FAST-LIO2
FAST-LIO2 延續 FAST-LIO 的緊耦合迭代卡爾曼濾波,但取消手工特徵擷取,直接以原始點對地圖中局部平面做點到平面(point-to-plane)配準,使系統較不依賴特定 LiDAR 掃描樣式。地圖以作者提出的增量式 k-d 樹(ikd-Tree)維護,支援逐點插入、刪除、樹上降採樣與平行重建,因而可在里程計頻率同步更新稠密點雲地圖。地圖只保留一個邊長 L 的立方體區域內的點,該區域初始以起點為中心,並在 LiDAR 偵測範圍觸及邊界時移動;系統不含迴圈偵測或全域修正。
本頁內容
Direct (feature-free) tightly-coupled iterated-Kalman LIO that registers raw points point-to-plane against a dense map kept in an incremental k-d tree (ikd-Tree), supporting spinning and solid-state LiDARs in real time 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 | 3D LiDAR (solid-state Livox Horizon/Avia and spinning Velodyne VLP-16/HDL-32E in the tested datasets)、IMU |
|---|---|
| 原文測試平台 | handheld、UAV、vehicle、wheeled UGV |
| 狀態估計 | tightly-coupled iterated Kalman filter on manifold (IKFOM toolbox) inherited from FAST-LIO, state includes LiDAR-IMU extrinsic (dimension 24) |
| 資料關聯 | direct: raw (downsampled) points registered without feature extraction; point-to-plane residual to a local plane fitted from 5 nearest map points found in ikd-Tree |
| 時間表示 | discrete poses with per-point back-propagation |
| 去畸變 | IMU forward/backward propagation per point (inherited from FAST-LIO) |
| 迴圈閉合 | none (authors state FAST-LIO2 is an odometry without loop detection or correction, Sec. VI-C) |
| 全域最佳化 | none |
| 地圖表示 | dense point map in an incremental k-d tree (ikd-Tree) with on-tree downsampling and box-wise deletion; the map region is a cube of side L initialized around the start position and moved when the LiDAR detection area reaches its border (default L = 1000 m) |
| 先驗資訊 | none |
| 可輸出幾何 | odometry and registered dense point map inserted at odometry rate; export format 原文未報告 in paper |
| 計算需求 | CPU real time; benchmark on DJI Manifold 2-C (1.8 GHz quad-core Intel i7-8550U, 8 GB RAM) and Khadas VIM3 ARM board (2.2 GHz quad-core Cortex-A73, 4 GB RAM); 1.82 ms (Intel) and 5.23 ms (ARM) mean total per scan on a 100 Hz handheld sequence; benchmark per-scan totals 11.47 to 31.56 ms on Intel with the 1000 m map |
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 | Livox Avia | 方法輸入 | 未標示 | solid-state, 70.4 deg (H) x 77.2 deg (V) circular FoV, non-repetitive scan pattern, built-in IMU; scan rate 100 Hz unless stated (10 Hz in aerial test) | (Xu et al., 2022, Sec. VII-A) |
| LiDAR | Livox Horizon | 資料集感測器 | LiLi-OM dataset (lili) | solid-state, non-repetitive, 81.7 deg (H) x 25.1 deg (V) FoV, 10 Hz | (Xu et al., 2022, Sec. VI; Table II) |
| LiDAR | Velodyne VLP-16 | 資料集感測器 | LIO-SAM dataset (liosam) | 16 lines, 10 Hz | (Xu et al., 2022, Sec. VI; Table II) |
| LiDAR | Velodyne HDL-32E | 資料集感測器 | UTBM robocar dataset (utbm) | two units at 10 Hz on a human-driven robocar (max 50 km/h); only the left LiDAR used | (Xu et al., 2022, Sec. VI; Table II) |
| LiDAR | Velodyne HDL-32E | 資料集感測器 | UrbanLoco HK (ulhk) | 10 Hz, human-driven vehicle | (Xu et al., 2022, Sec. VI; Table II) |
| LiDAR | Velodyne HDL-32E | 資料集感測器 | NCLT | 10 Hz, UGV | (Xu et al., 2022, Sec. VI; Table II) |
| 慣性量測單元(IMU) | BMI088 (built into Livox Avia) | 方法輸入 | 未標示 | built-in IMU of Livox Avia | (Xu et al., 2022, Sec. VII-A) |
| 慣性量測單元(IMU) | Xsens MTi-670 | 資料集感測器 | LiLi-OM dataset (lili) | 6-axis, 200 Hz | (Xu et al., 2022, Sec. VI; Table II) |
| 慣性量測單元(IMU) | MicroStrain 3DM-GX5-25 | 資料集感測器 | LIO-SAM dataset (liosam) | 9-axis, 1000 Hz | (Xu et al., 2022, Sec. VI; Table II) |
