在營建工地或以真實 BIM 模型驗證；（推論）以已知網格作為配準約束的做法可延伸到 BIM 先驗，但竣工幾何與 BIM 的偏差會直接影響結果。

原文驗證環境：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 4 個比較組，合計 75 筆紀錄。

Oelsch et al., 2021 · Table II 本方法 20 筆

表格設定（擷取紀錄原文）：Scenario 1, airplane as reference object; APE (cm) and RE (deg) against Gazebo ground truth for 2 to 35 correspondence and optimization iterations; LOAM = A-LOAM (Oelsch et al., 2021, Table II)

APE in cm, median，R-LOAM Gazebo simulated datasets · Dataset 1: VLP-16, 15111 scans, 0.35 m/s, 514 m, manual flight

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。

這些是 Oelsch 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。

統計量：中位數（median）；對齊方式：原文未報告；單位：cm；場景：Scenario 1 (B737 in hangar, airplane as reference)

資料來源作者報告值(Oelsch et al., 2021, Table II)

數值與出處
方法（原文寫法）報告值出處
LOAM [1], #Iter 2 (def)50.5 cm(Oelsch et al., 2021, Table II)
LOAM [1], #Iter 557.7 cm(Oelsch et al., 2021, Table II)
LOAM [1], #Iter 1551.1 cm(Oelsch et al., 2021, Table II)
LOAM [1], #Iter 2530.9 cm(Oelsch et al., 2021, Table II)
LOAM [1], #Iter 3531.9 cm(Oelsch et al., 2021, Table II)
R-LOAM, #Iter 2 (def)本方法原文提出10.5 cm(Oelsch et al., 2021, Table II)
R-LOAM, #Iter 5本方法原文提出6.6 cm(Oelsch et al., 2021, Table II)
R-LOAM, #Iter 15本方法原文提出3.1 cm(Oelsch et al., 2021, Table II)
R-LOAM, #Iter 25本方法原文提出2.8 cm(Oelsch et al., 2021, Table II)
R-LOAM, #Iter 35本方法原文提出2 cm(Oelsch et al., 2021, Table II)

Oelsch et al., 2021 · Table III 本方法 20 筆

表格設定（擷取紀錄原文）：Scenario 2, van as small reference object; APE (cm) and RE (deg) against Gazebo ground truth for 2 to 35 iterations; LOAM = A-LOAM (Oelsch et al., 2021, Table III)

APE in cm, median，R-LOAM Gazebo simulated datasets · Dataset 3: VLP-16, 9718 scans, 0.49 m/s, 474 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。

這些是 Oelsch 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。

統計量：中位數（median）；對齊方式：原文未報告；單位：cm；場景：Scenario 2 (B737 in hangar, van as reference)

資料來源作者報告值(Oelsch et al., 2021, Table III)

數值與出處
方法（原文寫法）報告值出處
LOAM [1], #Iter 2 (def)19 cm(Oelsch et al., 2021, Table III)
LOAM [1], #Iter 565 cm(Oelsch et al., 2021, Table III)
LOAM [1], #Iter 1524 cm(Oelsch et al., 2021, Table III)
LOAM [1], #Iter 2523.8 cm(Oelsch et al., 2021, Table III)
LOAM [1], #Iter 3523.9 cm(Oelsch et al., 2021, Table III)
R-LOAM, #Iter 2 (def)本方法原文提出17.9 cm(Oelsch et al., 2021, Table III)
R-LOAM, #Iter 5本方法原文提出7.3 cm(Oelsch et al., 2021, Table III)
R-LOAM, #Iter 15本方法原文提出6.3 cm(Oelsch et al., 2021, Table III)
R-LOAM, #Iter 25本方法原文提出13.5 cm(Oelsch et al., 2021, Table III)
R-LOAM, #Iter 35本方法原文提出13.5 cm(Oelsch et al., 2021, Table III)

Oelsch et al., 2021 · Table IV 本方法 20 筆

表格設定（擷取紀錄原文）：Scenario 3, Eiffel Tower as reference object; APE (cm) and RE (deg) against Gazebo ground truth for 2 to 35 iterations; LOAM = A-LOAM; authors state LOAM fails with the VLP-16 (Oelsch et al., 2021, Table IV)

