SLAM 點雲圖鑑Point clouds for construction
M-01 · 2022

VIRAL-Fusion

VIRAL-Fusion 以滑動視窗最佳化融合三類觀測：IMU 預積分、UWB 測距，以及機上既有自我定位系統（如 VINS-Fusion 與 A-LOAM）輸出的相鄰位姿變化量。UWB 測距模型同時考慮天線相對機體中心的偏移與量測時刻落在兩個狀態之間的時間差，因此能約束姿態與速度。由於 UWB 錨點在現場先行自我定位並定義場址座標系，估測結果不隨時間累積漂移，也不需要迴圈閉合。

里程計＋局部建圖近十年已讀全文同儕審查已出版LiDAR、視覺與慣性融合

閱讀版本
IEEE version of record (T-RO, HTML through NTU access; all sections I-VII and Tables II-VII images); arXiv v1 (2020-10-23) consulted as an earlier manuscript
全文取得途徑
國立臺灣大學機構訂閱
本頁內容

Sliding-window optimization that fuses IMU preintegration, body-offset UWB ranges with explicit temporal offsets, and pose displacements from existing onboard self-localization systems (VIO and LiDAR odometry), so that anchor ranging bounds drift in a site-fixed frame without loop closure.

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["IMU (VectorNav VN-100, per footnote link)", "body-offset UWB ranging (Humatics P440, per footnote link、two UAV nodes with two antennae each and three or four anchors)", "camera system (footnote links the ueye_cam driver、used through VINS-Fusion odometry)", "two LiDARs (Ouster OS1, per footnote link、horizontal and vertical, used through A-LOAM odometry)"]
原文測試平台["UAV in field flight tests (airframe type not stated in the version of record)", "simulated UAV in AirSim (Building_99)", "EuRoC MAV sequences with simulated UWB"]
狀態估計sliding-window nonlinear least squares in Ceres over orientation, position, velocity and IMU biases; cost combines UWB body-offset range factors, IMU preintegration factors and pose-displacement factors from one or more onboard self-localization (OSL) systems such as VIO or LiDAR odometry (Sec. IV-A, Eq. 31; Sec. V)
資料關聯loosely coupled at the OSL level: the relative pose change reported by each OSL system between two time steps is used as an observation (Sec. V-A); UWB ranges are checked by SNR, leading-edge quality, rate of change and an IMU-predicted range test before use (Sec. IV-B)
時間表示discrete time steps independent of sensor timing; IMU and OSL samples are interpolated at step boundaries and each UWB range keeps its time offset within the step, with orientation and velocity interpolated at that offset in the range model (Sec. IV-B, V-C, Eq. 85)
去畸變not described in the fusion method; LiDAR data enter only through the A-LOAM odometry estimates
迴圈閉合no; drift is bounded by ranging to fixed UWB anchors instead of loop closure
全域最佳化none beyond the sliding window; global consistency comes from the anchor-defined world frame
地圖表示UWB anchor positions in a site frame from anchor self-localization; no dense map is built by the fusion
先驗資訊UWB anchor positions estimated on site by an anchor self-localization step before flight (three anchors at 1 m height; anchor 0 as origin, x axis toward anchor 1, y axis toward the inspected structure) (Sec. IV-B, VI-C; Table V); fixed sensor extrinsics
可輸出幾何UAV pose and velocity in the anchor-defined world frame at the optimization rate plus IMU-propagated high-rate pose
計算需求ROS and Ceres implementation; optimization time per step shown in Fig. 16; the VoR does not name the onboard computer (the arXiv v1 manuscript mentions a small form-factor PC with an Intel i7-8650U)