| 慣性量測單元(IMU) | Xsens MTi-28A53G25歸入:Xsens MTi-28 | 資料集感測器 | UTBM robocar dataset (utbm) | 6-axis, 100 Hz | (Xu et al., 2022, Sec. VI; Table II) |
| 慣性量測單元(IMU) | Xsens MTi-10 | 資料集感測器 | UrbanLoco HK (ulhk) | 9-axis, 100 Hz | (Xu et al., 2022, Sec. VI; Table II) |
| 慣性量測單元(IMU) | Microstrain MS25 | 資料集感測器 | NCLT | 9-axis, 50 Hz (interpolated to 100 Hz for LIO-SAM) | (Xu et al., 2022, Sec. VI; Table II) |
| GNSS 接收器 | UAV onboard GPS/IMU navigation (model not reported) | 參考或真值量測 | 未標示 | used only for UAV navigation, not by FAST-LIO2; trajectories compared visually, GPS trajectories not available for quantitative evaluation | (Xu et al., 2022, Sec. VII-C) |
| 載具平台 | 280 mm wheelbase quadrotor UAV | 方法輸入 | 未標示 | forward-looking Livox Avia, indoor aggressive flight | (Xu et al., 2022, Fig. 6(a); Sec. VII-A) |
| 載具平台 | handheld platform | 方法輸入 | 未標示 | Livox Avia with DJI Manifold 2-C | (Xu et al., 2022, Fig. 6(b); Sec. VII-A) |
| 載具平台 | 750 mm wheelbase quadrotor UAV (developed by Ambit-Geospatial) | 方法輸入 | 未標示 | down-facing Livox Avia, GPS-navigated waypoint flight | (Xu et al., 2022, Fig. 6(c); Sec. VII-A; Sec. VII-C) |
| 運算硬體 | DJI Manifold 2-C歸入:DJI Manifold 2C | 執行運算平台 | 未標示 | 1.8 GHz quad-core Intel i7-8550U CPU, 8 GB RAM | (Xu et al., 2022, Sec. VI-A) |
| 運算硬體 | Khadas VIM3 | 執行運算平台 | 未標示 | 2.2 GHz quad-core Cortex-A73 CPU, 4 GB RAM | (Xu et al., 2022, Sec. VI-A) |
作者報告的優勢與限制
優勢
- Removing feature extraction makes the system adaptable to LiDARs with different scan patterns (abstract; Sec. VIII)
- ikd-Tree gives the best overall kNN/insert/delete performance among compared dynamic structures (octree, R-tree, nanoflann) in 18 sequences (Sec. VI-B)
- Robust pose estimation in cluttered indoor scenes with rotation up to 1000 deg/s (abstract)
- Authors report FAST-LIO2 or a variant best in 18 of 19 benchmark sequences, the exception being ulhk 4 where LILI-OM is slightly better (Sec. VI-C1; Table IV)
- Total processing time about 8, 10 and 6 times lower than LILI-OM, LIO-SAM and LINS respectively (Sec. VI-D; Table VI)
限制
- Odometry only: no loop detection or correction (Sec. VI-C)
- Only map points within a moving cube of side L are retained, so map revisits beyond this region are not re-associated (Sec. V-A) (inference)
- In degenerate ENWIDE sequences FAST-LIO2 avoided divergence only where vegetation offered weak structure and still showed large drift (COIN-LIO, Sec. IV-C)
- Diverged at start when initialized in strong motion because gravity is estimated from averaged acceleration (Voxel-SLAM arXiv v1, Sec. X-A)
- Enlarging the map beyond 2000 m does not persistently improve accuracy because drift can cause false matches with old map points (Sec. VI-C1)
- On the ARM board the per-scan time occasionally exceeds the 10 ms sampling period at 100 Hz (Sec. VII-B1)
- Airborne mapping has no quantitative ground truth; GPS trajectories were unavailable and comparison was visual (Sec. VII-C)
營建工程相關證據
論文本身未在施工現場或以獨立幾何參考評估點雲;後續研究常以其為比較基準,例如 COIN-LIO 以其為基礎並在隧道等退化場景比較(Pfreundschuh et al., 2024),Voxel-SLAM 在動態起始條件下與之比較(Liu et al., 2026)。
原文驗證環境:公開基準、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 99 個比較組,合計 907 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 95 組列在最後,並連到性能比較頁。
Hu et al., 2025 · Table V 本方法 42 筆
