D-LIOM
D-LIOM 把 Cartographer 式的直接配準改為與 IMU 緊耦合的 3D 版本:每個去畸變掃描不擷取特徵,直接以高斯牛頓法對齊到 3D 佔據機率子地圖,得到的位姿作為一元 LiDAR 因子,與 IMU 預積分及線上估計的重力先驗因子組成子地圖時間窗內的局部因子圖,同時更新 IMU 偏差並抑制橫滾與俯仰漂移。後端利用重力對齊,把 3D 子地圖投影成 2D 影像,以 SURF 特徵與 FLANN 比對偵測迴圈並由 RANSAC 求得平面位移與旋轉,再以分支定界搜尋高度差、做精細配準後最佳化全域位姿圖。系統也支援多 LiDAR 輸入與 6 軸 IMU 的靜態或動態初始化。
本頁內容
Tightly coupled direct LIO and mapping that registers raw points to Cartographer-style 3D probability submaps, fuses the resulting pose with IMU preintegration and an online gravity prior in a submap-window factor graph, supports multiple LiDARs and dynamic 6-axis IMU initialization, and closes loops by SURF matching of gravity-aligned 2D projections of submaps followed by 3D refinement.
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 | one or more 3D spinning LiDARs (16-line RoboSense in the authors' device; two Ouster OS1-16 in NTU VIRAL; two inclined 16-line Velodyne in Complex Urban)、6-axis IMU (built-in consumer-grade IMU at 400 Hz; VectorNav VN100 in VIRAL) |
|---|---|
| 原文測試平台 | handheld (authors' device, carried while cycling and walking)、UAV (NTU VIRAL, DJI M600)、vehicle (Complex Urban Dataset) |
| 狀態估計 | front-end local factor graph in GTSAM over the time window of the current submap with LiDAR odometry unary factors from direct scan-to-submap matching, IMU preintegration factors and an online gravity-prior factor on roll and pitch; back-end global sparse pose graph in Ceres (Sec. III-E, IV-A) |
| 資料關聯 | direct: raw deskewed points are registered to a 3D probability (occupancy log-odds) submap by Gauss-Newton maximization of voxel probability using map gradients; no feature extraction (Sec. III-E-2) |
| 時間表示 | discrete scan nodes with IMU preintegration; LiDAR-IMU time offset ignored (Sec. III-C, IV-F) |
| 去畸變 | IMU preintegration between scans deskews each point to the previous body frame; points of auxiliary LiDARs are merged by timestamp with the primary LiDAR (Sec. III-C-2) |
| 迴圈閉合 | yes; gravity-aligned 3D submaps projected to 2D images, SURF keypoints matched with FLANN, RANSAC 3-DoF transform, then branch-and-bound search of the vertical offset and fine scan-to-submap registration (Sec. III-F) |
| 全域最佳化 | global sparse pose graph over submaps and nodes with loop constraints (Sec. III-F) |
| 地圖表示 | 3D probability submaps stored as octrees of occupancy log-odds voxels (Cartographer-style), plus 2D projections for loop detection (Sec. III-E-2) |
| 先驗資訊 | LiDAR-IMU extrinsics calibrated offline; gravity magnitude assumed known (Sec. III-D, IV-A) |
| 可輸出幾何 | trajectory and probability submaps; dense point cloud maps shown qualitatively (Figs. 5-6) |
| 計算需求 | real time on a workstation with two Intel Xeon E5-2620V3 at 2.4 GHz and 32 GB RAM: 27.5 ms odometry and 33.5 ms mapping per 16-line scan, 54.5 and 60.1 ms per 64-line scan (Table V, Sec. IV-A) |
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 | RoboSense 16-line LiDAR (model not stated) | 方法輸入 | TONGJI dataset (authors' handheld device) | 10 Hz | (Wang et al., 2023b, Sec. IV-A-2; Fig. 4(a)) |
| LiDAR | Ouster OS1-16-gen-1 (two units, horizontal and vertical)歸入:Ouster OS1-16 | 資料集感測器 | NTU VIRAL | 原文未報告 | (Wang et al., 2023b, Sec. IV-A-2; Sec. IV-C-1) |
| LiDAR | two 16-line Velodyne LiDARs (inclined) | 資料集感測器 | Complex Urban Dataset | large tilt angles | (Wang et al., 2023b, Sec. IV-A-2; Sec. IV-B-4) |
| 慣性量測單元(IMU) | built-in consumer-grade IMU of a ZED camera (model not stated) | 方法輸入 | TONGJI dataset (authors' handheld device) | 400 Hz; noise statistics from Woodman's approach | (Wang et al., 2023b, Sec. IV-A-2; Fig. 4(a)) |
