SLAM 點雲圖鑑Point clouds for construction
M-01 · 2024

LTA-OM

LTA-OM 以 FAST-LIO2 作為光達慣性里程計、以 STD 作為迴圈偵測，整合迴圈校正、誤判迴圈剔除、長期關聯（long-term association, LTA）建圖與多時段定位建圖。其 LTA 建圖把校正後的歷史地圖直接作為 LIO 掃描對地圖配準的全域約束，使回到舊地點時里程計不再漂移；多時段模式可儲存校正後的地圖點、最佳化軌跡與描述子資料庫，供後續時段接續使用。

具全域修正的完整 SLAM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, Journal of Field Robotics 41(7):2455-2474, first published 15 April 2024 (Wiley Online Library HTML full text, open access; includes the 9 May 2024 correction notice on the corresponding author)
全文取得途徑
國立臺灣大學機構訂閱
本頁內容

LTA-OM combines FAST-LIO2 and STD loop detection with loop correction, false-loop rejection, and long-term association that feeds the corrected history map back into LIO registration, plus a multi-session mode storing maps and descriptors.

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public datasets MulRan and NCLT (carrier platforms not described in the paper)、multilevel building dataset collected with a Livox Avia LiDAR and its embedded IMU (carrier not described)
狀態估計FAST-LIO2 tightly coupled iterated Kalman filter LIO with long-term association (history map points reloaded into the ikd-Tree); separate loop optimization on a pose graph of submap poses with odometry factors plus neighbouring and loop-closure key-point factors (STD key points), solved with GTSAM and iSAM2; false-positive rejection by an optimize-and-recover graph consistency check (Sec. 4.1, 4.3)
資料關聯Direct point-to-plane scan-to-map registration on the ikd-Tree (plane fitted to five neighbours); STD-LCD on submaps of 20 accumulated scans: triangle descriptors of key points, hash-table rough detection, transform clustering with more than 4 supporters, and plane-to-plane overlap verification; neighbouring key-point pairs associated by kd-tree radius search (Sec. 4.1, 4.2, 4.3.1)
時間表示IMU forward propagation and backward propagation within each scan (FAST-LIO2); discrete submap poses in the pose graph (Sec. 4.1, 4.3.1)
去畸變FAST-LIO2 backward propagation with the IMU kinematic model compensates motion distortion of each scan (Sec. 4.1)
迴圈閉合STD-LCD loops on 20-scan submaps; only loops with overlap ratio above 0.5 are trusted; each loop is optimized after backing up the graph and rejected (graph restored) if key-point factor residuals exceed a threshold (2 in the benchmarks); the first loop or loops far from the current pose require two consecutive mutually consistent loops (Sec. 4.2, 4.3.2, 5.1)
全域最佳化iSAM2 pose graph with key-point factors, replaced inside each closed loop cycle by recalculated odometry factors to bound factor count; loop correction every 200 m: history submap points corrected with optimized poses, ikd-Tree rebuilt in a separate thread, scan re-registered by ICP (about 8 ms) and the LIO pose corrected (Sec. 4.3.3, 4.4)
地圖表示ikd-Tree live map holding recent scan points and corrected, on-tree downsampled history submap points loaded around the current position; stored multisession prior map with submap poses and STD descriptor database (Sec. 4.1, 4.4.1, 4.5)
先驗資訊optional prior session produced by LTA-OM itself (prestored map, submap poses and descriptor database); maps from other systems are not compatible (Sec. 4.5, Remark 4)
可輸出幾何corrected point-cloud map and optimized trajectory; multi-session stitched map (abstract)
計算需求Real time on an Intel i7-10700 CPU (2.90 GHz, 16 cores, 15.5 GB RAM): average total time of all threads 32.22 ms per scan on MulRan and 20.62 ms on NCLT (10 Hz LiDAR); loop correction 83.87 ms per run on average; loop detection 71.11 ms (MulRan) and 40.18 ms (NCLT) per run (Sec. 5, Tables 5-7)

