RO-LOAM
RO-LOAM 是可外掛在 LiDAR SLAM 上的「參考物件式軌跡與地圖最佳化」：LOAM 本身不修改，每隔 L 幅掃描便把最近 M+1 幅裁切後的掃描以 ICP 對齊到已知參考物件的稠密點雲模型，再以 EKF 運動先驗檢查最後一個對齊位姿是否與前序一致（0.05 m 與 0.5 度內），通過者以高權重加入兩次修正之間的位姿圖，最佳化後重新插入過去掃描以修正 LOAM 地圖。實驗以八旋翼無人機搭載由 Dynamixel 致動器旋轉的 Velodyne VLP-16，在機庫內沿 B737 單側飛行三次，以 Leica Nova MS60 追蹤稜鏡作為地面真值；啟用後 LOAM 的中位數 APE 由約 67 至 81 cm 降到約 6.4 至 6.8 cm，且可在邊緣雲端伺服器上線上執行。
本頁內容
RO-LOAM periodically aligns short sequences of isolated scans to a known reference point-cloud model with ICP, accepts the last pose only if it agrees with an EKF motion prior, and uses it as a high-weight pose-graph constraint to correct the trajectory since the previous correction and to rebuild the LOAM map; on three real UAV hangar flights it cut LOAM median APE from about 67-81 cm to about 6.4-6.8 cm while running online on a server.
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 | Velodyne VLP-16 (10 Hz) on a Dynamixel actuator rotating the LiDAR about the roll axis within +/-40 deg、actuator readings transform scans to the robot frame、Intel NUC logs data to SSD (Sec. III-A, IV-A) |
|---|---|
| 原文測試平台 | octocopter UAV flown by an operator inside an aircraft hangar along one side of a B737 (Sec. IV-A, Fig. 3) |
| 狀態估計 | Unmodified LOAM (A-LOAM) map-optimized poses; every L scans the latest M+1 isolated scans (each with more than 50 points, downsampled to 500) are aligned to the reference point-cloud model by ICP with Huber loss and Levenberg-Marquardt (1 m max correspondence distance, 100 iterations, 2 s cap); the last pose is a candidate if its MSE is below 0.001 and it lies within 0.05 m and 0.5 deg of an EKF motion prior (ROS robot_localization, R = 0.01 I) built from the previous aligned poses; accepted poses become high-confidence constraints in a Ceres pose graph; final setting M = 9, L = 15 (Sec. III-C to III-E, IV-B, V) |
| 資料關聯 | raw scans transformed to the world frame with actuator readings and LOAM poses, cropped to the reference object by a bounding box, then point-to-point nearest-neighbour correspondences to the dense model point cloud within 1 m (Sec. III-B, III-C) |
| 時間表示 | discrete per-scan poses; the LiDAR actuation is removed with actuator readings (Sec. III-A) |
| 去畸變 | actuator motion compensated with actuator readings before LOAM; no ego-motion deskew beyond LOAM defaults is described (Sec. III-A) |
| 迴圈閉合 | none; global correction comes from model-aligned trajectory and map optimization (TMO); authors state it may complement relocalization or loop closure methods (Sec. VI) |
| 全域最佳化 | subgraph between consecutive TMOs: first node (previous TMO) fixed, relative-pose edges with identity information, the new TMO constraint weighted 4000 I; Ceres Levenberg-Marquardt with Huber loss (Sec. III-E) |
| 地圖表示 | LOAM voxelized feature map in a cube structure; corrected by reinserting previous scans with pose-graph-optimized poses (Sec. III-A, III-F) |
| 先驗資訊 | known location and geometry of a large static reference object (CAD from BIM or dense point cloud); experiments use a dense B737 point cloud with over 1.6 M points and 1.4 cm mean nearest-neighbour distance scanned by a Leica Nova MS60; the object-to-LiDAR relative pose is estimated before each flight from the first 20 static scans with RMSE below 3 cm (Sec. I, IV-A, IV-B) |
