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

LI-Init

LI-Init 在 LiDAR 慣性里程計啟動前，自動判斷資料激勵是否足夠，並線上估計 LiDAR 與 IMU 的時間偏移、外參、重力向量與 IMU 偏差。時間偏移先以互相關粗估，再與旋轉外參聯合最佳化。作者強調若時間偏移未知，依賴 IMU 的運動畸變補償就無法正確執行。

感測、校正、同步與前處理近十年已讀全文同儕審查已出版誤差來源：校正、同步、畸變與退化

閱讀版本
arXiv 2202.11006v5 (2022-09-15), header 'Accepted for IROS 2022'; not the IEEE version of record
全文取得途徑
arXiv 預印本
本頁內容

Online initialization estimating LiDAR-IMU time offset, extrinsics, gravity and biases with excitation checking, feeding FAST-LIO2.

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3D LiDAR: Livox Avia (small FoV), Livox Mid360 (non-repetitive scanning), Hesai PandarXT (mechanical spinning); all set to 10 Hz、6-axis IMU Bosch BMI088, both inside the Livox LiDARs (factory hardware-synchronized) and inside a Pixhawk flight controller (unsynchronized with PandarXT); raw data 200 Hz
原文測試平台handheld
狀態估計LiDAR odometry with error-state iterated Kalman filter, then cross-correlation temporal alignment and joint temporal-spatial optimization
資料關聯LiDAR-only odometry uses scan-to-map point-to-plane residuals (modified from FAST-LIO2), chosen over NDT scan-to-scan so that non-repetitive LiDARs are supported; LiDAR and IMU data are associated by cross-correlation of angular-velocity magnitudes, then least-squares alignment of angular velocity and acceleration
時間表示Constant angular and linear velocity model between scans, with each input frame split into sub-frames to reduce model mismatch; IMU data linearly interpolated to LiDAR-odometry timestamps; the unknown constant time offset is estimated as an integer number of odometry intervals by cross-correlation and refined by a sub-interval residual using a constant angular acceleration approximation
去畸變During initialization, points are deskewed without the IMU: each point is projected to the scan-end frame using the constant-velocity prediction, because the unsynchronized IMU cannot yet be used; Fig. 3 compares Mid360 maps with and without this compensation; after initialization FAST-LIO2 uses the calibrated offset
迴圈閉合none
全域最佳化none
地圖表示Point map used for scan-to-map registration inside the FAST-LIO2-derived LiDAR odometry; the data structure is not described in this paper
先驗資訊none (initial extrinsic set to identity in tests)
可輸出幾何time offset, extrinsic, gravity, IMU bias as initial states for FAST-LIO2
計算需求Desktop Intel i7-10700 at 2.90 GHz with 32 GB RAM; LiDAR odometry about 8 ms per sub-frame; initialization solver below 500 ms once data are collected; 10.2 s total calibration on 40 s of PandarXT data versus 332.6 s for LI-Calib and 115.7 s for Target-Free

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 Avia方法輸入未標示small FoV, non-repetitive scanning; output set to 10 Hz; built-in BMI088 IMU hardware-synchronized in factory(Zhu et al., 2022b, Fig. 1, Sec. IV, Sec. IV-A, Table II to III)
LiDARLivox Mid360歸入：Livox MID-360方法輸入未標示non-repetitive scanning; output set to 10 Hz; built-in BMI088 IMU hardware-synchronized in factory(Zhu et al., 2022b, Fig. 1, Sec. IV, Sec. IV-A, Table II to III, Fig. 3, Fig. 8)
LiDARHesai PandarXT方法輸入未標示mechanical spinning; output set to 10 Hz; not synchronized with the Pixhawk IMU(Zhu et al., 2022b, Fig. 1, Sec. IV, Sec. IV-A to IV-C, Table III to IV, Fig. 9)
慣性量測單元（IMU）Bosch BMI088方法輸入未標示6-axis IMU inside both the Pixhawk flight controller and the Livox LiDARs; raw data 200 Hz(Zhu et al., 2022b, Sec. IV (setup))
運算硬體Intel i7-10700執行運算平台未標示desktop computer, 2.90 GHz, 32 GB RAM; all experiments(Zhu et al., 2022b, Sec. IV (setup))
其他Pixhawk flight controller方法輸入未標示houses the BMI088 IMU; fixed at two poses I1 and I2 with CAD relative pose of 0 deg rotation and 0.25 m translation (CAD accuracy stated as 0.01 deg and millimetres)(Zhu et al., 2022b, Fig. 1, Sec. IV, Sec. IV-B.1)

作者報告的優勢與限制

優勢

  • With artificially shifted IMU timestamps (0.05 to 0.5 s), time-offset RMSE about 1.6 to 3.4 ms on Livox sensors (Table II)
  • Extrinsic translation error about 0.6 to 1.3 cm mean and rotation error below 1 deg against CAD reference across three LiDARs (Table III)
  • On two PandarXT sequences (first 40 s): rotation error 0.6208 deg versus 1.0375 deg (LI-Calib) and 0.8483 deg (Target-Free); translation 0.0162 m versus diverged (LI-Calib) and 0.0187 m (Target-Free); 10.2 s versus 332.6 s and 115.7 s (Table IV)
  • FAST-LIO2 with the LI-Init time offset: end-to-end drift 0.0102 m over an 11.36 m handheld loop, versus 0.246 m with FAST-LIO2 internal synchronization (Sec. IV-A, Fig. 6)
  • No calibration target, extra sensor, hardware synchronization or stationary start required; an excitation assessment tells the user how to move the device (Sec. I, Sec. II, Sec. III-C.5)

