Voxel-SLAM
Voxel-SLAM 以同一種自適應體素地圖貫穿初始化、里程計、局部建圖、迴圈與全域建圖五個模組，並依作者所稱的短期、中期、長期與多地圖四類資料關聯設計。局部建圖以滑動視窗 LiDAR-慣性光束法平差（bundle adjustment, BA）同時修正狀態與地圖；迴圈偵測以 BTC 描述子並加上平面約束與漂移比例檢查，觸發位姿圖最佳化與地圖重建。全域建圖以階層式 BA 從關鍵影格視窗到子地圖逐層最佳化，以提升多次作業地圖的一致性；其地圖表示明言沿用 VoxelMap。
本頁內容
A complete LiDAR-inertial SLAM that uses one adaptive voxel map for initialization, EKF odometry, sliding-window LiDAR-inertial BA, BTC-based multi-session loop closure, and hierarchical global BA.
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、IMU |
|---|---|
| 原文測試平台 | handheld、UAV、vehicle |
| 狀態估計 | Odometry: IMU propagation with motion compensation, then scan-to-map registration against the adaptive voxel map following VoxelMap (point-to-plane distances with per-point noise plus IMU residuals update the current state); the rough LIO inside initialization follows FAST-LIO; local mapping: sliding-window LiDAR-inertial BA (window 10) combining BALM2 point-cluster BA factors with IMU preintegration, solved by Levenberg-Marquardt with analytic Jacobian and Hessian; PGO via GTSAM after loops; hierarchical global BA |
| 資料關聯 | point-to-plane to adaptive voxel-map planes; BA over point clusters per voxel |
| 時間表示 | discrete poses |
| 去畸變 | IMU propagation compensates in-scan motion distortion within the odometry module (Sec. VI-A) |
| 迴圈閉合 | BTC descriptor place recognition with geometric verification, plus checks for plane constraints in three independent directions and a drift-to-travel-distance ratio (e.g., at most 1%); works within and across sessions |
| 全域最佳化 | Pose graph optimization (GTSAM) with map rebuilding when the drift distance exceeds a threshold (e.g., 0.1 m); keyframe-window BA (size 10, stride 5) in real time; after a session ends, global BA over submaps with coarse-to-fine voxelization and a top-down PGO (hierarchical BA), also across sessions |
| 地圖表示 | adaptive voxel map (planes with point clusters) shared by initialization, odometry, local mapping, loop closure and global mapping |
| 先驗資訊 | none (previous sessions used for multi-map association) |
| 可輸出幾何 | globally optimized multi-session point-cloud map and scan poses |
| 計算需求 | Laptop Intel i7-10750H (3.5 GHz, 32 GB) and onboard Intel i3-N305 (3.0 GHz, 16 GB); mean total per-scan thread time 0.027 s (Hilti), 0.095 s (MARS-LVIG), 0.051 s (UrbanNav), 0.044 s (MulRan), 0.037 s onboard; global mapping after session end 3.2 s (Hilti mean) to 62.1 s (MulRan mean), 14 s for the 11-min onboard sequence; memory 1.1 to 5.3 GB |
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 | Hesai XT-32 | 資料集感測器 | Hilti (handheld sequences, Table C1) | handheld sequences; LiDAR at 10 Hz | (Liu et al., 2026, Sec. 10) |
| LiDAR | Livox Avia | 資料集感測器 | MARS-LVIG | downward-looking on a UAV at about 100 m height; speed up to 12 m/s | (Liu et al., 2026, Sec. 10; Sec. 10.2.2) |
| LiDAR | HDL-32 E (as written)歸入：Velodyne HDL-32E | 資料集感測器 | UrbanNav | robotcar urban dataset | (Liu et al., 2026, Sec. 10) |
| LiDAR | OS1-64歸入：Ouster OS1-64 | 資料集感測器 | MulRan | robotcar; LiDAR loses about 70 deg of FoV | (Liu et al., 2026, Sec. 10) |
| LiDAR | Livox Avia | 方法輸入 | private dataset (private1, private2) | on the authors' handheld device | (Liu et al., 2026, Sec. 10) |
