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

Faster-LIO

Faster-LIO 以 FAST-LIO2 為基礎，將 ikd-Tree 換成增量式稀疏體素（iVox），以雜湊表與 LRU 快取管理體素，並以近似 k 近鄰查詢取代嚴格 k 近鄰，以換取大幅加速。作者提出線性與偽希爾伯特曲線（PHC）兩種體素內結構，並指出近似鄰點雖會帶來不精確，但在 LIO 配準中影響不大。其貢獻集中在資料結構效率，而非估計架構或全域一致性。

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
Version of record, IEEE RA-L 7(2):4861-4868 (open access, CC BY 4.0): IEEE Xplore full-text HTML plus the identical published-layout PDF posted in the authors' GitHub repository (used for the table images)
全文取得途徑
出版者開放取用（open access）
本頁內容

Replaces FAST-LIO2's ikd-Tree with incremental sparse voxels (iVox) supporting incremental insertion and parallel approximate k-NN, greatly increasing LIO throughput at similar accuracy.

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 (solid-state LiDAR in the AVIA dataset from FastLIO2 and 32-line spinning LiDARs; sensor models not named in the paper)、IMU
原文測試平台not described in the paper; evaluation on public datasets AVIA (from FastLIO2), NCLT, ULHK, UTBM and liosam_1; the LeGO-LOAM comparison uses data gathered from a ground vehicle (Sec. V; Sec. V-D)
狀態估計iterated EKF pipeline inherited from FAST-LIO2 (paper states Faster-LIO is developed under FastLIO2 with code refactoring)
資料關聯point-to-plane residuals using approximate k-NN from incremental sparse voxels
時間表示discrete poses (as FAST-LIO2)
去畸變preprocessing plus undistortion inherited from FAST-LIO2 (Table I footnote)
迴圈閉合none
全域最佳化none
地圖表示incremental sparse voxels (iVox) in a hash map with LRU cache; linear or pseudo-Hilbert-curve (PHC) in-voxel layout
先驗資訊none
可輸出幾何odometry and point-cloud map; export format 原文未報告
計算需求CPU with parallel k-NN, threads set to the CPU maximum: AMD R7-5800X (8 cores, 3.8 GHz) and Intel Xeon Gold 5128 (16 cores, 2.3 GHz; Fig. 1 caption says 5218) desktops, plus an Intel i7-10750H laptop for accuracy runs; 1000-2000 Hz per scan for solid-state and over 200 Hz for 32-line spinning LiDAR (abstract)

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
LiDARspinning LiDAR (model not named)資料集感測器NCLT; UTBM robocar datasetabstract reports over 200 Hz for 32-line spinning lidars; line count and model per dataset not stated(Bai et al., 2022, Abstract; Sec. V-B)
LiDARsolid-state LiDAR (model not named)資料集感測器AVIAAVIA dataset from FastLIO2(Bai et al., 2022, Sec. V)
GNSS 接收器RTK (model not reported)參考或真值量測NCLT; UTBM robocar datasetreference ground-truth trajectories of the selected sequences(Bai et al., 2022, Sec. V-C; Fig. 11)
運算硬體AMD R7-5800X執行運算平台未標示8 cores, 3.8 GHz desktop ('iVox AMD')(Bai et al., 2022, Sec. V; Fig. 1)
運算硬體Intel Xeon Gold 5128執行運算平台未標示16 cores, 2.3 GHz desktop ('iVox Intel'); Fig. 1 caption gives Xeon Gold 5218(Bai et al., 2022, Sec. V; Fig. 1)
運算硬體Intel i7-10750H執行運算平台未標示laptop CPU, 6 cores, boost 5 GHz; used for accuracy experiments(Bai et al., 2022, Sec. V)

作者報告的優勢與限制

優勢

  • Per-scan rate of 1000-2000 Hz (solid-state) and over 200 Hz (32-line) at the same level of accuracy (abstract; Sec. VI)
  • iVox can be integrated into other SLAM systems such as LeGO-LOAM mapping to reduce mapping time (Sec. V-D)
  • Speed increase against FastLIO2 of 1.52 to 3.49 times across nine sequences, and lower pose-computation (opt) time than FastLIO2 on all nine sequences, e.g. utbm_2 5.47 ms versus 19.35 ms (Table I)
  • In LeGO-LOAM mapping, iVox reduced total per-scan runtime from 69.41 ms to 51.706 ms, mainly by removing keyframe extraction (Table III)

