SLAM 點雲圖鑑Point clouds for construction
M-01 · 2023

HBA

HBA 針對大場景下原始光達 BA 計算量過大的問題，採「由下而上」分層 BA：在小視窗內做局部 BA 並把視窗內各幀合併為上一層的關鍵影格，逐層向上，最後在頂層做全域 BA；再「由上而下」以位姿圖最佳化把結果平滑回傳到所有原始幀位姿，並以局部 BA 的 Hessian 作為資訊矩陣。作者依計算複雜度推導最佳層數。

離線地圖精修近十年已讀全文同儕審查已出版全域一致性、場所辨識與長期建圖

閱讀版本
Version of record, IEEE RA-L 8(3):1523-1530 (IEEE Xplore HTML; tables read from the IEEE table images), cross-checked against arXiv 2209.11939v1 (PDF read in full)
全文取得途徑
國立臺灣大學機構訂閱
本頁內容

HBA splits large-scale LiDAR BA into a bottom-up pyramid of small parallel BA problems and a top-down pose graph that propagates the refinement to all frames, trading a single huge Hessian for many small ones.

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 (mechanical spinning in the public datasets; solid-state LiDAR of ref. [26], retina-like incommensurable scanning, in the self-collected data)
原文測試平台vehicle (public datasets KITTI and MulRan; platform described in the dataset papers, not in HBA)、handheld (public New College and Newer College datasets; platform described in the dataset papers)、self-collected solid-state LiDAR sequences (carrier platform not described)
狀態估計bottom-up hierarchical local BA in sliding windows (window 10, stride 5, parallel threads) plus global BA on top layer, followed by top-down pose-graph optimization using BA Hessians as information matrices
資料關聯plane features via adaptive voxelization (BALM) in each layer
時間表示discrete poses
去畸變input may be raw or deskewed scans (Sec. III-A); deskew not performed by HBA
迴圈閉合no place recognition module; can close gaps when the initial trajectory lacks loop closure if overlapping geometry is associated (Sec. IV-A2)
全域最佳化hierarchical BA plus pose graph, iterated until convergence
地圖表示layered keyframe point clouds; adaptive voxel plane features
先驗資訊initial pose trajectory from any LiDAR odometry or SLAM
可輸出幾何globally consistent point-cloud map and optimized poses
計算需求Offline, CPU parallel processing with n = 8 threads (Table I); CPU model and RAM size not reported. On MulRan DCC01, DCC02 and DCC03 the total optimization took 226.10 s, 362.59 s and 248.42 s versus 2830.91 s, 2631.83 s and 3655.10 s for the original BA and 4390.15 s, 4615.90 s and 7522.20 s for a reduced BA (version of record Table VIII). With the optimal layer number the method converges within about 12% of the data time (Sec. IV-C; Fig. 12 in the version of record, Fig. 11 in arXiv v1).

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
LiDARsolid-state LiDAR of ref. [26] (retina-like, incommensurable scanning; model not named)方法輸入self-collected scene-1 (indoor factory) and scene-2 (outdoor park)原文未報告(Liu et al., 2023b, Sec. IV-A2)
LiDARmechanical spinning LiDAR (model not named)資料集感測器KITTI, MulRan, New College, Newer College原文未報告(Liu et al., 2023b, Abstract; Sec. IV-A)
運算硬體CPU with 8 parallel threads (model not named)執行運算平台未標示n = 8 threads for parallel processing(Liu et al., 2023b, Table I; Table VIII)

作者報告的優勢與限制

優勢

  • Combines map-consistency benefit of BA with pose-graph efficiency (abstract)
  • Improved ATE even when initial trajectories were already loop-closed (Sec. IV-A1, Table II)
  • Lower time and RAM than original BA on New College/Newer College as layers increase (Sec. IV-C; Fig. 12 in the version of record, Fig. 11 in arXiv v1)

