FAST-LIO
FAST-LIO 以緊耦合迭代擴展卡爾曼濾波(iterated extended Kalman filter, iEKF)融合 LiDAR 特徵點與 IMU,並以 IMU 前向傳播與反向傳播(back-propagation)將掃描內每個點補償到掃描結束時刻,以處理運動畸變。作者提出與傳統等價、但計算量取決於狀態維度而非量測維度的卡爾曼增益公式,使大量特徵點可在機載電腦上即時融合。其前端仍沿用 LOAM 式邊緣與平面特徵,地圖為特徵點集合,無迴圈偵測。
本頁內容
A tightly-coupled iEKF LiDAR-inertial odometry with IMU-based back-propagation deskewing and a Kalman-gain formula whose cost scales with state dimension, enabling real-time fusion of many feature points on a small UAV computer.
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 Avia; Velodyne VLP-16 in LINS data)、IMU (model on the authors' rig not reported; Xsens MTiG-710 in LINS data) |
|---|---|
| 原文測試平台 | UAV、handheld |
| 狀態估計 | tightly-coupled iterated extended Kalman filter on manifold, with an equivalent Kalman-gain formula whose matrix inversion scales with state dimension rather than measurement dimension |
| 資料關聯 | LOAM-style edge and planar feature points; point-to-edge / point-to-plane residuals to nearest map features found with a k-d tree |
| 時間表示 | discrete poses (scan-end state) with per-point back-propagation |
| 去畸變 | IMU forward propagation plus backward propagation that projects each feature point to the scan-end time |
| 迴圈閉合 | none |
| 全域最佳化 | none |
| 地圖表示 | accumulated feature-point map (edge and plane points) organized by k-d tree |
| 先驗資訊 | none |
| 可輸出幾何 | odometry and registered feature-point map; raw-point maps shown in figures; export format 原文未報告 |
| 計算需求 | real-time CPU on DJI Manifold 2-C (Intel i7-8550U); all iEKF iterations within 25 ms with >1,200 effective features (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 |
|---|---|---|---|---|---|
| LiDAR | Livox AVIA歸入:Livox Avia | 方法輸入 | 未標示 | solid-state, 70 deg FoV, installed directly on the airframe | (Xu & Zhang, 2021, Fig. 1; Sec. IV-B) |
| LiDAR | Velodyne VLP-16 | 資料集感測器 | LINS dataset (seaport) | 原文未報告 | (Xu & Zhang, 2021, Sec. IV-D; Fig. 7) |
| 慣性量測單元(IMU) | IMU (model not reported) | 方法輸入 | 未標示 | rigidly attached to the LiDAR with known extrinsic | (Xu & Zhang, 2021, Sec. III-B-2) |
| 慣性量測單元(IMU) | Xsens MTiG-710歸入:Xsens MTi-G-710 | 資料集感測器 | LINS dataset (seaport) | 原文未報告 | (Xu & Zhang, 2021, Sec. IV-D; Fig. 7) |
| 載具平台 | customized small-scale quadrotor UAV | 方法輸入 | 未標示 | 280 mm wheelbase | (Xu & Zhang, 2021, Fig. 1; Sec. IV-B) |
| 載具平台 | handheld sensor suite (same LiDAR platform) | 方法輸入 | 未標示 | 原文未報告 | (Xu & Zhang, 2021, Sec. IV-C; Sec. IV-D; Fig. 6) |
| 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 & Zhang, 2021, Fig. 1; Sec. IV-B; Sec. IV-D) |
作者報告的優勢與限制
優勢
- Kalman gain computed with complexity tied to state dimension, lowering cost when many LiDAR points are fused (abstract; Sec. III-C)
- Stable odometry under hand-held shaking with angular rates often above 100 deg/s where compared LOAM variants degraded (Sec. IV-C)
- New Kalman gain formula took 1.16 ms versus 1621 ms for the conventional formula at 1802 feature points (Table II)
- On the LINS seaport data FAST-LIO used 7.3 ms per scan versus 34.5 ms for LINS while keeping 784 instead of 147 feature points, with better mapping accuracy shown qualitatively (Sec. IV-D; Fig. 7)
限制
- Front end depends on hand-engineered edge/plane feature extraction (Sec. III-A); the successor FAST-LIO2 motivates removing it because feature extraction depends on LiDAR scan pattern (FAST-LIO2 Sec. I, VIII)
- No loop closure or global optimization; drift is reported only as start-end return error (Sec. IV) (inference: global consistency of maps not addressed)
營建工程相關證據
作者以手持方式繞行香港大學主樓(既有建築外部)約 140 m 後回到起點,以起訖點差估計漂移;未於施工中工地或以獨立參考量測評估點雲幾何。
原文驗證環境:受控實驗、已完工建築
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 12 個比較組,合計 88 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 8 組列在最後,並連到性能比較頁。
Reinke et al., 2022 · Table III 本方法 30 筆