限制

  • Nearest neighbours are approximate; authors note iVox is not a good choice when exact neighbours are required (Sec. V-A)
  • No loop closure detection (author-stated, Sec. V-C; loop closure of LIO-SAM and LiLi-OM was disabled for comparison)
  • Linear Faster-LIO has higher APE than FastLIO2 on all nine sequences in Table II (e.g. liosam_1 1.78 m versus 0.83 m) while RPE is similar; authors describe accuracy as comparable (Table II; Sec. V-C; table-derived)
  • k-NN time of iVox grows faster than tree structures as points per voxel increase, so small to medium local maps are preferred (Sec. V-A)

營建工程相關證據

原文未報告

原文驗證環境：公開基準

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 15 個比較組，合計 190 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 11 組列在最後，並連到性能比較頁。

Bai et al., 2022 · Table I 本方法 45 筆

表格設定（擷取紀錄原文）：Time evaluation: 'pre' = preprocessing + undistortion + downsampling per scan, 'opt' = pose computation; Spd inc = speed increase against FastLIO2; LIO-SAM and LiLi-OM columns omitted because they are measured per keyframe on distributed ROS nodes (Table I footnote) (Bai et al., 2022, Table I)

pre (ms): preprocessing+undistortion+downsampling per scan，NCLT · nclt_2

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 Bai 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：平均值（mean）；對齊方式：不適用；單位：ms；場景：NCLT public dataset (scene and platform not described in the paper)

資料來源作者報告值(Bai et al., 2022, Table I)

數值與出處
方法（原文寫法）報告值出處
Faster-LIO本方法原文提出硬體：not stated for Table I (paper platforms: AMD R7-5800X, Intel Xeon Gold 5128, Intel i7-10750H laptop)2.74 ms(Bai et al., 2022, Table I)
Faster-LIO PHC本方法原文提出硬體：not stated for Table I (paper platforms: AMD R7-5800X, Intel Xeon Gold 5128, Intel i7-10750H laptop)0.52 ms(Bai et al., 2022, Table I)
FastLIO2硬體：not stated for Table I (paper platforms: AMD R7-5800X, Intel Xeon Gold 5128, Intel i7-10750H laptop)2.73 ms(Bai et al., 2022, Table I)

Bai et al., 2022 · Table II 本方法 36 筆

表格設定（擷取紀錄原文）：Accuracy in APE (m) over whole trajectories and translational RPE (%) per 100 m; loop closure of LIO-SAM and LiLi-OM disabled; parameters of LIO-SAM and LiLi-OM not adjusted; reference mostly RTK (Bai et al., 2022, Table II)

APE (m)，NCLT · nclt_2 (0.26 km)

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 Bai 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：原文未報告；對齊方式：原文未報告；單位：m；場景：NCLT public dataset (scene and platform not described in the paper)

資料來源作者報告值(Bai et al., 2022, Table II)

數值與出處
方法（原文寫法）報告值出處
Faster-LIO本方法原文提出0.94 m(Bai et al., 2022, Table II)
Faster-LIO PHC本方法原文提出1.03 m(Bai et al., 2022, Table II)
FastLIO20.91 m(Bai et al., 2022, Table II)
LIO-SAM1.11 m(Bai et al., 2022, Table II)
LiLi-OM無數值失敗註記（擷取紀錄）：'-' in table (Table I footnote: failed due to large drift or lack of necessary input)(Bai et al., 2022, Table II)

Lim et al., 2023 · Table II 本方法 28 筆

表格設定（擷取紀錄原文）：HILTI-Oxford validation sequences with millimetre-level marker poses; each marker scored 10, 6 or 3 if the closest estimated pose is within 1, 10 or 100 cm; x = trajectory diverged (Lim et al., 2023, Table II)

number of markers within 1 cm，HILTI-Oxford dataset (HILTI SLAM Challenge 2022 validation set) · Exp01