限制

  • Requires an initial pose trajectory from LiDAR odometry or SLAM (Sec. III-A)
  • IMU pre-integration and a LiDAR measurement noise model are not included; left to future work (Sec. V)
  • Bottom-up local BA ignores features co-visible across different local windows; the top-down PGO is needed to compensate (Sec. I, Sec. III-C)
  • False feature correspondences in the loosely parameterized top-layer global BA can add incorrect factors; the authors rely on dense lower-layer factors to contain them (Sec. IV-A2)
  • Self-collected scenes have no ground truth; map quality is shown only by mean map entropy (Sec. IV-A2, Table VI)
  • Not best on every KITTI sequence with loop-closed input: translation ATE is lower for CT-ICP on Seq. 02, 04, 08, 09 and 10, for LiTAMIN2 on Seq. 02, 04 and 08, and for SuMa and LOAM on Seq. 04 (VoR Table II)

營建工程相關證據

未見工地驗證。公開資料為 KITTI、MulRan、New College 與 Newer College；自行蒐集的固態光達資料有兩段：約 14 m×16 m×8 m、含不規則管線與機台的室內工廠（7339 幀），以及約 95 m×195 m 的戶外公園（3407 幀），兩段都沒有真值，只以平均地圖熵（MME）比較（Sec. IV-A2、Table VI）。可作為工地掃描軌跡的離線全域一致化模組（推論），但缺少以獨立參考點雲量化的幾何誤差。

原文驗證環境：公開基準

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 8 個比較組，合計 61 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 4 組列在最後，並連到性能比較頁。

Tao et al., 2025 · Table 3 本方法 14 筆

指標RMS of ATE

表格設定（擷取紀錄原文）：ATE RMS (m) against LiDAR-to-TLS ground truth after SE(3) Umeyama alignment; online: VILENS-SLAM, Fast-LIO-SLAM, SC-LIO-SAM, ImMesh, Fast-LIVO2; offline: HBA (input VILENS-SLAM), COLMAP (images only). VILENS-SLAM = VILENS with pose-graph optimisation; Fast-LIO-SLAM and SC-LIO-SAM add Scan Context loop closures to Fast-LIO2 and LIO-SAM. 'x' in the table = failed or incomplete. Authors note methods could improve with further tuning. (Tao et al., 2025, Table 3)

RMS of ATE，Oxford Spires · Keble College 02 (290 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。

這些是 Tao et al., 2025 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。

統計量：均方根誤差（RMSE）；對齊方式：SE(3) 剛體對齊；單位：m；場景：historic site, outdoor and indoor parts (Keble College, Oxford)

資料來源作者報告值(Tao et al., 2025, Table 3)

數值與出處
方法（原文寫法）報告值出處
VILENS-SLAM0.06 m(Tao et al., 2025, Table 3)
Fast-LIO-SLAM0.25 m(Tao et al., 2025, Table 3)
SC-LIO-SAM1.26 m(Tao et al., 2025, Table 3)
ImMesh0.08 m(Tao et al., 2025, Table 3)
Fast-LIVO20.95 m(Tao et al., 2025, Table 3)
HBA本方法0.11 m(Tao et al., 2025, Table 3)
COLMAP0.05 m(Tao et al., 2025, Table 3)

Liu et al., 2023b · Table II 本方法 12 筆

指標RMSE of the ATE, translation part of deg/m pair

表格設定（擷取紀錄原文）：KITTI with loop-closed MULLS poses as HBA input; RMSE of ATE printed as rotation (deg)/translation (m); only the translation part is extracted; the '(loops)' sequence labels follow the asterisks of arXiv v1 Table II (the version-of-record table image carries no asterisks); version of record adds CT-ICP and LOAM rows (Liu et al., 2023b, Table II)

RMSE of the ATE, translation part of deg/m pair，KITTI · Seq. 00 (loops)

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。

這些是 Liu 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。

統計量：均方根誤差（RMSE）；對齊方式：原文未報告；單位：m；場景：urban and rural driving

資料來源作者報告值(Liu et al., 2023b, Table II)