| 行動掃描設備 | light handheld device with Livox Avia and its internal IMU | 方法輸入 | private dataset (private1, private2) | used for private1 (forest, fast initial motion) and private2 (campus with elevator and escalator, 11 min 34 s) | (Liu et al., 2026, Sec. 10; Fig. 16a; Table C1) |
| 慣性量測單元（IMU） | Bosch BMI085 | 資料集感測器 | Hilti (handheld sequences, Table C1) | 400 Hz | (Liu et al., 2026, Sec. 10) |
| 慣性量測單元（IMU） | BMI088 (built-in IMU of Livox Avia) | 資料集感測器 | MARS-LVIG | 200 Hz | (Liu et al., 2026, Sec. 10) |
| 慣性量測單元（IMU） | Xsens-MTI-30 (as written)歸入：Xsens MTi-30 | 資料集感測器 | UrbanNav | 400 Hz | (Liu et al., 2026, Sec. 10) |
| 慣性量測單元（IMU） | IMU (model not reported) | 資料集感測器 | MulRan | 100 Hz | (Liu et al., 2026, Sec. 10) |
| 運算硬體 | laptop with Intel i7-10750H (written i7-10 750 H) at 3.5 GHz, 32 GB memory | 執行運算平台 | 未標示 | used for initialization, single-session and multisession experiments | (Liu et al., 2026, Sec. 10) |
| 運算硬體 | onboard computer with Intel i3-N305 at 3.0 GHz, 16 GB memory | 執行運算平台 | 未標示 | used for the online relocalization experiment (private2) | (Liu et al., 2026, Sec. 10; Sec. 10.4) |
作者報告的優勢與限制
優勢
- Initializes within about 1 s of data even from a highly dynamic initial state where FAST-LIO2 diverged (Sec. 1; Sec. 10.1)
- Multi-session merging demonstrated on Hilti sequences recorded at the same construction site: multisession ATE 7.6 cm with PGO only and 4.9 cm with global mapping (Table 5)
- Full system has the lowest ATE on all 13 Hilti sequences among compared odometry and SLAM methods; global mapping improves ATE even without detected loops (Table 2)
- Global mapping after a session costs 0.6% to 4.7% of the data collection time (Sec. 10.5)
限制
- Missed loop between overlapping Hilti sequences (narrow corridor) led to insufficient constraints in one multi-session case, later mitigated by global mapping (Sec. 10.3.1; Fig. 12b)
- Relocalization does not remove the root cause of divergence; in long degenerate conditions such as dynamic scenes, rain or fog, or long feature-poor tunnels, frequent reinitialization wastes data, so the system suits non-extreme scenarios (Sec. 10.4, Remark 1)
- Relies only on planar features from the adaptive voxel map (Sec. 1; Sec. 3.3)
- Future work: fuse images for degenerate scenes and colour, and GPU acceleration for global mapping (Sec. 11)
營建工程相關證據
作者使用 Hilti 資料集手持序列（Hesai XT-32 與 Bosch BMI085 IMU），並描述為室內外結構化施工環境。依期刊版附錄表 C1，hilti01 至 hilti08 對應 exp01、exp02、exp03、exp07、exp09、exp11、exp15 與 exp21 序列，其中只有 exp01 至 exp03 以 construction 命名，其餘為 long-corridor、cupola、gallery 與 outside；hilti09 至 hilti13 對應 site1 手持序列，作者明言為同一施工現場，用於多次作業地圖合併：多次作業 ATE 由僅位姿圖最佳化的 7.6 cm 降至全域建圖後的 4.9 cm（表 5）。單次作業時完整系統在 hilti01 至 hilti13 的 ATE 為 0.62 至 13.8 cm（表 2），誤差最大者為含狹窄樓梯的序列。評估僅為軌跡 ATE，未見點雲幾何對參考量測的評估。
原文驗證環境：公開基準、施工中工地
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 64 筆紀錄。
Liu et al., 2026 · Table 2 (full SLAM with LC) 本方法 26 筆
指標absolute trajectory error (RMSE, centimeters)
表格設定（擷取紀錄原文）：Hilti handheld sequences; ATE exported from the Hilti evaluation website; full SLAM with loop closure (Our (Full) adds global mapping) (Liu et al., 2026, Table 2 (full SLAM with LC))
