GLIO
GLIO 在因子圖中緊耦合 GNSS 原始量測、LiDAR 與 IMU：第一階段以滑動視窗融合基準站差分後的雙差虛擬距離、都卜勒、IMU 預積分與 LiDAR 掃描對地圖平面因子；第二階段在獨立執行緒上對關鍵影格做批次最佳化，每個關鍵影格與 12 個相鄰影格建立掃描對多掃描約束，並逐步排除離群量測。GNSS 提供全域約束消除漂移，LiDAR 與 IMU 則在高樓遮蔽、GNSS 品質差時維持連續估測。
本頁內容
Tightly coupled GNSS raw-measurement, LiDAR and IMU factor graph with a sliding-window first stage (DD pseudorange, Doppler, preintegration, scan-to-map factors) and a batch second stage with scan-to-multiscan LiDAR factors, giving continuous and drift-free vehicle state estimates in urban canyons.
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 | low-cost GNSS receiver raw pseudorange and Doppler (u-blox F9P) with reference-station corrections、IMU (Xsens Ti-10)、3D LiDAR (Velodyne HDL-32E) |
|---|---|
| 原文測試平台 | vehicle (UrbanNav dataset, Hong Kong) |
| 狀態估計 | two-stage factor graph optimization in Ceres on a LILI-OM base: (1) sliding-window fusion of double-differenced pseudorange, Doppler, IMU preintegration, scan-to-map planar LiDAR factors and marginalization; (2) batch optimization over keyframes with scan-to-multiscan LiDAR factors and relative attitude constraints, run on a separate thread (max 50 iterations or 3 s) with outlier exclusion (Sec. III) |
| 資料關聯 | planar LiDAR features (100 per keyframe in the first stage; 25 random planar features per frame pair with 12 adjacent keyframes in the second stage); GNSS raw measurements double-differenced with a reference station; GNSS epochs associated to LiDAR keyframes by interpolation (Sec. III) |
| 時間表示 | discrete LiDAR keyframes at 10 Hz; GNSS at 10 Hz and IMU at 100 Hz linked through preintegration and interpolation |
| 去畸變 | not described; the LiDAR front end is adopted from LILI-OM |
| 迴圈閉合 | no explicit loop closure; global drift is removed by GNSS factors |
| 全域最佳化 | second-stage batch factor graph over keyframes with GNSS and scan-to-multiscan LiDAR factors (Sec. III) |
| 地圖表示 | LiDAR keyframe feature map in the global (GNSS) frame |
| 先驗資訊 | reference-station GNSS corrections; lever arm from prior calibration (Fig. 6 caption) |
| 可輸出幾何 | vehicle trajectory in a global frame and a LiDAR feature map |
| 計算需求 | preprocessing under 40 ms, first stage about 30 ms, total under 100 ms per frame, real time at 10 Hz; computer not specified (Sec. IV) |
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 | HDL-32E歸入：Velodyne HDL-32E | 方法輸入 | UrbanNav (Hong Kong) | 10 Hz | (Liu et al., 2024, Sec. IV) |
| 慣性量測單元（IMU） | Ti-10 | 方法輸入 | UrbanNav (Hong Kong) | 100 Hz | (Liu et al., 2024, Sec. IV) |
| GNSS 接收器 | F9P | 方法輸入 | UrbanNav (Hong Kong) | low-cost receiver; raw single-frequency GPS, BeiDou, Galileo and GLONASS at 10 Hz | (Liu et al., 2024, Sec. IV) |
| GNSS 接收器 | SPAN-CPT | 參考或真值量測 | UrbanNav (Hong Kong) | multi-frequency multi-constellation GNSS RTK with a tactical-grade IMU; post-processed ground truth | (Liu et al., 2024, Sec. IV) |
| GNSS 接收器 | GNSS reference station | 方法輸入 | UrbanNav (Hong Kong) | corrections used to remove systematic errors from pseudoranges | (Liu et al., 2024, Abstract; Sec. III) |
作者報告的優勢與限制
優勢
- On UrbanNav TST, GLIO-DS 2D mean error 1.21 m and 3D mean 2.07 m versus 7.53 and 18.02 m for RTKLIB and 3.24 and 5.28 m for LIO-SAM with DGNSS (Table I)
- On UrbanNav Whampoa (over 25 min and 4.5 km), GLIO-DS 3D mean 3.34 m versus 7.44 m for LIO-SAM with DGNSS and 28.36 m for LILI-OM (Table II)
- The second stage reduces errors strongly compared with the single-stage variant (GLIO-SS) on both sequences (Tables I-II)
- Real-time at 10 Hz (Sec. IV)
- Author-reported improvement of more than 70 % over GNSS-only positioning and standalone LIO (Abstract)