表格設定(擷取紀錄原文):Underground datasets A, C, F, H, I, J (Table I); LOCUS 2.0 versus FAST-LIO and LINS; column labels reproduced as printed (APE max [m], APE mean [%], CPU [%] max and mean, max memory [GB]); many printed 'max' values are below 'mean' values; ground truth from LOCUS 1.0 against survey-grade maps (Reinke et al., 2022, Table III)
APE max [m],NeBula odometry dataset (DARPA SubT, Team CoSTAR) · A: power plant, Elma WA (urban), Husky, 631.53 m
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 Reinke 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(Reinke et al., 2022, Table III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| LOCUS 2.0原文提出 | 0.19 m | (Reinke et al., 2022, Table III) |
| FAST-LIO本方法 | 0.79 m | (Reinke et al., 2022, Table III) |
| LINS | 0.43 m | (Reinke et al., 2022, Table III) |
Lin et al., 2021 · Table I 本方法 24 筆
表格設定(擷取紀錄原文):Version of record Table I: relative rotation error (RRE, deg) and relative translation error (RTE, %) over all sub-sequences of each length, two fast-rotating handheld sequences (130 and 200 deg/s, mapping to (a) and (b) not stated) with D-GPS RTK ground truth; rows R2LIVE-LIO, R2LIVE-VIO and R2LIVE-LC (the latter undefined in the text) omitted here. Supersedes the median values in the arXiv v1 Fig. 11 caption. (Lin et al., 2021, Table I)
RRE (deg) over 50 m sub-sequences,R2LIVE Experiment-4 (authors' data, D-GPS RTK) · Experiment-4 (a), 50 m sub-sequences
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 Lin et al., 2021 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(Lin et al., 2021, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| VINS-Mono | 0.24 deg | (Lin et al., 2021, VoR Table I) |
| Fast-Lio本方法 | 0.34 deg | (Lin et al., 2021, VoR Table I) |
| Camvox | 0.67 deg | (Lin et al., 2021, VoR Table I) |
| R2LIVE原文提出 | 0.25 deg | (Lin et al., 2021, VoR Table I) |
Xu & Zhang, 2021 · Table II 本方法 12 筆
指標Kalman gain computation time
表格設定(擷取紀錄原文):Running time of the Kalman gain computation with the conventional versus the proposed formula, same pipeline and number of feature points (Xu & Zhang, 2021, Table II)
Kalman gain computation time,原文未報告 (own data) · feature number = 307
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 & Zhang, 2021 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 & Zhang, 2021, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Old Formula (conventional Kalman gain)本方法 | 7.1 ms | (Xu & Zhang, 2021, Table II) |
| New Formula (FAST-LIO)本方法原文提出 | 0.07 ms | (Xu & Zhang, 2021, Table II) |
He et al., 2023b · Table I 本方法 6 筆
指標odometry drift (%)
表格設定(擷取紀錄原文):Odometry drift (%) of the FAST-LIO based LiDAR-inertial system implemented with IKFoM (with online extrinsics) versus the hand-derived IESEKF of FAST-LIO, from start and end pose coincidence, six datasets, at most 5 iterations (He et al., 2023b, Table I)
odometry drift (%),own datasets (trial 01 by the authors; trial 02 from the FAST-LIO paper) · 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。
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 He et al., 2023b 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(He et al., 2023b, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| IKFoM-based原文提出 | 0.414% | (He et al., 2023b, Table I) |
| Hand-derived [14] (FAST-LIO)本方法 | 0.527% | (He et al., 2023b, Table I) |
其他比較組
來源
Xu & Zhang, 2021
(2021)FAST-LIO: A Fast, Robust LiDAR-Inertial Odometry Package by Tightly-Coupled Iterated Kalman FilterIEEE Robotics and Automation Letters, 6(2):3317-3324
DOI 10.1109/lra.2021.3064227arXiv 2010.08196程式碼
同儕審查已出版已讀全文近十年
相關版本
- 預印本:FAST-LIO: A Fast, Robust LiDAR-inertial Odometry Package by Tightly-Coupled Iterated Kalman Filter (arXiv v3) https://arxiv.org/abs/2010.08196
- 程式碼釋出:hku-mars/FAST_LIO 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; same repository now also hosts FAST-LIO2))。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。