SLAM 點雲圖鑑Point clouds for construction
M-01 · 2023

FF-LINS

FF-LINS 認為把掃描配準到自建全域地圖（frame-to-map）會讓 LiDAR 慣性估計器把原本不可觀的全域偏航與位置錯誤地當成可觀，造成不一致。它採以 INS 為中心的架構：先以 INS 機械編排的高頻位姿去畸變並選取關鍵影格，再把兩關鍵影格間所有非重複掃描的影格累積成較稠密的關鍵影格點雲地圖；最新關鍵影格的點直接與滑動視窗內其他關鍵影格地圖做點到平面關聯，形成相對位姿約束，與 IMU 預積分一起在因子圖中最佳化，並線上估計 LiDAR 與 IMU 外參及時間延遲。

里程計＋局部建圖近十年已讀全文同儕審查已出版當代 LiDAR（慣性）里程計與 SLAM

閱讀版本
arXiv v1 (2023-07-13; only version); RA-L version of record not read
全文取得途徑
arXiv 預印本
本頁內容

Consistent solid-state LiDAR-inertial estimator that replaces frame-to-map registration with direct frame-to-frame point-to-plane factors between the newest keyframe and INS-accumulated keyframe point-cloud maps in a sliding-window factor graph, with online LiDAR-IMU extrinsic and time-delay calibration.

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solid-state non-repetitive LiDAR (Livox Mid-70 in the Robot dataset; Livox Horizon and Livox AVIA in public datasets)、MEMS IMU (ADI ADIS16465 in the Robot dataset; built-in Livox IMUs in public datasets)
原文測試平台wheeled UGV (low-speed robot, about 1.5 m/s)、原文未報告 (LiLi-OM and R3LIVE public datasets; carrier not described in this paper)
狀態估計sliding-window factor graph (10 IMU preintegration intervals) solved with Levenberg-Marquardt in Ceres, tightly coupling LiDAR frame-to-frame point-to-plane factors with IMU preintegration and marginalization prior; LiDAR-IMU extrinsics and time delay estimated online; INS-centric update only at LiDAR keyframes (Sec. II, III-C)
資料關聯direct frame-to-frame: each point of the newest keyframe is projected into the accumulated keyframe point-cloud maps of the other keyframes in the window; plane fitted to 5 nearest points and accepted if all lie within 0.1 m; Huber loss plus chi-square rejection between two optimizations (Sec. III-B, III-C-4)
時間表示discrete keyframe states with IMU preintegration and an estimated LiDAR-IMU time delay (Sec. III-C-1)
去畸變interpolated INS poses from mechanization undistort each frame before 0.5 m voxel downsampling (Sec. III-A-2)
迴圈閉合none (authors state it could be added for large-scale mapping, Sec. V)
全域最佳化none
地圖表示per-keyframe point-cloud maps accumulated from all frames since the previous keyframe with INS poses, voxel-downsampled at 0.5 m; no global map is used for state estimation (Sec. III-A-3)
先驗資訊INS initialization with zero position and yaw, roll and pitch from accelerometers; extrinsics and time delay assumed uncalibrated and estimated online (Sec. II, IV-A)
可輸出幾何continuous INS-rate poses between keyframes, keyframe states, online LiDAR-IMU extrinsics and time delay; keyframe point clouds
計算需求real time on a desktop AMD R7-3700X: about 0.6 ms preprocessing per frame, 2.6 to 2.7 ms frame-to-frame association and 32.6 to 45.6 ms factor-graph optimization per keyframe (keyframes every 300 to 370 ms); runs at 4 to 6 times real-time speed (Table IV, Sec. IV-D)

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
LiDARLivox Mid-70歸入：Livox MID70方法輸入FF-LINS Robot datasetsolid-state LiDAR, 10 Hz; described as containing only one scanning line(Tang et al., 2023, Sec. IV-A; Fig. 5)
LiDARLivox Horizon資料集感測器LiLi-OM datasetwith built-in IMU(Tang et al., 2023, Sec. IV-A)
LiDARLivox AVIA歸入：Livox Avia資料集感測器R3LIVE datasetwith built-in IMU(Tang et al., 2023, Sec. IV-A)
慣性量測單元（IMU）ADI ADIS16465方法輸入FF-LINS Robot datasetindustrial-grade MEMS IMU, gyroscope bias instability 2 deg/hr, 200 Hz; hardware-triggered synchronization with the LiDAR(Tang et al., 2023, Sec. IV-A)
GNSS 接收器GNSS/INS integrated navigation system with GNSS-RTK and a navigation-grade IMU (models not stated)參考或真值量測FF-LINS Robot datasetpost-processed ground truth, 0.02 m position and 0.01 deg attitude(Tang et al., 2023, Sec. IV-A)
載具平台low-speed wheeled robot方法輸入FF-LINS Robot datasetaverage speed around 1.5 m/s(Tang et al., 2023, Sec. IV-A; Fig. 5)
運算硬體AMD R7-3700X執行運算平台未標示desktop PC, ROS, multi-threaded(Tang et al., 2023, Sec. IV-A; Sec. IV-D)

作者報告的優勢與限制

優勢

  • On the Robot dataset FF-LINS had lower ARE and ATE than FAST-LIO2 on all four sequences, e.g. campus 0.41 deg and 1.51 m versus 3.55 deg and 4.42 m (Table III)
  • Yaw standard deviation grows over time as expected for an unobservable state, whereas FAST-LIO2's does not, which the authors present as evidence of consistency (Fig. 1)
  • Online extrinsic and time-delay calibration converged and strongly improved accuracy: R3LIVE end-to-end errors 1.20, 2.41 and 2.51 m with calibration versus 12.18, 14.16 and 17.14 m without (Table II, Fig. 7)
  • Succeeded in narrow indoor passages of hku_main_building where LIO-SAM failed (Sec. IV-B-2)
  • Runs 4 to 6 times faster than real time on a desktop CPU (Sec. IV-D)