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
LiDAROS1 (two units)方法輸入VIRAL field flight tests (authors)one horizontal and one vertical LiDAR, each processed by A-LOAM as an OSL input(Nguyen et al., 2022b, Sec. VI-C; Fig. 10)
慣性量測單元（IMU）VN-100歸入：VN100方法輸入VIRAL field flight tests (authors)IMU on the flight platform (VoR footnote)(Nguyen et al., 2022b, Sec. VI-C; Fig. 10)
相機camera system (footnote links the ueye_cam ROS driver)方法輸入VIRAL field flight tests (authors)camera system processed by VINS-Fusion as an OSL input; stereo configuration not stated for the field platform(Nguyen et al., 2022b, Sec. VI-C; footnote)
UWB 測距P440方法輸入VIRAL field flight tests (authors)UWB ranging nodes; two UAV nodes with two antennae each; three anchors in field tests(Nguyen et al., 2022b, Sec. VI-C; Figs. 10-11)
UWB 測距simulated UWB network (ranges generated from Vicon data)方法輸入EuRoC MAV (simulated UWB)4 anchors at (3,3,3), (3,-3,0.5), (-3,-3,3), (-3,3,0.5) m; 2 UAV nodes with 2 antennae; 80 Hz; 0.05 m noise(Nguyen et al., 2022b, Sec. VI-A1)
全測站Leica MS60參考或真值量測VIRAL field flight tests (authors)millimeter-level accuracy; position ground truth for the flight tests(Nguyen et al., 2022b, Sec. VI-C)
其他AirSim (Building_99 environment)資料集感測器AirSim Building_99 (authors' simulation)simulated sensors: 10 Hz stereo camera, two 10 Hz LiDARs, 400 Hz IMU, two UWB nodes with two antennae each, four anchors at 11 m height(Nguyen et al., 2022b, Sec. VI-B)
其他VICON motion capture (EuRoC ground truth)參考或真值量測EuRoC MAV (simulated UWB)position and orientation ground truth of the EuRoC V1 and V2 sequences; also used to generate the simulated UWB ranges(Nguyen et al., 2022b, Sec. VI-A1)
其他onboard attitude and heading reference system (AHRS) of the drone (model not reported)參考或真值量測VIRAL field flight tests (authors)fuses magnetometer measurements; used as orientation ground truth for Table VII because the Leica station gives no orientation(Nguyen et al., 2022b, Sec. VI-C)

作者報告的優勢與限制

優勢

  • ["On EuRoC V1/V2 with simulated UWB, positional RMSE is lower than VINS-Fusion and ORB-SLAM3 with loop closure and bundle adjustment on all six sequences (Table II)", "Fusing OSL data improves both position and rotation over the UWB-IMU-only variant (Tables II, III, IV, VI, VII)", "In three field inspection flights with Leica MS60 ground truth, translational RMSE 0.1593, 0.3334 and 0.1745 m versus 0.40 to 14.27 m for single LOAM or VINS-Fusion estimates (Table VI)", "An earlier UWB-VIO method could not run on the body-offset ranging setup, while VIRAL handled it (Sec. VI-A1
  • Remark 2)", "Anchor network of three anchors can be deployed and self-localized on site before flight (Sec. IV-B
  • Fig. 11)"]

限制

  • Requires UWB anchors to be deployed and surveyed or self-localized on site; accuracy depends on anchor geometry
  • Rotational RMSE on EuRoC is not always lower than ORB-SLAM3 with bundle adjustment (Table III)
  • LiDAR and camera information enter only through OSL pose displacements, so no map is optimized
  • Orientation ground truth in field tests is the onboard AHRS rather than an external system (Sec. VI-C)
  • Field positional error is larger than in EuRoC and AirSim, attributed to anchor self-localization error, cable-induced range bias and an obtuse three-anchor geometry (Sec. VI-C; Remark 4)

營建工程相關證據

現地測試是無人機在平行於待檢測垂直結構物的平面內飛行，模擬環境則是室內大廳；三個 UWB 錨點可在 10 至 15 分鐘內布設到 60 m 以上範圍並自動定位，同時定義朝向結構物的場址座標系。這與施工現場需要固定座標系、重複巡檢或分段作業的需求相近（推論）。不過本文只評估軌跡，沒有點雲或地圖精度；作者也指出錨點幾何與高度會影響精度，實際工地的遮蔽與布設限制並未驗證。

原文驗證環境：公開基準、模擬、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 5 個比較組，合計 54 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 1 組列在最後，並連到性能比較頁。

Nguyen et al., 2022b · Table IV 本方法 18 筆

表格設定（擷取紀錄原文）：AirSim Building_99 indoor foyer simulation: 10 Hz stereo, two 10 Hz LiDARs (horizontal and vertical, A-LOAM), 400 Hz IMU, two UWB nodes with two antennae each and four ceiling anchors at 11 m; three simulated flights (Nguyen et al., 2022b, Table IV)

Translational RMSE [m]，AirSim Building_99 (authors' simulation) · Test 01

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 Nguyen et al., 2022b 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：均方根誤差（RMSE）；對齊方式：原文未報告；單位：m；場景：simulated indoor foyer

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

數值與出處
方法（原文寫法）報告值出處
Vertical LOAM0.2047 m(Nguyen et al., 2022b, Table IV; Sec. VI-B)
Horizontal LOAM0.1089 m(Nguyen et al., 2022b, Table IV; Sec. VI-B)
VINS (w/o BA)0.4709 m(Nguyen et al., 2022b, Table IV; Sec. VI-B)
VINS (with BA)0.0942 m(Nguyen et al., 2022b, Table IV; Sec. VI-B)
ORB-SLAM30.1275 m(Nguyen et al., 2022b, Table IV; Sec. VI-B)
VIRAL (w/o OSL)本方法原文提出0.0518 m(Nguyen et al., 2022b, Table IV; Sec. VI-B)
VIRAL (with OSL)本方法原文提出0.0518 m(Nguyen et al., 2022b, Table IV; Sec. VI-B)

