GLIM
GLIM 以 GPU 加速的體素化 GICP 配準誤差因子（matching cost factor）取代傳統的掃描對模型配準與以高斯近似的相對位姿約束。里程計以固定延遲平滑（fixed-lag smoothing）在約數秒視窗內持續修正過去狀態，並以關鍵影格作為配準目標，使短暫幾何退化仍可藉由後續觀測回推修正。全域最佳化直接最小化各子地圖之間的配準誤差並緊耦合 IMU，可約束重疊很小的子地圖，但運算量高而需 GPU。
本頁內容
Range-inertial SLAM that uses GPU-accelerated voxelized-GICP registration-error factors both in fixed-lag-smoothing odometry and in global submap optimization, replacing pose-graph relative constraints.
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 (spinning and non-repetitive)、depth cameras (ToF, active stereo, stereo)、IMU、optional multi-camera |
|---|---|
| 原文測試平台 | handheld、UAV |
| 狀態估計 | fixed-lag smoothing (iSAM2 in GTSAM) with GPU voxelized-GICP matching-cost factors, IMU preintegration and keyframes; global factor graph minimizing registration errors between submaps with IMU constraints |
| 資料關聯 | voxelized GICP (distribution-to-distribution) with surface-orientation-based correspondence validation and multi-resolution voxelmaps |
| 時間表示 | discrete poses (authors judge continuous-time unnecessary when IMU prediction cancels distortion, Sec. II) |
| 去畸變 | IMU-based motion prediction transforms points to the IMU frame before covariance estimation (Sec. IV-B) |
| 迴圈閉合 | implicit: global matching-cost factors between overlapping submaps rather than explicit place recognition |
| 全域最佳化 | global multi-scan registration-error minimization over submaps with tightly coupled IMU constraints and submap endpoints |
| 地圖表示 | submaps of points with covariances and GPU voxelmaps |
| 先驗資訊 | none |
| 可輸出幾何 | globally optimized submap point clouds and trajectory; export format 原文未報告 in sections read |
| 計算需求 | Real time on a single consumer-grade GPU; park-sequence timing on an Intel Core i7 8700K with an NVIDIA 'RTX 1660 Ti' (as written): preprocessing 15.1 ms and odometry 29.3 ms per frame, local-map optimization 295.0 ms per submap (about every 2 s), global optimization 55.8 ms per submap with a maximum of about 250 ms; linear solving on the CPU is about 5% of global optimization time (Sec. VI-C; Table VII) |
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 VLP-16 | 方法輸入 | 未標示 | LiDAR model used to generate simulated point clouds; observation range limited to 15 m in a 40 m wide environment | (Koide et al., 2024, Sec. VI-A; Fig. 8) |
| LiDAR | Livox Avia | 方法輸入 | 未標示 | non-repetitive scan LiDAR; used for the eight flat-wall degeneration sequences and the cross-sensor test | (Koide et al., 2024, Sec. VI-A; Sec. VI-B; Table III) |
| LiDAR | Ouster OS0-32 | 方法輸入 | 未標示 | spinning LiDAR | (Koide et al., 2024, Sec. VI-B; Table III) |
| LiDAR | Ouster OS0-64 | 方法輸入 | 未標示 | spinning LiDAR | (Koide et al., 2024, Sec. VI-B; Table III) |
| LiDAR | Intel Realsense L515歸入：Intel RealSense L515 | 方法輸入 | 未標示 | described as solid-state LiDAR | (Koide et al., 2024, Sec. VI-B; Table III) |
| LiDAR | Ouster OS0-128 | 資料集感測器 | Multi-Camera Newer College | 10 Hz point clouds with 100 Hz IMU; PTP-synchronized with cameras | (Koide et al., 2024, Sec. VI-C) |
| LiDAR | Ouster OS1-16 | 資料集感測器 | NTU VIRAL | two units on a UAV | (Koide et al., 2024, Sec. VI-D) |
| 地面雷射掃描儀（TLS） | FARO Focus | 參考或真值量測 | 未標示 | survey-grade LiDAR; environment point cloud to which sensor trajectories are aligned for ground truth | (Koide et al., 2024, Sec. VI-B) |
| 慣性量測單元（IMU） | VectorNav VN100 | 資料集感測器 | NTU VIRAL | 原文未報告 | (Koide et al., 2024, Sec. VI-D) |
| 相機 | OMRON SENTECH STC-MBS202POE | 參考或真值量測 | 未標示 | images recorded with LiDAR-IMU data; synchronized via IEEE 1588 PTP; with wall AprilTags gives ground-truth trajectories | (Koide et al., 2024, Sec. VI-A) |