| 慣性量測單元(IMU) | VectorNav-VN100歸入:VectorNav VN100 | 資料集感測器 | NTU VIRAL | 原文未報告 | (Wang et al., 2023b, Sec. IV-A-2) |
| 慣性量測單元(IMU) | AHRS IMU (model not stated) | 資料集感測器 | Complex Urban Dataset | 原文未報告 | (Wang et al., 2023b, Sec. IV-A-2) |
| GNSS 接收器 | GPS (model not stated) | 參考或真值量測 | Complex Urban Dataset | unstable under high-rise buildings; approximate reference only | (Wang et al., 2023b, Sec. IV-B-4) |
| 載具平台 | self-developed handheld device | 方法輸入 | TONGJI dataset (authors' handheld device) | carried while cycling (TJ-1 to TJ-4) and walking (TJ-5); aggressive rotations up to 232 deg/s (TJ-6, TJ-7) | (Wang et al., 2023b, Sec. IV-A-2; Table I) |
| 載具平台 | DJI M600 UAV | 資料集感測器 | NTU VIRAL | 原文未報告 | (Wang et al., 2023b, Sec. IV-A-2; Fig. 4(b)) |
| 載具平台 | car (Complex Urban vehicle-mounted platform) | 資料集感測器 | Complex Urban Dataset | two inclined LiDARs and an IMU; each sequence about 1 hour and nearly 2 square kilometres | (Wang et al., 2023b, Sec. IV-A-2; Sec. IV-B-4; Fig. 4(c)) |
| 運算硬體 | two Intel Xeon E5-2620V3 | 執行運算平台 | 未標示 | 2.4 GHz, 32 GB RAM workstation | (Wang et al., 2023b, Sec. IV-A-1) |
作者報告的優勢與限制
優勢
- With both VIRAL LiDARs, weighted average positioning error 0.10 m versus 0.15 m for LIO-SAM and 1.82 m for LIOM (Table III)
- Revisiting errors below 1 m on four 3 to 3.5 km TONGJI sequences where Carto3D, LOAM, LIOM and LIO-SAM had errors from 1.08 m up to hundreds of metres or failed (Table IV)
- Submap-to-submap loop detection reached 90.94% precision and 95.45% recall (Table VII)
- Ran on Complex Urban Urban-09 and Urban-10 with strongly inclined LiDARs where the compared methods produced no meaningful results (Sec. IV-B-4)
- Sharper reconstructions than LIO-SAM under rotations above 223 deg/s (Sec. IV-B-2, Fig. 5)
限制
- With a single horizontal LiDAR on VIRAL it was slightly less accurate than LIO-SAM, which the authors relate to submap resolution in small, slow scenes (Table III, Sec. IV-C-1)
- Loop detection can fail through feature mismatches and similar structures (false negatives and false positives) (Sec. IV-E, Fig. 9)
- The time offset between IMU and LiDAR is ignored (Sec. IV-F)
- TONGJI revisiting-error reference comes from NDT registration of revisited scans, not independent ground truth; Complex Urban GPS is only an approximate reference (Sec. IV-B-4, IV-C-2)
- The published Sec. III-D contained errors later corrected by the authors in the repository (README)
營建工程相關證據
論文未在營建場域測試;資料為同濟校園手持(騎車與步行)、NTU VIRAL 無人機(作者稱為室內資料)與 Complex Urban 城市車載序列。Ghadimzadeh Alamdari 等人(Ghadimzadeh Alamdari et al., 2025)在針對基礎設施非破壞檢測的比較中,將 D-LIOM 列為因程式庫不一致而未能執行。其直接機率子地圖配準不依賴特徵,並支援多 LiDAR 與傾斜安裝,理論上適合特徵稀少或安裝角度受限的工地平台,但實際可重現性受程式庫狀態限制(推論)。
原文驗證環境:公開基準、獨立參考量測、跨場域
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 42 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。
Wang et al., 2023b · Table III 本方法 20 筆
指標absolute positioning error
表格設定(擷取紀錄原文):NTU VIRAL absolute positioning errors; LIOM, LIO-SAM and Ours (H) use the horizontal LiDAR only, Ours (HV) uses horizontal and vertical LiDARs; w-avg weighted by number of poses (Wang et al., 2023b, Table III)
absolute positioning error,NTU VIRAL · eee01
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 Wang 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(Wang et al., 2023b, Table III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| LIOM [9] | 1.06 m | (Wang et al., 2023b, Table III) |