限制
- Maximum error of nearly 10 m remains because unhealthy GNSS observations are not fully mitigated (Sec. IV)
- Depends on reference-station corrections
- Evaluated only on two vehicle sequences of one dataset
營建工程相關證據
GLIO 針對高樓林立、GNSS 訊號受遮蔽的都市峽谷，與城市施工現場周邊的 GNSS 條件相近（推論）。不過誤差為公尺級，適合提供全域參考或抑制長距離漂移，不足以單獨支撐公分級點雲量測；論文也沒有評估地圖精度。
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 3 個比較組，合計 27 筆紀錄。
Liu et al., 2024 · Table I 本方法 12 筆
資料集與序列UrbanNav (Hong Kong) · TST
表格設定（擷取紀錄原文）：UrbanNav dataset; positioning error against NovAtel SPAN-CPT RTK/INS ground truth in metres; UrbanNav TST: starts under an overpass, dense tall office buildings, heavy traffic, narrow roads with limited sky view (Liu et al., 2024, Table I)
2D MEAN positioning error，UrbanNav (Hong Kong) · TST
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 Liu 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。
資料來源作者報告值(Liu et al., 2024, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| RTKLIB (GNSS RTK) | 7.53 m | (Liu et al., 2024, Table I; Sec. IV) |
| LIO (LILI-OM, aligned to the world frame by ground truth) | 2.21 m | (Liu et al., 2024, Table I; Sec. IV) |
| LIO-GNSS (LIO-SAM with DGNSS from RTKLIB) | 3.24 m | (Liu et al., 2024, Table I; Sec. IV) |
| GLIO-SS (single-stage only)本方法原文提出 | 1.53 m | (Liu et al., 2024, Table I; Sec. IV) |
| GLIO-DS (full two-stage)本方法原文提出 | 1.21 m | (Liu et al., 2024, Table I; Sec. IV) |
Liu et al., 2024 · Table II 本方法 12 筆
資料集與序列UrbanNav (Hong Kong) · Whampoa
表格設定（擷取紀錄原文）：UrbanNav dataset; positioning error against NovAtel SPAN-CPT RTK/INS ground truth in metres; UrbanNav Whampoa: over 25 min and 4.5 km from open sky into dense urban areas with overpasses, billboards and dynamic objects (Liu et al., 2024, Table II)
2D MEAN positioning error，UrbanNav (Hong Kong) · Whampoa
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 Liu 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。
資料來源作者報告值(Liu et al., 2024, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| RTKLIB (GNSS RTK) | 7.74 m | (Liu et al., 2024, Table II; Sec. IV) |
| LIO (LILI-OM, aligned to the world frame by ground truth) | 12.63 m | (Liu et al., 2024, Table II; Sec. IV) |
| LIO-GNSS (LIO-SAM with DGNSS from RTKLIB) | 2.68 m | (Liu et al., 2024, Table II; Sec. IV) |
| GLIO-SS (single-stage only)本方法原文提出 | 4.4 m | (Liu et al., 2024, Table II; Sec. IV) |
| GLIO-DS (full two-stage)本方法原文提出 | 1.68 m | (Liu et al., 2024, Table II; Sec. IV) |
Liu et al., 2024 · Sec. IV timing 本方法 3 筆
資料集與序列UrbanNav (Hong Kong)
表格設定（擷取紀錄原文）：Timing stated in the text (computer not specified) (Liu et al., 2024, Sec. IV timing)
preprocessing of GNSS, IMU and LiDAR features for both stages per frame，UrbanNav (Hong Kong)
這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Liu et al., 2024 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| GLIO本方法原文提出 | 40 ms僅報告範圍註記（擷取紀錄）：upper bound as printed ('less than 40ms') | (Liu et al., 2024, Sec. IV) |
來源
Liu et al., 2024
（2024）GLIO: Tightly-Coupled GNSS/LiDAR/IMU Integration for Continuous and Drift-Free State Estimation of Intelligent Vehicles in Urban AreasIEEE Transactions on Intelligent Vehicles, 9(1), pp. 1412-1422
DOI 10.1109/tiv.2023.3323648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：GLIO author manuscript (submitted version) in the GitHub repository https://github.com/XikunLiu-huskit/GLIO
- 程式碼釋出：XikunLiu-huskit/GLIO https://github.com/XikunLiu-huskit/GLIO
程式碼：https://github.com/XikunLiu-huskit/GLIO（授權：not specified (no license file detected by the GitHub API)）。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。