| 相機 | Sevensense Alphasense Core | 資料集感測器 | Multi-Camera Newer College | 4 hardware-synchronized cameras at 30 Hz | (Koide et al., 2024, Sec. VI-C) |
| 相機 | uEye 1221 LE | 資料集感測器 | NTU VIRAL | two cameras on a UAV | (Koide et al., 2024, Sec. VI-D) |
| 雙目相機 | Intel Realsense D455歸入：Intel RealSense D455 | 方法輸入 | 未標示 | active stereo camera | (Koide et al., 2024, Sec. VI-B; Table III) |
| 雙目相機 | Stereolabs ZED2i | 方法輸入 | 未標示 | passive stereo camera | (Koide et al., 2024, Sec. VI-B; Table III) |
| RGB-D 相機 | Microsoft Azure Kinect | 方法輸入 | 未標示 | time-of-flight depth camera | (Koide et al., 2024, Sec. VI-B; Table III) |
| UWB 測距 | Humatic P440 | 資料集感測器 | NTU VIRAL | 原文未報告 | (Koide et al., 2024, Sec. VI-D) |
| 運算硬體 | Intel Core i7 8700K | 執行運算平台 | 未標示 | CPU used for the processing-time measurement | (Koide et al., 2024, Sec. VI-C; Table VII) |
| 運算硬體 | NVIDIA RTX 1660 Ti | 執行運算平台 | 未標示 | consumer-grade GPU used for the processing-time measurement; model string printed exactly as 'NVIDIA RTX 1660 Ti' in Sec. VI-C | (Koide et al., 2024, Sec. VI-C; Table VII) |
作者報告的優勢與限制
優勢
- Tolerates a few seconds of completely degenerate range data where filter-based frame-to-model methods drift or diverge (abstract; Sec. VI-A)
- Real flat-wall degeneration test: ATE 0.017 to 0.389 m versus 0.297 to 15.551 m for FAST-LIO2 and 0.146 to 0.950 m for VoxelMap (Table II)
- Works with LiDARs and depth or stereo cameras under one parameter set: ATE 0.007 to 0.042 m for LiDARs and ToF camera, 0.139 and 0.206 m for stereo cameras (Sec. VI-B; Table III)
- Best ATE on the indoor stairs sequence (0.106 m odometry, 0.046 m with loop closure, versus 1.320 m for FAST-LIO2) (Table V)
- Global registration-error minimization constrains submaps with small overlap (Sec. I)
限制
- Long-term degeneration beyond the bounded optimization window remains challenging (Sec. VII)
- Requires much more computation than conventional methods and relies on a GPU for real time (abstract; Sec. I)
- Stereo-based cameras (Realsense D455, ZED2i) give distorted maps and larger errors due to distorted input clouds (Sec. VI-B; Table III)
- Multi-resolution voxelmap slightly lowers accuracy in small indoor scenes (stairs ATE 0.073 m without it versus 0.106 m with it) (Sec. VI-C; Table VIII)
- City- or nation-scale mapping would need a more scalable optimizer (Sec. VII)
- In the real degeneration test VoxelMap was better on Seq. 02 (0.146 versus 0.299 m) (Table II)
營建工程相關證據
原文未報告（驗證包括模擬走廊、面對平坦牆面的實測退化序列、以 FARO Focus 建立參考點雲的室內跨感測器實驗、Newer College 校園與 NTU VIRAL 空拍資料；牆面退化情境與施工中室內場景相似屬推論，論文未於工地驗證）
原文驗證環境：公開基準、受控實驗、模擬
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 6 個比較組，合計 54 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 2 組列在最後，並連到性能比較頁。
Koide et al., 2024 · Table V 本方法 20 筆
表格設定（擷取紀錄原文）：Multi-Camera Newer College (Ouster OS0-128, Alphasense Core); translational ATE; unlabeled rows are the no-loop-closure variant of the method printed in the next row (checked against the PDF layout and the ablation baseline in Table VIII) (Koide et al., 2024, Table V)
Absolute Trajectory Error [m] (no loop closure)，Multi-Camera Newer College · quad-easy
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 Koide 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(Koide et al., 2024, Table V)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LINS [3] | 0.16 m | (Koide et al., 2024, Table V) |
| LIO-SAM [10] (without loop closure; unlabeled row above LIO-SAM) | 0.086 m | (Koide et al., 2024, Table V) |
| FAST-LIO2 [5] | 0.068 m | (Koide et al., 2024, Table V) |
| CLINS [11] (without loop closure; unlabeled row above CLINS) | 0.197 m | (Koide et al., 2024, Table V) |