數值與出處
方法（原文寫法）報告值出處
Proposed本方法原文提出0.8 m(Liu et al., 2023b, Table II (version of record))
CT-ICP1.7 m(Liu et al., 2023b, Table II (version of record))
MULLS1.1 m(Liu et al., 2023b, Table II (version of record))
LiTAMIN21.3 m(Liu et al., 2023b, Table II (version of record))
SuMa1 m(Liu et al., 2023b, Table II (version of record))
LOAM1.5 m(Liu et al., 2023b, Table II (version of record))

Liu et al., 2023b · Table V 本方法 12 筆

指標RMSE of the ATE, translation part of deg/m pair

表格設定（擷取紀錄原文）：KITTI with MULLS poses without loop closure as HBA input; RMSE of ATE printed as rotation (deg)/translation (m); only translation extracted; baselines run without loop closure; the '(loops)' sequence labels follow the asterisks of arXiv v1 Table V (the version-of-record table image carries no asterisks); version of record adds CT-ICP and LOAM rows (Liu et al., 2023b, Table V)

RMSE of the ATE, translation part of deg/m pair，KITTI · Seq. 00 (loops)

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。

這些是 Liu 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。

統計量：均方根誤差（RMSE）；對齊方式：原文未報告；單位：m；場景：urban and rural driving

資料來源作者報告值(Liu et al., 2023b, Table V)

數值與出處
方法（原文寫法）報告值出處
Proposed本方法原文提出1.2 m(Liu et al., 2023b, Table V (version of record))
CT-ICP4.5 m(Liu et al., 2023b, Table V (version of record))
MULLS6.1 m(Liu et al., 2023b, Table V (version of record))
Voxel Map2.8 m(Liu et al., 2023b, Table V (version of record))
SuMa2.9 m(Liu et al., 2023b, Table V (version of record))
LiTAMIN25.8 m(Liu et al., 2023b, Table V (version of record))
LOAM2.3 m(Liu et al., 2023b, Table V (version of record))

Wang et al., 2026 · Table III 本方法 12 筆

表格設定（擷取紀錄原文）：Single-robot study on self-collected Mid360 data; spatial BA versus BALM2 (sliding window) and HBA, all on raw odometry without loop-based refinement; z-drift and z-RMSE relative to the z-value of the first frame; MME via MapEval (lower is better) (Wang et al., 2026, Table III)

MME，self-collected (Mid360) · Garage

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。

這些是 Wang 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。

統計量：平均值（mean）；對齊方式：未對齊；單位：無單位；場景：indoor, 334 m

資料來源作者報告值(Wang et al., 2026, Table III)

數值與出處
方法（原文寫法）報告值出處
HBA本方法-6.83(Wang et al., 2026, Table III)
BALM2-6.93(Wang et al., 2026, Table III)
Ours (spatial BA)原文提出-6.86(Wang et al., 2026, Table III)

其他比較組

列出其餘 4 個比較組

來源

  • Liu et al., 2023b

    Xiyuan Liu, Zheng Liu, Fanze Kong, Fu Zhang（2023）Large-Scale LiDAR Consistent Mapping Using Hierarchical LiDAR Bundle AdjustmentIEEE Robotics and Automation Letters, 8(3):1523-1530

    DOI 10.1109/lra.2023.3238902arXiv 2209.11939程式碼

    同儕審查已出版已讀全文近十年

程式碼：https://github.com/hku-mars/HBA（授權：GPL-2.0 (LICENSE file)）。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。

回到方法圖鑑

SLAM 點雲圖鑑

本站為綜述論文《SLAM-Based Point Clouds for Construction: A Review》研究過程的互動說明文件。互動圖為教學模擬，不代表任何特定系統的實測表現；方法描述、數值與結論以所引原始文獻為準。

內容基準日 2026-09-25 · 研究與網站製作使用 AI 輔助工具，見 證據標記與方法

選擇開啟Esc關閉