| 可輸出幾何 | corrected trajectory and corrected LOAM feature map (Fig. 5); no map accuracy metric is reported |
| 計算需求 | designed to be offloaded to an edge cloud in parallel threads; experiments online with 10 Hz scans on a server with 32 Intel Xeon E5-2690 @2.90 GHz CPUs and 132 GB memory; LOAM mapping frequency 2.6 to 3.3 Hz across configurations (Sec. IV-B, V, Table 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 | Velodyne VLP-16 | 方法輸入 | 未標示 | 10 Hz update rate; continuously rotated about the roll axis within +/-40 deg | (Oelsch et al., 2022, Sec. IV-A) |
| 地面雷射掃描儀（TLS） | Leica Nova MS60 MultiStation | 方法輸入 | 未標示 | its scanning functionality generated the B737 reference point cloud: over 1.6 M points, mean nearest-neighbour distance 1.4 cm (Sec. IV-A, Fig. 1) | (Oelsch et al., 2022, Sec. IV-A, Fig. 1) |
| 全測站 | Leica Nova MS60 MultiStation | 參考或真值量測 | 未標示 | measures ground-truth 3-DoF position (no rotation) of a mini prism on a rod | (Oelsch et al., 2022, Sec. IV-A, Fig. 3b) |
| 載具平台 | Octocopter UAV (model not stated) | 方法輸入 | 未標示 | controlled by an operator inside a hangar along one side of a B737 | (Oelsch et al., 2022, Sec. IV-A, Fig. 3a) |
| 運算硬體 | Intel NUC (onboard) | 資料集感測器 | 未標示 | saves LiDAR data on an SSD for offline evaluation | (Oelsch et al., 2022, Sec. IV-A) |
| 運算硬體 | Server with 32 Intel Xeon E5-2690 CPUs | 執行運算平台 | 未標示 | 2.90 GHz, 132 GB memory; all experiments run online with 10 Hz scans | (Oelsch et al., 2022, Sec. IV-B) |
| 其他 | Dynamixel actuator (model not stated) | 方法輸入 | 未標示 | rotates the LiDAR about the roll axis between +/-40 deg; readings used for the LiDAR-to-robot transform | (Oelsch et al., 2022, Sec. III-A, IV-A) |
| 其他 | Leica mini prism on a rod | 參考或真值量測 | 未標示 | low-weight prism mounted on top of a rod for good visibility throughout the flight; its 3-DoF position measured by the MS60 is the ground truth | (Oelsch et al., 2022, Sec. IV-A) |
作者報告的優勢與限制
優勢
- Median APE of map-optimized poses fell from 71.1, 67.2 and 80.6 cm (LOAM) to 6.5, 6.4 and 6.8 cm (LOAM + RO), and from 11.5, 9.6 and 15.3 cm (R-LOAM) to 5.9, 5.2 and 6.7 cm (R-LOAM + RO) on datasets 1 to 3 (Table IV, means of 5 runs)
- over 50 successful TMOs per dataset with max APE below 30 cm and median below 5 cm (Sec. V)
- nearly no extra onboard computation when run remotely (Sec. V)
限制
- Relies on an accurate initial relative pose between robot and reference object
- any error directly harms TMO (Sec. VI)
- deviations between model and actual geometry increase scan-to-model error (Sec. VI)
- not a standalone SLAM algorithm (Sec. VI)
- only 7-16% of scan-to-model aligned poses reach APE below 10 cm, so pure model-based localization would fail (Table I, Sec. V)
- online results not fully reproducible, reported as means of 5 runs (Sec. V)
- ground truth is 3-DoF position only
- evaluated in one hangar scenario with three flights
營建工程相關證據
作者在引言指出，建築 CAD 模型或以 BIM 相關高精度設備取得的稠密點雲可作為參考物件，並回顧營建工地 In situ Fabricator 與以 CAD 模型定位施工機器人的研究（Sec. I, II）。本文驗證僅在機庫內 B737 單側目視檢測的三組 UAV 飛行資料，參考模型由 Leica Nova MS60 掃描建立，並非 BIM 模型，也不是營建工地；地面真值為全測站追蹤稜鏡的三自由度位置。（推論）以參考模型間歇修正軌跡與地圖的架構可延伸到以 BIM 或既有點雲作為先驗的營建掃描，但竣工幾何與模型的偏差會直接影響結果（Sec. VI）。