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
LiDARLivox Avia LiDAR歸入：Livox Avia方法輸入multilevel building (authors' own data)原文未報告(Zou et al., 2024, Sec. 5.4)
LiDAROuster LiDAR (model not named)資料集感測器MulRan10 Hz(Zou et al., 2024, Sec. 5)
LiDARVelodyne LiDAR (model not named)資料集感測器NCLT10 Hz(Zou et al., 2024, Sec. 5)
慣性量測單元（IMU）embedded IMU of the Livox Avia方法輸入multilevel building (authors' own data)原文未報告(Zou et al., 2024, Sec. 5.4)
慣性量測單元（IMU）IMU (model not named)資料集感測器MulRan; NCLTmore than 100 Hz(Zou et al., 2024, Sec. 5)
運算硬體Intel i7-10700 CPU執行運算平台未標示2.90 GHz, 16 cores, 15.5 GB RAM(Zou et al., 2024, Sec. 5)

作者報告的優勢與限制

優勢

  • Lowest ATE RMSE on 8 of 9 MulRan sequences (average 4.83 m vs 5.67 m FAST-LIO-SC and 5.82 m LIO-SAM-SC) and on all 9 NCLT sequences (average 1.56 m vs 2.83 m FAST-LIO-SC) (Tables 2-3)
  • Fewer 5 cm map voxels than FAST-LIO-SC and LIO-SAM-SC on all nine MulRan sequences, and fewer than LTA-OM without LTA on eight of nine (not riverside03) (Table 4)
  • Average all-thread time 32.22 ms per scan on MulRan and 20.62 ms on NCLT, about 2 and 1.5 times faster than FAST-LIO-SC (Tables 5-6)
  • Loop-closing convergence 2.9 s versus 39.5 s for LIO-SAM-SC with a robust estimator on riverside02 (Sec. 4.3.2, Fig. 4)
  • No ICP verification of loop candidates; loop correction averages 83.87 ms per run (Sec. 5.2, Table 7)
  • Multisession mode lowered ATE in 5 of 6 MulRan tests and relocalized on the prior map within 4.9 s to 7.5 s (Tables 8-9)
  • Consistent map in a multilevel building where half of the detected loops were false positives, whereas a Cauchy robust estimator failed (Sec. 5.4, Fig. 11)

限制

  • Loop detection accumulates about 20 scans per submap, adding delay before loop optimization and correction (Remark 1)
  • Only loops with overlap ratio above 0.5 are trusted, discarding STD's small-overlap detections (Sec. 4.3.2)
  • Loop detection time (71.11 ms MulRan, 40.18 ms NCLT) is higher than the Scan Context variants of the baselines (Sec. 5.2)
  • Multisession mode requires a prior map produced by LTA-OM; maps from other systems are incompatible (Remark 4)
  • LTA did not reduce the voxel count on riverside03 because of a large non-revisited area, and ATE on DCC01 was slightly worse than FAST-LIO-SC (Sec. 5.1, Tables 2 and 4)
  • Map consistency evaluated only by voxel counts and comparison with a satellite image; no reference point cloud (Sec. 5.1)
  • Future work: GNSS fusion and edge or plane factors for global map optimization (Sec. 6)

營建工程相關證據

無工地測試。公開資料為 MulRan（DCC、KAIST、Riverside）與 NCLT 校園；另以 Livox Avia 與其內建 IMU 在多層建築中蒐集資料，該建築三個樓層有尺寸相同的相似矩形走廊，偵測到的迴圈約半數為誤判，FPR 仍建出一致地圖，而以 Cauchy 穩健估計取代時迴圈閉合失敗（Sec. 5.4、Fig. 11）；此情境與施工中各樓層結構重複相近（推論）。多時段模式可載入前次地圖接續建圖，並在 4.9 至 7.5 s 內於前次地圖上重新定位（Table 9），與工地重複掃描相關（推論）；但地圖一致性只以體素數與衛星影像比對評估，未以參考點雲量化。

原文驗證環境：公開基準、已完工建築

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 9 個比較組，合計 67 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 5 組列在最後，並連到性能比較頁。

Liu et al., 2026 · Table 2 (full SLAM with LC) 本方法 13 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。

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。

統計量：均方根誤差（RMSE）；對齊方式：原文未報告；單位：cm；場景：construction environment (sequence named construction)

資料來源作者報告值(Liu et al., 2026, Table 2 (full SLAM with LC))

數值與出處
方法（原文寫法）報告值出處
LeGO-LOAM8.8 cm(Liu et al., 2026, Table 2)
LiLi-OM6.2 cm(Liu et al., 2026, Table 2)
LIO-SAM6.1 cm(Liu et al., 2026, Table 2)
LTA-OM本方法1.27 cm(Liu et al., 2026, Table 2)
Our (Odom+LM+LC)原文提出0.78 cm(Liu et al., 2026, Table 2)
Our (Full)原文提出0.62 cm(Liu et al., 2026, Table 2)