| LIO-SAM [10] | 0.1 m | (Wang et al., 2023b, Table III) |
| Ours (H)本方法原文提出 | 0.23 m | (Wang et al., 2023b, Table III) |
| Ours (HV)本方法原文提出 | 0.09 m | (Wang et al., 2023b, Table III) |
Wang et al., 2023b · Table VI 本方法 12 筆
指標revisiting error
表格設定(擷取紀錄原文):Ablation on TONGJI: D-LIOM without gravity factor (WoG) and without submap-to-submap loop detection (WoL); revisiting error (Wang et al., 2023b, Table VI)
revisiting error,TONGJI dataset · TJ-1 (3014.68 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 Wang 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(Wang et al., 2023b, Table VI)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| D-LIOM WoG (without gravity factor)本方法 | 7.14 m | (Wang et al., 2023b, Table VI) |
| D-LIOM WoL (without loop detection)本方法 | 12.14 m | (Wang et al., 2023b, Table VI) |
| D-LIOM本方法原文提出 | 0.81 m | (Wang et al., 2023b, Table VI) |
Wang et al., 2023b · Table IV 本方法 4 筆
指標revisiting error
表格設定(擷取紀錄原文):TONGJI self-collected handheld data; revisiting error = difference between each method's relative pose of two revisiting nodes and the relative pose from NDT registration of the revisited scans (Wang et al., 2023b, Table IV)
revisiting error,TONGJI dataset · TJ-1 (3014.68 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。
- 失敗
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 Wang 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(Wang et al., 2023b, Table IV)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Carto3D | 14.7 m | (Wang et al., 2023b, Table IV) |
| LOAM | 90.33 m | (Wang et al., 2023b, Table IV) |
| LIOM | 無數值失敗註記(擷取紀錄):failed | (Wang et al., 2023b, Table IV) |
| LIO-SAM | 1.08 m | (Wang et al., 2023b, Table IV) |
| D-LIOM本方法原文提出 | 0.81 m | (Wang et al., 2023b, Table IV) |
Wang et al., 2023b · Table V 本方法 4 筆
表格設定(擷取紀錄原文):Time cost per scan of the Odometry and Mapping modules for 16-line and 64-line LiDAR input; D-LIOM back-end time summed and averaged per frame (Wang et al., 2023b, Table V)
Odometry time cost,原文未報告 · 16 lines LiDAR scan
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 Wang 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(Wang et al., 2023b, Table V)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| LOAM硬體:two Intel Xeon E5-2620V3 2.4 GHz, 32 GB RAM | 55.6 ms | (Wang et al., 2023b, Table V) |
| LIOM硬體:two Intel Xeon E5-2620V3 2.4 GHz, 32 GB RAM | 684.7 ms | (Wang et al., 2023b, Table V) |
| LIO-SAM硬體:two Intel Xeon E5-2620V3 2.4 GHz, 32 GB RAM | 25.9 ms | (Wang et al., 2023b, Table V) |
| D-LIOM本方法原文提出硬體:two Intel Xeon E5-2620V3 2.4 GHz, 32 GB RAM | 27.5 ms | (Wang et al., 2023b, Table V) |
其他比較組
列出其餘 1 個比較組
來源
Wang et al., 2023b
(2023)D-LIOM: Tightly-Coupled Direct LiDAR-Inertial Odometry and MappingIEEE Transactions on Multimedia, 25:3905-3920
DOI 10.1109/tmm.2022.3168423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
- correction statement in article:Author-revised version in the code repository correcting typos and errors in Sec. III-D (system initialization) of the published article https://github.com/peterWon/D-LIOM/blob/master/paper/TMM_D_LIOM_FINAL_VERSION.pdf
- 程式碼釋出:peterWon/D-LIOM (built on Cartographer; TONGJI demo data) https://github.com/peterWon/D-LIOM
- 專案頁面:D-LIOM project page named in the paper footnote (not opened) https://cslinzhang.github.io/D-LIOM/D-LIOM.html
程式碼:https://github.com/peterWon/D-LIOM(授權:not_stated (no LICENSE file found; builds on Cartographer))。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。