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 Lim 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：count；場景：construction-site building interiors, handheld

資料來源作者報告值(Lim et al., 2023, Table II)

數值與出處
方法（原文寫法）報告值出處
Faster-LIO [17]本方法0 count(Lim et al., 2023, Table II)
AdaLIO (Ours)原文提出0 count(Lim et al., 2023, Table II)

Chen et al., 2024 · Table III 本方法 16 筆

表格設定（擷取紀錄原文）：Absolute pose error (RMSE, m); identical iG-LIO parameters for all sequences; BG sequences evaluated with origin alignment, others with SE(3) alignment; '*' marks Livox avia sequences (Chen et al., 2024, Table III)

Absolute pose error (RMSE, meters)，NCLT · nclt_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 Chen et al., 2024 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：campus (Velodyne HDL-32E)

資料來源作者報告值(Chen et al., 2024, Table III)

數值與出處
方法（原文寫法）報告值出處
iG-LIO原文提出1.673 m(Chen et al., 2024, Table III)
iG-LIO* (kd-tree surface covariance variant, ablation)1.795 m(Chen et al., 2024, Table III)
NDT-LIO (ablation)2.365 m(Chen et al., 2024, Table III)
Faster-LIO本方法1.855 m(Chen et al., 2024, Table III)
FastLIO21.734 m(Chen et al., 2024, Table III)
DLIO2.104 m(Chen et al., 2024, Table III)

其他比較組

列出其餘 11 個比較組
  • Liu et al., 2026 · Table 2 (odometry without LC)本方法 13 筆；Hilti handheld sequence exp01-construction (name per Table C1)、Hilti handheld sequence exp02-construction (name per Table C1)、Hilti handheld sequence exp03-construction (name per Table C1)、Hilti handheld sequence exp07-long-corridor (name per Table C1)、Hilti handheld sequence exp09-cupola (name per Table C1)、Hilti handheld sequence exp11-lower-gallery (name per Table C1)、Hilti handheld sequence exp15-upper-gallery (name per Table C1)、Hilti handheld sequence exp21-outside (name per Table C1)、Hilti handheld sequence site1-handheld-1 (name per Table C1)、Hilti handheld sequence site1-handheld-2 (name per Table C1)、Hilti handheld sequence site1-handheld-3 (name per Table C1)、Hilti handheld sequence site1-handheld-4 (name per Table C1)、Hilti handheld sequence site1-handheld-5 (name per Table C1)
  • Yuan et al., 2022 · Table V本方法 8 筆；authors' Livox Avia datasets
  • Wu et al., 2024b · Table I本方法 7 筆；M2DGR
  • Chen et al., 2024 · Table II本方法 6 筆；AVIA、BG、NCD、NCLT、ULHK
  • Feng et al., 2025 · Table 3本方法 5 筆；Feng et al. simulated construction-site dataset (Gazebo)
  • Feng et al., 2025 · Table 4本方法 5 筆；Feng et al. real construction-site dataset (Xi'an hospital)
  • Wu et al., 2024b · Table II本方法 5 筆；VoxelMap++ own datasets
  • Wu et al., 2024b · Table III本方法 5 筆；VoxelMap++ own datasets
  • Chen et al., 2024 · Table IV本方法 4 筆；AVIA、GDUT (self-collected)
  • Wu et al., 2024b · Table IV本方法 4 筆；VoxelMap++ own datasets
  • Lee et al., 2024b · Table 4本方法 3 筆；ConSLAM、HeLiPR、NTU VIRAL

來源

  • Bai et al., 2022

    Chunge Bai, Tao Xiao, Yajie Chen, Haoqian Wang, Fang Zhang, Xiang Gao（2022）Faster-LIO: Lightweight Tightly Coupled Lidar-Inertial Odometry Using Parallel Sparse Incremental VoxelsIEEE Robotics and Automation Letters, 7(2):4861-4868

    DOI 10.1109/lra.2022.3152830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/gaoxiang12/faster-lio（授權：GPL-2.0 (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