LOG-LIO
LOG-LIO 在 FAST-LIO2 的迭代誤差狀態卡爾曼濾波架構上,加入即時的局部幾何資訊估計。作者提出 Ring FALS:預先依 LiDAR 的環編號與方位角建立方位向量查找表,新掃描到達時只需距離值即可以近似最小平方求得每點法向量,避免鄰域搜尋。地圖以擴充的 ikd-Tree 管理,每個體素節點遞增維護點分布的平均與共變異數,並在收斂後固定。資料關聯先做可見性與法向量一致性檢查,再依序嘗試大尺度面元、小尺度面元,最後才退回點到平面。
本頁內容
FAST-LIO2-style iterated EKF LIO that estimates per-point normals from range only with a ring-indexed lookup table (Ring FALS), maintains incremental per-voxel point distributions in an extended ikd-Tree, and uses hierarchical association that prefers large then small point-to-surfel constraints over point-to-plane after visibility and normal-consistency checks.
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 spinning LiDAR with ring index (Velodyne 32-beam in M2DGR; Ouster OS1 16-channel in NTU VIRAL)、9-axis IMU (VectorNav VN100 in NTU VIRAL) |
|---|---|
| 原文測試平台 | ground robot (M2DGR ground platform; locomotion type not stated)、UAV (NTU VIRAL) |
| 狀態估計 | error-state iterated EKF adopted from FAST-LIO2, with the MAP update augmented by point-to-surfel residuals besides point-to-plane residuals (Eq. 13, Sec. V-C) |
| 資料關聯 | hierarchical: k nearest map points and voxels; visibility check (map normal versus ray) and normal consistency check (mean angle below 60 deg); then large merged surfel, else small fixed surfel of the voxel, else LOAM-style point-to-plane; surfel if planarity above 1.0 and lambda2/lambda1 above 100 (Sec. III-E, V-B) |
| 時間表示 | discrete scan poses with IMU backward propagation for undistortion (Sec. V-A) |
| 去畸變 | IMU backward propagation (FAST-LIO2 style) after normal estimation and voxel downsampling (Sec. V-A) |
| 迴圈閉合 | none (listed as future work, Sec. VII) |
| 全域最佳化 | none |
| 地圖表示 | ikd-Tree extended so that each node also stores an incrementally updated point distribution (mean and covariance) of its voxel; distributions are fixed once the Ring FALS normal and the distribution eigenvector agree within 20 deg or after 2 eta = 50 points (Sec. V-D, VI-A) |
| 先驗資訊 | LiDAR-specific lookup table of bearing vectors per ring and azimuth, precomputed for Ring FALS (Sec. III-B, IV) |
| 可輸出幾何 | odometry and a voxelized point map with per-voxel normals and surfels |
| 計算需求 | CPU real time: mean 28.1 ms per scan over all sequences versus 20.5 ms for FAST-LIO2 on an Intel Xeon Gold 6248R 3.00 GHz with 32 GB RAM; Ring FALS normals take 7.8 ms per 57,600-point Velodyne-32 scan versus 79.8 ms for single-thread PCL (Tables I and IV, Sec. VI-B) |
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 | Velodyne-32 (as written)歸入:Velodyne-32 | 資料集感測器 | M2DGR | about 57,600 points per scan | (Huang et al., 2024b, Sec. VI-B; Table I) |
| LiDAR | Ouster OS1 16-channel (horizontal)歸入:Ouster OS1-16 | 資料集感測器 | NTU VIRAL | 16,384 points per scan | (Huang et al., 2024b, Sec. VI-B; Table I) |
| 慣性量測單元(IMU) | VectorNav VN100 | 資料集感測器 | NTU VIRAL | 9-axis IMU | (Huang et al., 2024b, Sec. VI-B) |
| 全測站 | laser-tracker total station (model not stated) | 參考或真值量測 | NTU VIRAL | centimetre-level ground truth | (Huang et al., 2024b, Sec. VI-B) |
| 載具平台 | ground robot platform (not further specified) | 資料集感測器 | M2DGR | indoor and outdoor campus scenes, night street sequences | (Huang et al., 2024b, Sec. VI-B) |
| 載具平台 | UAV (not further specified) | 資料集感測器 | NTU VIRAL | 原文未報告 | (Huang et al., 2024b, Sec. VI-B) |
| 運算硬體 | Intel Core Xeon(R) Gold 6248R (as written)歸入:Intel Core Xeon(R) Gold 6248R | 執行運算平台 | 未標示 | 3.00 GHz, 32 GB RAM, Ubuntu 18.04 | (Huang et al., 2024b, Sec. VI-B) |
| 其他 | laser 3D tracking, motion capture and RTK receivers (models not stated) | 參考或真值量測 | M2DGR | ground-truth trajectories | (Huang et al., 2024b, Sec. VI-B) |
作者報告的優勢與限制
優勢
- Lowest mean translation RMSE on M2DGR (0.684 m versus 0.799 m FAST-LIO2 and 1.095 m LIO-SAM) and best on 10 of 21 sequences (Table II)
- Lowest mean RMSE on NTU VIRAL (0.330 m versus 0.423 m FAST-LIO2), with clearer gains on the high-altitude spms sequences where map overlap is limited (Table III, Sec. VI-D-2)
- Ring FALS normal estimation is about one tenth of single-thread PCL time on Velodyne-32 scans and four times faster than PCL with 10 OpenMP threads (Table I)
- Visibility check removes double-sided wall associations common indoors (Sec. V-B-2)
限制
- About 8 ms more per scan than FAST-LIO2 on average (Table IV, Sec. VI-D-3)
- Ring FALS assumes similar range within a small neighbourhood, which fails at wall edges, occlusions and missing returns; such normals rely on smoothing and outlier checks (Sec. IV, VI-C)