| DLO [14] | 0.08 m | (Koide et al., 2024, Table V) |
| GLIM (odometry, without loop closure; unlabeled row above GLIM)本方法原文提出 | 0.07 m | (Koide et al., 2024, Table V) |
Koide et al., 2024 · Table III 本方法 14 筆
表格設定（擷取紀錄原文）：Cross-sensor test with one parameter set; reference trajectories from alignment to a FARO Focus environment point cloud; RTE sub-trajectory length 2 m; FAST-LIO2 gave no reasonable result except for Ouster and Livox (Koide et al., 2024, Table III)
ATE [m] (± spread omitted)，authors' cross-sensor sequences · Ouster OS0-32
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 Koide 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 |
|---|---|---|
| GLIM本方法原文提出 | 0.037 m | (Koide et al., 2024, Table III) |
Koide et al., 2024 · Table II 本方法 8 筆
指標Absolute Trajectory Error [m], reported as value ± (± not defined in the paper)
表格設定（擷取紀錄原文）：Eight real sequences (path 2.2-4.8 m) with a Livox Avia moved between two pillars while facing a flat wall; ground truth from AprilTag bundle adjustment and camera-IMU batch optimization; LIO-SAM gave no decent result (Koide et al., 2024, Table II)
Absolute Trajectory Error [m], reported as value ± (± not defined in the paper)，authors' flat-wall degeneration sequences · Seq. 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 Koide 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(Koide et al., 2024, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| FAST-LIO2 [5] | 0.815 m原文指標寫法：Absolute Trajectory Error [m], reported as value ± 0.5 (± not defined in the paper) | (Koide et al., 2024, Table II) |
| VoxelMap [61] | 0.577 m原文指標寫法：Absolute Trajectory Error [m], reported as value ± 0.157 (± not defined in the paper) | (Koide et al., 2024, Table II) |
| GLIM本方法原文提出 | 0.118 m原文指標寫法：Absolute Trajectory Error [m], reported as value ± 0.047 (± not defined in the paper) | (Koide et al., 2024, Table II) |
Koide et al., 2024 · Table I 本方法 5 筆
指標Absolute Trajectory Error [m], reported as value ± (± not defined in the paper)
表格設定（擷取紀錄原文）：Simulated corridor 40 m wide with LiDAR range limited to 15 m so range data fully degenerate mid-corridor; five IMU noise levels; ATE via evo; loop closure disabled for all methods (Koide et al., 2024, Table I)
Absolute Trajectory Error [m], reported as value ± (± not defined in the paper)，simulation (Velodyne VLP-16 model, OpenVINS IMU synthesis) · IMU noise 1.0e-3 [m/s^2] and [deg/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 Koide 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(Koide et al., 2024, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LIO-SAM [10] | 1.473 m原文指標寫法：Absolute Trajectory Error [m], reported as value ± 0.696 (± not defined in the paper) | (Koide et al., 2024, Table I) |
| FAST-LIO2 [5] | 1.294 m原文指標寫法：Absolute Trajectory Error [m], reported as value ± 0.691 (± not defined in the paper) | (Koide et al., 2024, Table I) |
| GLIM本方法原文提出 | 0.099 m原文指標寫法：Absolute Trajectory Error [m], reported as value ± 0.012 (± not defined in the paper) | (Koide et al., 2024, Table I) |
其他比較組
來源
Koide et al., 2024
（2024）GLIM: 3D range-inertial localization and mapping with GPU-accelerated scan matching factorsRobotics and Autonomous Systems, 179:104750
DOI 10.1016/j.robot.2024.104750arXiv 2407.10344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：GLIM (arXiv v1) https://arxiv.org/abs/2407.10344
- 程式碼釋出：koide3/glim https://github.com/koide3/glim
程式碼：https://github.com/koide3/glim（授權：MIT (LICENSE file checked); authors state the majority of the code is released）。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。