Zou et al., 2024 · Table 8 本方法 12 筆

指標absolute trajectory error (root-mean-square error, meters)

表格設定（擷取紀錄原文）：Multisession mode on MulRan: the sequence with the smallest RMSE of each scene provides the prior map; 'R2-1' = testing on R1 with prior from R2; single-session values copied from Table 2 (Zou et al., 2024, Table 8)

absolute trajectory error (root-mean-square error, meters)，MulRan · R2-1

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 Zou 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。

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 campus driving

資料來源作者報告值(Zou et al., 2024, Table 8)

數值與出處
方法（原文寫法）報告值出處
LTA-OM single-session本方法6.98 m(Zou et al., 2024, Table 8)
LTA-OM multisession本方法原文提出6.16 m(Zou et al., 2024, Table 8)

Zou et al., 2024 · Table 2 本方法 10 筆

指標absolute trajectory error (RMSE, meters)

表格設定（擷取紀錄原文）：ATE RMSE (m) from the EVO tool after Umeyama alignment to ground truth; each value is the mean of five trials; identical LTA-OM parameters for all sequences of a dataset (FPR residual threshold 2, overlap threshold 0.5); baselines with default or minimally adjusted parameters (Zou et al., 2024, Table 2)

absolute trajectory error (RMSE, meters)，MulRan · R1 riverside01

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。

  • 未報告（沒有數值，不是 0）

這些是 Zou 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。

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 campus driving with buildings, mountain, river, bridges

資料來源作者報告值(Zou et al., 2024, Table 2)

數值與出處
方法（原文寫法）報告值出處
LTA-OM (proposed)本方法原文提出6.98 m(Zou et al., 2024, Table 2)
FAST-LIO-SC7.96 m(Zou et al., 2024, Table 2)
LIO-SAM-SC9.27 m(Zou et al., 2024, Table 2)
LeGO-LOAM-SC24.6 m(Zou et al., 2024, Table 2)
LILIOM無數值未報告註記（擷取紀錄）：other: printed as 999; the paper does not define this value(Zou et al., 2024, Table 2)

Zou et al., 2024 · Table 3 本方法 10 筆

指標absolute trajectory error (RMSE, meters)

表格設定（擷取紀錄原文）：ATE RMSE (m) from the EVO tool after Umeyama alignment to ground truth; each value is the mean of five trials; identical LTA-OM parameters for all sequences of a dataset (FPR residual threshold 2, overlap threshold 0.5); baselines with default or minimally adjusted parameters (Zou et al., 2024, Table 3)

absolute trajectory error (RMSE, meters)，NCLT · N1 2012-01-15

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。

  • 未報告（沒有數值，不是 0）

這些是 Zou 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。

統計量：均方根誤差（RMSE）；對齊方式：原文未報告；單位：m；場景：University of Michigan North Campus, different seasons

資料來源作者報告值(Zou et al., 2024, Table 3)

數值與出處
方法（原文寫法）報告值出處
LTA-OM (proposed)本方法原文提出1.46 m(Zou et al., 2024, Table 3)
FAST-LIO-SC2.44 m(Zou et al., 2024, Table 3)
LIO-SAM-SC42.03 m(Zou et al., 2024, Table 3)
LeGO-LOAM-SC288.59 m(Zou et al., 2024, Table 3)
LILIOM無數值未報告註記（擷取紀錄）：other: printed as 999; the paper does not define this value(Zou et al., 2024, Table 3)

其他比較組

列出其餘 5 個比較組

來源

  • Zou et al., 2024

    Zuhao Zou, Chongjian Yuan, Wei Xu, Haotian Li, Shunbo Zhou, Kaiwen Xue, Fu Zhang（2024）LTA‐OM: Long‐term association LiDAR–IMU odometry and mappingJournal of Field Robotics, 41(7):2455-2474

    DOI 10.1002/rob.22337程式碼

    同儕審查已出版已讀全文近十年查證後修正

程式碼：https://github.com/hku-mars/LTAOM（授權：conflicting: LICENSE file shows Apache-2.0 while README states GPLv2 (unresolved)）。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