表格設定(擷取紀錄原文):Map metrics and ATE for FAST-LIO2 and PALoc maps against TLS or high-precision ground-truth maps (FusionPortable MCR room sequences and MS-dataset parking lot); estimated map registered to the GT map by point-to-plane ICP (SE(3)), tau = 0.2 m, voxel 3.0 m, MME radius 0.1 m; values identical in arXiv v2 and the RA-L version of record (Hu et al., 2025, Table V)
ATE (computed with evo [5]; statistic not stated),FusionPortable · S5 (MCR slow)
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 Hu et al., 2025 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(Hu et al., 2025, Table V)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| FAST-LIO2 (FL2) [2]本方法 | 14.24 cm | (Hu et al., 2025, Table V) |
| PALoc [10] (loop closure and prior-map constraints) | 13.03 cm | (Hu et al., 2025, Table V) |
Xu et al., 2022 · Table VI 本方法 38 筆
指標average processing time per scan, Total
表格設定(擷取紀錄原文):Average total processing time per scan (odometry plus mapping) of FAST-LIO2 with 1000 m map; competitor Odo./Map. columns (LILI-OM, LIO-SAM, LINS) and map-size variants omitted for row cap (Xu et al., 2022, Table VI)
average processing time per scan, Total,LiLi-OM dataset (lili) · lili 6
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 Xu 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(Xu et al., 2022, Table VI)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| FAST-LIO2 (1000)本方法原文提出硬體:DJI Manifold 2-C (1.8 GHz quad-core Intel i7-8550U, 8 GB RAM) | 12.56 ms | (Xu et al., 2022, Table VI) |
| FAST-LIO2 (ARM)本方法原文提出硬體:Khadas VIM3 (2.2 GHz quad-core Cortex-A73, 4 GB RAM) | 45.58 ms | (Xu et al., 2022, Table VI) |
Hu et al., 2025 · Table VI 本方法 36 筆
表格設定(擷取紀錄原文):Map metrics for FAST-LIO2 and PALoc maps in larger scenes (FusionPortable corridor, canteen, escalator, building; Newer College math easy and parkland0); same registration and parameters as Table V; values identical in arXiv v2 and the RA-L version of record (Hu et al., 2025, Table VI)
AC (mean point-to-point error of correspondences within tau = 0.2 m),FusionPortable · S0 (corridor day)
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 Hu et al., 2025 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(Hu et al., 2025, Table VI)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| FAST-LIO2 (FL2) [2]本方法 | 7.53 cm | (Hu et al., 2025, Table VI) |
| PALoc [10] (loop closure and prior-map constraints) | 4.04 cm | (Hu et al., 2025, Table VI) |
Wu et al., 2024a · Table I 本方法 32 筆
表格設定(擷取紀錄原文):Default parameters for FAST-LIO2 and LIO-SAM, LIO-SAM loop closure disabled; one LIO-EKF configuration for all data; KITTI relative errors and ATE; LIO-SAM not run on Newer College (needs IMU attitude) (Wu et al., 2024a, Table I)
Avg. tra. (KITTI relative translation error),UrbanNav · 20210517
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 Wu et al., 2024a 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(Wu et al., 2024a, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| FAST-LIO2本方法 | 4.11% | (Wu et al., 2024a, Table I) |
| LIO-SAM | 3.18% | (Wu et al., 2024a, Table I) |
| LIO-EKF原文提出 | 3.2% | (Wu et al., 2024a, Table I) |
其他比較組
列出其餘 95 個比較組
- Lin & Zhang, 2024 · Table III
- Xu et al., 2022 · Table IV
- Lin & Zhang, 2022 · Table III
- Huang et al., 2024b · Table II
- Tang et al., 2026 · Table II
- Huang et al., 2024b · Table III
- Bai et al., 2022 · Table I
- Bai et al., 2022 · Table II
- Cao et al., 2025 · Table II (NTU VIRAL)
- Nguyen et al., 2023 · Table I