限制

  • FAST-LIO2 obtained a much lower end-to-end error on hkust_campus_01 (0.14 m versus 2.51 m), which the authors attribute to FAST-LIO2 matching its own earlier map (Table II)
  • LiLi-OM dataset evaluation uses meter-level GPS start and end points, which the authors call inaccurate and only qualitative (Sec. IV-B-1)
  • Measurement covariance was set offline from error statistics with ground-truth poses (sigma about 0.1 m) (Sec. III-C-2)
  • Odometry only without loop closure; designed around non-repetitive solid-state LiDARs, although authors claim applicability to others (Sec. V)

營建工程相關證據

論文未涉及營建場域；自建資料是約 1.5 m/s 低速輪式機器人在校園、建物周邊、操場與公園的序列，參考為 GNSS-RTK 加導航級 IMU 後處理軌跡（0.02 m、0.01 deg）。PA-LVIO （Tang et al., 2026）在 i2Nav-Robot 與 MARS-LVIG 資料上以 FF-LINS 為比較基準。其一致性估計與線上外參及時間延遲校正，對需要再融合 GNSS 或 BIM 等絕對資訊的工地機器人有參考價值（推論）。

原文驗證環境：公開基準、獨立參考量測、跨場域

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 37 筆紀錄。

Tang et al., 2023 · Table III 本方法 16 筆

表格設定（擷取紀錄原文）：Private Robot dataset (Livox Mid-70, ADIS16465) with post-processed GNSS-RTK/INS ground truth; ARE and ATE; LiLi-OM and LIO-SAM could not be run (Tang et al., 2023, Table III)

ATE，FF-LINS Robot dataset · campus (1.33 km, 934 s)

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 Tang et al., 2023 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：原文未報告；對齊方式：原文未報告；單位：m；場景：campus, building surroundings, playground and park with pedestrians, bicycles and vehicles; wheeled robot

資料來源作者報告值(Tang et al., 2023, Table III)

數值與出處
方法（原文寫法）報告值出處
FAST-LIO24.42 m(Tang et al., 2023, Table III)
FF-LINS-WO (without online calibration)本方法2.17 m(Tang et al., 2023, Table III)
FF-LINS本方法原文提出1.51 m(Tang et al., 2023, Table III)

Tang et al., 2023 · Table IV 本方法 12 筆

表格設定（擷取紀錄原文）：Average running times of FF-LINS on the Robot dataset; frame preprocessing about 0.6 ms per frame (text) (Tang et al., 2023, Table IV)

Keyframe interval，FF-LINS Robot dataset · campus

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 Tang et al., 2023 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。

統計量：平均值（mean）；對齊方式：不適用；單位：ms；場景：per keyframe

數值與出處
方法（原文寫法）報告值出處
FF-LINS本方法原文提出硬體：AMD R7-3700X desktop300 ms(Tang et al., 2023, Table IV)

Tang et al., 2023 · Table II 本方法 6 筆

指標end-to-end error

表格設定（擷取紀錄原文）：R3LIVE dataset (Livox AVIA) end-to-end errors; LiLi-OM could not be run; FF-LINS-WO disables online extrinsic and time-delay calibration (Tang et al., 2023, Table II)

end-to-end error，R3LIVE dataset · hku_main_building

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 Tang et al., 2023 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：原文未報告；對齊方式：未對齊；單位：m；場景：HKU main building narrow indoor passages and HKUST campus

資料來源作者報告值(Tang et al., 2023, Table II)

數值與出處
方法（原文寫法）報告值出處
LIO-SAM無數值失敗註記（擷取紀錄）：failed(Tang et al., 2023, Table II)
FAST-LIO22.5 m(Tang et al., 2023, Table II)
FF-LINS-WO (without online calibration)本方法12.18 m(Tang et al., 2023, Table II)
FF-LINS本方法原文提出1.2 m(Tang et al., 2023, Table II)

Tang et al., 2023 · Table I 本方法 3 筆

指標distance error of starting-ending distance versus GPS

表格設定（擷取紀錄原文）：LiLi-OM dataset without ground truth: error of the estimated starting-ending distance relative to meter-level GPS start and end fixes; LiLi-OM and LIO-SAM without loop closure; extrinsics and time delay treated as uncalibrated (Tang et al., 2023, Table I)

distance error of starting-ending distance versus GPS，LiLi-OM dataset · Schloss-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 Tang et al., 2023 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：原文未報告；對齊方式：未對齊；單位：m；場景：outdoor sequences (Schloss and East)

資料來源作者報告值(Tang et al., 2023, Table I)

數值與出處
方法（原文寫法）報告值出處
LiLi-OM1.36 m(Tang et al., 2023, Table I)
LIO-SAM0.47 m(Tang et al., 2023, Table I)
FAST_LIO21.1 m(Tang et al., 2023, Table I)
FF-LINS本方法原文提出0.23 m(Tang et al., 2023, Table I)

來源

  • Tang et al., 2023

    Hailiang Tang, Tisheng Zhang, Xiaoji Niu, Liqiang Wang, Linfu Wei, Jingnan Liu（2023）FF-LINS: A Consistent Frame-to-Frame Solid-State-LiDAR-Inertial State EstimatorIEEE Robotics and Automation Letters, 8(12):8525-8532

    DOI 10.1109/lra.2023.3329625arXiv 2307.06632程式碼

    同儕審查已出版已讀全文近十年查證後修正

程式碼：https://github.com/i2Nav-WHU/FF-LINS（授權：GPL-3.0 (LICENSE file checked)）。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。

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