原文驗證環境：已完工建築、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 4 個比較組，合計 109 筆紀錄。
Oelsch et al., 2022 · Table IV 本方法 60 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。
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 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。
資料來源作者報告值(Oelsch et al., 2022, Table IV)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LOAM [1] | 466.3 cm | (Oelsch et al., 2022, Table IV) |
| LOAM + RO本方法原文提出 | 84.2 cm | (Oelsch et al., 2022, Table IV) |
| R-LOAM [2] | 50.7 cm | (Oelsch et al., 2022, Table IV) |
| R-LOAM + RO本方法原文提出 | 57.8 cm | (Oelsch et al., 2022, Table IV) |
Oelsch et al., 2022 · Table II 本方法 20 筆
資料集與序列RO-LOAM hangar datasets (octocopter, VLP-16, Leica MS60 ground truth) · Dataset 1
表格設定（擷取紀錄原文）：Parameter study on Dataset 1 for LOAM + RO with L = 50; APE of map-optimized poses (cm) and number of successful TMOs (Oelsch et al., 2022, Table II)
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。
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 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。
資料來源作者報告值(Oelsch et al., 2022, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LOAM + RO (M = 4, L = 50)本方法原文提出 | 79.5 cm | (Oelsch et al., 2022, Table II) |
| LOAM + RO (M = 9, L = 50)本方法原文提出 | 94.9 cm | (Oelsch et al., 2022, Table II) |
| LOAM + RO (M = 19, L = 50)本方法原文提出 | 96.7 cm | (Oelsch et al., 2022, Table II) |
| LOAM + RO (M = 29, L = 50)本方法原文提出 | 58.5 cm | (Oelsch et al., 2022, Table II) |
Oelsch et al., 2022 · Table III 本方法 20 筆
資料集與序列RO-LOAM hangar datasets (octocopter, VLP-16, Leica MS60 ground truth) · Dataset 1
表格設定（擷取紀錄原文）：Parameter study on Dataset 1 for LOAM + RO with M = 9; APE of map-optimized poses (cm) and number of successful TMOs (Oelsch et al., 2022, Table III)
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。
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 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。
資料來源作者報告值(Oelsch et al., 2022, Table III)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LOAM + RO (L = 15, M = 9)本方法原文提出 | 84.2 cm | (Oelsch et al., 2022, Table III) |
| LOAM + RO (L = 100, M = 9)本方法原文提出 | 96.3 cm | (Oelsch et al., 2022, Table III) |
| LOAM + RO (L = 200, M = 9)本方法原文提出 | 91.9 cm | (Oelsch et al., 2022, Table III) |
| LOAM + RO (L = 300, M = 9)本方法原文提出 | 92.6 cm | (Oelsch et al., 2022, Table III) |
Oelsch et al., 2022 · Table I 本方法 9 筆
表格設定（擷取紀錄原文）：Share of scan-to-model aligned poses below an APE threshold, using LOAM map-optimized poses as initial guess (Oelsch et al., 2022, Table I)
scan-to-model aligned poses with APE < 10 cm，RO-LOAM hangar datasets (octocopter, VLP-16, Leica MS60 ground truth) · Dataset 1
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 Oelsch et al., 2022 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 |
|---|---|---|
| scan-to-model alignment (ICP) initialized with LOAM poses本方法原文提出 | 9% | (Oelsch et al., 2022, Table I) |
來源
Oelsch et al., 2022
（2022）RO-LOAM: 3D Reference Object-based Trajectory and Map Optimization in LiDAR Odometry and MappingIEEE Robotics and Automation Letters, 7(3): 6806-6813
同儕審查已出版已讀全文近十年查證後修正