- Wei et al., 2025a · Table 6
- Zheng et al., 2025 · Table II
- Zhou et al., 2025 · Table I
- Jiao et al., 2022 · Table IV
- Pfreundschuh et al., 2024 · Table II
- Chen et al., 2024 · Table III
- Chen et al., 2025b · Table 5
- Xu et al., 2022 · Table V
- Xie et al., 2025 · Supp. Table 7
- Liu et al., 2026 · Table 2 (odometry without LC)
- Yuan et al., 2022 · Table II
- Wei et al., 2025b · Table V
- Huang et al., 2024b · Table IV
- Jung et al., 2023 · Table III
- Jung et al., 2023 · Table IV
- Zheng & Zhu, 2024 · Table III
- Lang et al., 2023 · Table III
- Koide et al., 2024 · Table V
- Jung et al., 2023 · Table II
- Cramariuc et al., 2023 · Table II
- Malladi et al., 2026 · Table I
- Zheng et al., 2022 · Table II
- Yan et al., 2026a · Table 1
- Pfreundschuh et al., 2024 · Table I
- Chen et al., 2023 · Table II
- Tang et al., 2023 · Table III
- Koide et al., 2024 · Table II
- Lee et al., 2025b · Table 7
- Cao et al., 2025 · Table III (GrandTour)
- Wei et al., 2025a · Table 5
- Ramezani et al., 2022 · Table II
- Yuan et al., 2022 · Table V
- Tang et al., 2026 · Table III
- He et al., 2023a · Table 6
- Wu et al., 2024b · Table I
- Chen et al., 2023 · Table I
- Chen et al., 2024 · Table II
- Wei et al., 2025b · Table II
- He et al., 2023a · Table 4
- Malladi et al., 2026 · Table III
- Malladi et al., 2026 · Table IV
- Schillberg et al., 2025 · Table 2
- Schillberg et al., 2025 · Table 3
- Burnett et al., 2025 · Table II
- Feng et al., 2025 · Table 3
- Feng et al., 2025 · Table 4
- Koide et al., 2024 · Table I
- He et al., 2023a · Table 5
- Malladi et al., 2026 · Table II
- Nguyen et al., 2023 · Table II
- Wu et al., 2024b · Table II
- Wu et al., 2024b · Table III
- Chen et al., 2024 · Table IV
- Wu et al., 2024b · Table IV
- Yan et al., 2026a · Table 2
- Lang et al., 2023 · Table IV
- Lang et al., 2023 · Table V
- Xu et al., 2022 · Text Sec.VII-C
- Tang et al., 2023 · Table I
- Tang et al., 2023 · Table II
- Lee et al., 2024b · Table 4
- Wu et al., 2024a · Table II
- He et al., 2023a · Table 1
- Nguyen et al., 2023 · Table III
- Xu et al., 2022 · Table VII
- Liu et al., 2023b · Table VI
- He et al., 2023a · Table 2
- Lin & Zhang, 2024 · Text Sec.VI-E
- Ramezani et al., 2022 · Text Sec.VI-C
- Yuan et al., 2022 · Table III
- Liu et al., 2023a · Supplementary Table V
- Xu et al., 2022 · Text Sec.VII-B1
- Xu et al., 2022 · Text Sec.VII-B2
- Xu et al., 2022 · Text Sec.VII-B3
- Ghadimzadeh Alamdari et al., 2025 · Table 3
- Ghadimzadeh Alamdari et al., 2025 · Text Sec.7.1.2
- Koide et al., 2024 · Table X
- Liu et al., 2023b · Table IV
- Lee et al., 2025b · Table 8 (Total column)
- He et al., 2023a · Table 7
- Cao et al., 2025 · Text Sec. V-C (R-Campus)
- Burnett et al., 2025 · Text Sec.V-B
- Yan et al., 2026a · Table 3
- Yan et al., 2026a · Table 4
- Zhu et al., 2022b · Text Sec.IV-A
來源
Xu et al., 2022
(2022)FAST-LIO2: Fast Direct LiDAR-Inertial OdometryIEEE Transactions on Robotics, 38(4):2053-2073
DOI 10.1109/tro.2022.3141876arXiv 2107.06829程式碼
同儕審查已出版已讀全文近十年查證後修正
相關版本
- 預印本:FAST-LIO2: Fast Direct LiDAR-inertial Odometry (arXiv v1) https://arxiv.org/abs/2107.06829
- 程式碼釋出:hku-mars/FAST_LIO (FAST-LIO2) and hku-mars/ikd-Tree https://github.com/hku-mars/FAST_LIO
程式碼:https://github.com/hku-mars/FAST_LIO(授權:GPL-2.0 (LICENSE file checked); ikd-Tree repository 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。