- Requires ring index and a LiDAR-specific lookup table, so it targets spinning LiDARs (Sec. III-B) (inference for non-repetitive solid-state scanners)
- No loop closure or dynamic-object handling (Sec. VII)
- LIO-SAM had lower RMSE on several NTU VIRAL sequences (eee_01-03, nya_02, sbs_02, rtp_02-03) and on M2DGR street06, street07, street09 and hall05 (Tables II-III)
營建工程相關證據
原論文僅在 M2DGR 校園地面機器人與 NTU VIRAL 無人機資料上評估,未涉及營建場域。Feng 等人(Feng et al., 2025)以預設參數在施工中醫院大樓的實際工地(1,004 m)與依 BIM 建立的 Gazebo 模擬工地測試十種方法,LOG-LIO 的 APE RMSE 在實際工地為 5.33 m、模擬工地為 20.81 m,均高於 LIO-SAM(2.26 m、3.19 m)與 FAST-LIO2(3.52 m、10.1 m)。該研究未說明實際工地參考軌跡來源,數值只能視為作者報告值;但它顯示在公開資料集上的小幅改進,未必能移轉到大型、結構重複的施工樓層(推論)。
原文驗證環境:公開基準、獨立參考量測
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 55 筆紀錄。
Huang et al., 2024b · Table II 本方法 22 筆
指標translation RMSE
表格設定(擷取紀錄原文):M2DGR; translation RMSE of ATE; loop closure disabled; map and scan downsampling 0.4 m; first and last 100 s of street07 and street10 discarded (RTK instability); LOG-C ablation column not stored (Huang et al., 2024b, Table II)
translation RMSE,M2DGR · gate01 (172 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。
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 Huang et al., 2024b 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(Huang et al., 2024b, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| LOG-LIO本方法原文提出 | 0.097 m | (Huang et al., 2024b, Table II) |
| FAST-LIO2 | 0.091 m | (Huang et al., 2024b, Table II) |
| LIO-SAM | 0.122 m | (Huang et al., 2024b, Table II) |
Huang et al., 2024b · Table III 本方法 19 筆
指標translation RMSE
表格設定(擷取紀錄原文):NTU VIRAL (horizontal OS1-16, VN100); translation RMSE of ATE; loop closure disabled; map and scan downsampling 0.5 m; x = failed; LOG-C ablation column not stored (Huang et al., 2024b, Table III)
translation RMSE,NTU VIRAL · eee_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 Huang et al., 2024b 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(Huang et al., 2024b, Table III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| LOG-LIO本方法原文提出 | 0.084 m | (Huang et al., 2024b, Table III) |
| FAST-LIO2 | 0.084 m | (Huang et al., 2024b, Table III) |
| LIO-SAM | 0.049 m | (Huang et al., 2024b, Table III) |
Huang et al., 2024b · Table IV 本方法 12 筆
指標average time consumption
表格設定(擷取紀錄原文):Average processing time per scan for each sequence group (Huang et al., 2024b, Table IV)
average time consumption,M2DGR and NTU VIRAL · M2DGR gate
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 Huang et al., 2024b 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(Huang et al., 2024b, Table IV)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| LOG-LIO本方法原文提出硬體:Intel Xeon Gold 6248R 3.00 GHz, 32 GB RAM | 46.563 ms | (Huang et al., 2024b, Table IV) |
| FAST-LIO2硬體:Intel Xeon Gold 6248R 3.00 GHz, 32 GB RAM | 31.378 ms | (Huang et al., 2024b, Table IV) |
Huang et al., 2024b · Table I 本方法 2 筆
指標mean running time of normal estimation
表格設定(擷取紀錄原文):Mean running time of normal estimation for a single scan; Ring FALS includes projection, box-filtering and smoothing; PCL least squares with k-d tree, without smoothing (Huang et al., 2024b, Table I)
mean running time of normal estimation,M2DGR · Velodyne-32 (57,600 points)
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 Huang et al., 2024b 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(Huang et al., 2024b, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Ring FALS (total)本方法原文提出硬體:Intel Xeon Gold 6248R 3.00 GHz, 32 GB RAM | 7.784 ms | (Huang et al., 2024b, Table I) |
| PCL single thread硬體:Intel Xeon Gold 6248R 3.00 GHz, 32 GB RAM | 79.811 ms | (Huang et al., 2024b, Table I) |
| PCL OMP 10 threads硬體:Intel Xeon Gold 6248R 3.00 GHz, 32 GB RAM | 26.355 ms | (Huang et al., 2024b, Table I) |
來源
Huang et al., 2024b
(2024)LOG-LIO: A LiDAR-Inertial Odometry With Efficient Local Geometric Information EstimationIEEE Robotics and Automation Letters, 9(1):459-466
DOI 10.1109/lra.2023.3332020arXiv 2307.09531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:arXiv 2307.09531 (v1 2023-07-18, v2 2023-08-14, v3 2023-10-11) https://arxiv.org/abs/2307.09531
- 程式碼釋出:tiev-tongji/LOG-LIO https://github.com/tiev-tongji/LOG-LIO
- 程式碼釋出:tiev-tongji/RingFalsNormal (stand-alone Ring FALS normal estimator, not opened) https://github.com/tiev-tongji/RingFalsNormal
程式碼:https://github.com/tiev-tongji/LOG-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。