absolute trajectory error (RMSE, centimeters)，Hilti handheld sequence exp01-construction (name per Table C1) · hilti01
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., 2026 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., 2026, Table 2 (full SLAM with LC))
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LeGO-LOAM | 8.8 cm | (Liu et al., 2026, Table 2) |
| LiLi-OM | 6.2 cm | (Liu et al., 2026, Table 2) |
| LIO-SAM | 6.1 cm | (Liu et al., 2026, Table 2) |
| LTA-OM | 1.27 cm | (Liu et al., 2026, Table 2) |
| Our (Odom+LM+LC)本方法原文提出 | 0.78 cm | (Liu et al., 2026, Table 2) |
| Our (Full)本方法原文提出 | 0.62 cm | (Liu et al., 2026, Table 2) |
Liu et al., 2026 · Table 2 (odometry without LC) 本方法 26 筆
指標absolute trajectory error (RMSE, centimeters)
表格設定（擷取紀錄原文）：Hilti handheld sequences (Hesai XT-32, BMI085 400 Hz); ATE exported from the Hilti evaluation website; odometry without loop closure; all methods with default parameters (Liu et al., 2026, Table 2 (odometry without LC))
absolute trajectory error (RMSE, centimeters)，Hilti handheld sequence exp01-construction (name per Table C1) · hilti01
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., 2026 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., 2026, Table 2 (odometry without LC))
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LeGO-LOAM | 9.1 cm | (Liu et al., 2026, Table 2) |
| LiLi-OM | 6.2 cm | (Liu et al., 2026, Table 2) |
| LINS | 6.5 cm | (Liu et al., 2026, Table 2) |
| LIO-SAM | 7.4 cm | (Liu et al., 2026, Table 2) |
| FAST-LIO2 | 1.3 cm | (Liu et al., 2026, Table 2) |
| Faster-LIO | 1.1 cm | (Liu et al., 2026, Table 2) |
| Point-LIO | 1.1 cm | (Liu et al., 2026, Table 2) |
| Our (Odom)本方法原文提出 | 1.3 cm | (Liu et al., 2026, Table 2) |
| Our (Odom+LM)本方法原文提出 | 0.8 cm | (Liu et al., 2026, Table 2) |
Liu et al., 2026 · Table 5 本方法 12 筆
指標ATE (RMSE, centimeters)
表格設定（擷取紀錄原文）：Multisession SLAM: sequences fed in the order hilti13, 12, 11, 10, 09; ATE of each session and of the merged multisession trajectory, pose graph optimization only versus with global mapping (hierarchical BA) (Liu et al., 2026, Table 5)
ATE (RMSE, centimeters)，Hilti handheld sequence site1-handheld-1 (name per Table C1) · hilti09
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., 2026 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., 2026, Table 5)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Voxel-SLAM (PGO only)本方法原文提出 | 1 cm | (Liu et al., 2026, Table 5) |
| Voxel-SLAM (global mapping)本方法原文提出 | 0.6 cm | (Liu et al., 2026, Table 5) |
來源
Liu et al., 2026
（2026）Voxel‐SLAM: A Complete, Accurate, and Versatile Light Detection and Ranging‐Inertial Simultaneous Localization and Mapping SystemAdvanced Intelligent Systems, 8(4):e202501081
DOI 10.1002/aisy.202501081arXiv 2410.08935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：Voxel-SLAM: A Complete, Accurate, and Versatile LiDAR-Inertial SLAM System (arXiv v1) https://arxiv.org/abs/2410.08935
- 程式碼釋出：hku-mars/Voxel-SLAM https://github.com/hku-mars/Voxel-SLAM
程式碼：https://github.com/hku-mars/Voxel-SLAM（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。