限制

  • Assumes an unknown but constant time offset (Sec. I contributions, Sec. III-C.2)
  • Extrinsic error larger for the spinning PandarXT than the Livox units, because frame splitting shrinks the sub-frame FoV so the device must move slowly, lowering IMU excitation and SNR (Table III, Sec. IV-B.1)
  • Motion-based: needs sufficient excitation, judged by a singular-value threshold, and LiDAR odometry degeneration prevents data from being used (Sec. III-A, Sec. III-C.5)
  • Ground-truth time offset is unavailable for unsynchronized rigs; temporal accuracy is shown with artificial offsets on factory-synchronized Livox units and indirectly through LIO drift (Sec. IV-A)
  • Extrinsic accuracy is validated indirectly through a CAD-designed relative pose between two Pixhawk mounting poses; the baseline comparison uses only two PandarXT sequences (Sec. IV-B)

營建工程相關證據

未在工地測試；自組手持設備在建築量測常見，未同步時的時間偏移估計直接影響去畸變（推論）。

原文驗證環境：受控實驗

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 36 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。

Zhu et al., 2022b · Table II 本方法 18 筆

表格設定（擷取紀錄原文）：Temporal initialization with artificial IMU timestamp shifts on Livox LiDARs with built-in IMUs; 5 laboratory sequences per LiDAR; true offset about 0 s (factory sync) plus the added shift (Zhu et al., 2022b, Table II)

calibrated time offset Mean[s]，authors' handheld sequences (Livox Avia, built-in IMU) · artificial offset 0.05 s

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 Zhu et al., 2022b 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：未對齊；單位：s；場景：laboratory scene

數值與出處
方法（原文寫法）報告值出處
LI-Init (proposed)本方法原文提出0.049 s(Zhu et al., 2022b, Table II)

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

表格設定（擷取紀錄原文）：Extrinsic initialization: absolute error of the relative pose between two Pixhawk IMU mounting poses (from two LI-Init calibrations) against the CAD design (0 deg, 0.25 m); mean and SD over 5 sequences per LiDAR; initial extrinsic set to identity (Zhu et al., 2022b, Table III)

relative error Rot(deg), mean，authors' handheld sequences (Pixhawk IMU at poses I1 and I2) · 5 sequences, Livox Mid360

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 Zhu et al., 2022b 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：未對齊；單位：deg

數值與出處
方法（原文寫法）報告值出處
LI-Init (proposed)本方法原文提出0.2472 deg(Zhu et al., 2022b, Table III)

Zhu et al., 2022b · Table IV 本方法 3 筆

資料集與序列two PandarXT plus Pixhawk sequences from Sec. IV-B.1 · first 40 s of each sequence

表格設定（擷取紀錄原文）：Extrinsic calibration comparison on two Hesai PandarXT plus Pixhawk sequences, first 40 s (400 scans) each; average relative IMU pose error against CAD; time offset pre-compensated for the baselines; default baseline parameters; all on the i7-10700 desktop (Zhu et al., 2022b, Table IV)

Translation(m)，two PandarXT plus Pixhawk sequences from Sec. IV-B.1 · first 40 s of each sequence

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。

  • 發散

這些是 Zhu 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：平均值（mean）；對齊方式：未對齊；單位：m；場景：calibration scene of Fig. 7(e) (not described further)

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

數值與出處
方法（原文寫法）報告值出處
Proposed本方法原文提出0.0162 m(Zhu et al., 2022b, Table IV)
LI-Calib [14]無數值發散註記（擷取紀錄）：diverged (refinement fails; authors attribute it to missing gravity initialization)(Zhu et al., 2022b, Table IV and note)
Target-Free [15]0.0187 m(Zhu et al., 2022b, Table IV)

Zhu et al., 2022b · Text Sec.IV-C 本方法 2 筆

資料集與序列原文未報告

表格設定（擷取紀錄原文）：Runtime statements in the time consumption evaluation; desktop Intel i7-10700 (Zhu et al., 2022b, Text Sec.IV-C)

average processing time of a sub-frame (LiDAR odometry)，原文未報告

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 Zhu et al., 2022b 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

數值與出處
方法（原文寫法）報告值出處
LI-Init LiDAR odometry本方法原文提出硬體：desktop Intel i7-10700 @2.90 GHz, 32 GB RAM8 ms有附註註記（擷取紀錄）：other: stated as 'about 8 ms'(Zhu et al., 2022b, Sec. IV-C)

其他比較組

列出其餘 1 個比較組

來源

  • Zhu et al., 2022b

    Fangcheng Zhu, Yunfan Ren, Fu Zhang（2022）Robust Real-time LiDAR-inertial Initialization2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 3948-3955

    DOI 10.1109/iros47612.2022.9982225arXiv 2202.11006程式碼

    同儕審查已出版已讀全文近十年

程式碼：https://github.com/hku-mars/LiDAR_IMU_Init（授權：GPL-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。

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