Nguyen et al., 2022b · Table II 本方法 12 筆

指標Translational RMSE

表格設定（擷取紀錄原文）：EuRoC Vicon Room V1/V2 with UWB simulated from Vicon (4 anchors, 80 Hz, 0.05 m noise); VINS = VINS-Fusion stereo-inertial, BA = loop closure and global bundle adjustment; VIRAL fuses the VINS-Fusion estimate without BA; OSL baselines aligned by the initial ground-truth pose, VIRAL in the anchor frame; unit m (Nguyen et al., 2022b, Table II)

Translational RMSE，EuRoC MAV (simulated UWB) · V1_01

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 Nguyen et al., 2022b 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：均方根誤差（RMSE）；對齊方式：未對齊；單位：m；場景：indoor Vicon room

資料來源作者報告值(Nguyen et al., 2022b, Table II)

數值與出處
方法（原文寫法）報告值出處
Ours (w/o OSL)本方法原文提出0.0764 m(Nguyen et al., 2022b, Table II; Sec. VI-A)
Ours本方法原文提出0.0715 m(Nguyen et al., 2022b, Table II; Sec. VI-A)

Nguyen et al., 2022b · Table III 本方法 12 筆

指標Rotational RMSE

表格設定（擷取紀錄原文）：EuRoC Vicon Room V1/V2 with UWB simulated from Vicon (4 anchors, 80 Hz, 0.05 m noise); VINS = VINS-Fusion stereo-inertial, BA = loop closure and global bundle adjustment; VIRAL fuses the VINS-Fusion estimate without BA; OSL baselines aligned by the initial ground-truth pose, VIRAL in the anchor frame; unit degrees (Nguyen et al., 2022b, Table III)

Rotational RMSE，EuRoC MAV (simulated UWB) · V1_01

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 Nguyen et al., 2022b 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：均方根誤差（RMSE）；對齊方式：未對齊；單位：deg；場景：indoor Vicon room

資料來源作者報告值(Nguyen et al., 2022b, Table III)

數值與出處
方法（原文寫法）報告值出處
Ours (w/o OSL)本方法原文提出2.4428 deg(Nguyen et al., 2022b, Table III; Sec. VI-A)
Ours本方法原文提出2.3104 deg(Nguyen et al., 2022b, Table III; Sec. VI-A)

Nguyen et al., 2022b · Table VI 本方法 6 筆

指標Translational RMSE

表格設定（擷取紀錄原文）：Field flight tests (three flights; UAV moving in a vertical plane parallel to the inspected structure) with camera, VN-100 IMU, two Ouster OS1 LiDARs and Humatics P440 UWB with three self-localized anchors at 1 m height; VINS-Fusion and A-LOAM outputs fused as OSL; position ground truth from a Leica MS60 (periods of lost ground truth in datasets 01 and 02 excluded); all estimates aligned to ground truth by a rotation and translation fit [50]; orientation reference from the onboard AHRS; ORB-SLAM3 failed on these datasets and is absent (Nguyen et al., 2022b, Table VI)

Translational RMSE，VIRAL field flight tests (authors) · Test 01

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 Nguyen et al., 2022b 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：均方根誤差（RMSE）；對齊方式：SE(3) 剛體對齊；單位：m；場景：outdoor UAV inspection flights in a vertical plane parallel to a structure

資料來源作者報告值(Nguyen et al., 2022b, Table VI)

數值與出處
方法（原文寫法）報告值出處
Vertical LOAM14.2658 m(Nguyen et al., 2022b, Table VI; Sec. VI-C)
Horizontal LOAM1.594 m(Nguyen et al., 2022b, Table VI; Sec. VI-C)
VINS (w/o BA)1.6231 m(Nguyen et al., 2022b, Table VI; Sec. VI-C)
VINS (with BA)1.5915 m(Nguyen et al., 2022b, Table VI; Sec. VI-C)
VIRAL (w/o OSL)本方法原文提出0.1951 m(Nguyen et al., 2022b, Table VI; Sec. VI-C)
VIRAL (with OSL)本方法原文提出0.1593 m(Nguyen et al., 2022b, Table VI; Sec. VI-C)

其他比較組

列出其餘 1 個比較組

來源

  • Nguyen et al., 2022b

    Thien-Minh Nguyen, Muqing Cao, Shenghai Yuan, Yang Lyu, Thien Hoang Nguyen, Lihua Xie（2022）VIRAL-Fusion: A Visual-Inertial-Ranging-Lidar Sensor Fusion ApproachIEEE Transactions on Robotics, 38(2), pp. 958-977

    DOI 10.1109/tro.2021.3094157arXiv 2010.12274

    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