APE in cm, median，R-LOAM Gazebo simulated datasets · Dataset 5: VLP-16, 5674 scans, 0.51 m/s, 291 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。

  • 失敗

這些是 Oelsch 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。

統計量：中位數（median）；對齊方式：原文未報告；單位：cm；場景：Scenario 3 (Eiffel Tower, tower as reference)

資料來源作者報告值(Oelsch et al., 2021, Table IV)

數值與出處
方法（原文寫法）報告值出處
LOAM [1], #Iter 2 (def)499.3 cm失敗註記（擷取紀錄）：failed (authors state LOAM fails in this scenario with a VLP-16)(Oelsch et al., 2021, Table IV)
LOAM [1], #Iter 5380.2 cm失敗註記（擷取紀錄）：failed (authors state LOAM fails in this scenario with a VLP-16)(Oelsch et al., 2021, Table IV)
LOAM [1], #Iter 151420.9 cm失敗註記（擷取紀錄）：failed (authors state LOAM fails in this scenario with a VLP-16)(Oelsch et al., 2021, Table IV)
LOAM [1], #Iter 251326.4 cm失敗註記（擷取紀錄）：failed (authors state LOAM fails in this scenario with a VLP-16)(Oelsch et al., 2021, Table IV)
LOAM [1], #Iter 351338.5 cm失敗註記（擷取紀錄）：failed (authors state LOAM fails in this scenario with a VLP-16)(Oelsch et al., 2021, Table IV)
R-LOAM, #Iter 2 (def)本方法原文提出31.5 cm(Oelsch et al., 2021, Table IV)
R-LOAM, #Iter 5本方法原文提出2.4 cm(Oelsch et al., 2021, Table IV)
R-LOAM, #Iter 15本方法原文提出0.5 cm(Oelsch et al., 2021, Table IV)
R-LOAM, #Iter 25本方法原文提出0.3 cm(Oelsch et al., 2021, Table IV)
R-LOAM, #Iter 35本方法原文提出0.3 cm(Oelsch et al., 2021, Table IV)

Oelsch et al., 2022 · Table IV 本方法 15 筆

表格設定（擷取紀錄原文）：Three hangar datasets with M = 9 and L = 15; means of 5 online runs; APE mapping = map-optimized poses, APE TMO = poses used for TMO; 3-DoF position ground truth from Leica MS60 (Oelsch et al., 2022, Table IV)

APE mapping (cm), max，RO-LOAM hangar datasets (octocopter, VLP-16, Leica MS60 ground truth) · Dataset 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。

這些是 Oelsch 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。

統計量：最大值（max）；對齊方式：原文未報告；單位：cm；場景：aircraft hangar, one side of a B737 (visual inspection)

資料來源作者報告值(Oelsch et al., 2022, Table IV)

數值與出處
方法（原文寫法）報告值出處
LOAM [1]466.3 cm(Oelsch et al., 2022, Table IV)
LOAM + RO原文提出84.2 cm(Oelsch et al., 2022, Table IV)
R-LOAM [2]本方法50.7 cm(Oelsch et al., 2022, Table IV)
R-LOAM + RO原文提出57.8 cm(Oelsch et al., 2022, Table IV)

來源

  • Oelsch et al., 2021

    Martin Oelsch, Mojtaba Karimi, Eckehard Steinbach（2021）R-LOAM: Improving LiDAR Odometry and Mapping With Point-to-Mesh Features of a Known 3D Reference ObjectIEEE Robotics and Automation Letters, 6(2): 2068-2075

    DOI 10.1109/lra.2021.3060413

    同儕審查已出版已讀全文近十年查證後修正

回到方法圖鑑

SLAM 點雲圖鑑

本站為綜述論文《SLAM-Based Point Clouds for Construction: A Review》研究過程的互動說明文件。互動圖為教學模擬，不代表任何特定系統的實測表現；方法描述、數值與結論以所引原始文獻為準。

內容基準日 2026-09-25 · 研究與網站製作使用 AI 輔助工具，見 證據標記與方法

選擇開啟Esc關閉