SLAM 點雲圖鑑Point clouds for construction
M-01 · 2019

MC2SLAM

MC2SLAM 以兩個連續 LiDAR 掃描一起估計第一個掃描期間的運動：先以 IMU 積分（無 IMU 時以線性外推）預測兩掃描的軌跡，再在其上加一個隨時間線性增長的六自由度偏差，用點到平面殘差把稀疏取樣的查詢點配準到最近約 100 個掃描構成的局部地圖，完成配準與去畸變；殘差尺度以中位數絕對偏差穩健估計。每次只補償前一個掃描並插入近似 Poisson 圓盤取樣的局部地圖，得到的相對位姿再與 IMU 預積分因子放入因子圖，估計速度與偏差並使軌跡與重力對齊。

里程計＋局部建圖近十年已讀全文同儕審查已出版當代 LiDAR（慣性）里程計與 SLAM

閱讀版本
Springer version of record (LNCS 11269, pp. 60-72, 2019)
全文取得途徑
國立臺灣大學機構訂閱
本頁內容

Inertial LiDAR odometry that estimates motion over two consecutive sweeps as a linear-in-time bias on an IMU-predicted trajectory, registering Poisson-disk-sampled query points point-to-plane to a Poisson-disk local map with MAD-scaled robust weights, then fuses the resulting relative poses with IMU preintegration in a factor graph.

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 multi-beam LiDAR (Velodyne HDL-32 with built-in IMU in the authors' data; Velodyne HDL-64 in KITTI without IMU)、IMU (built-in HDL-32 IMU)
原文測試平台vehicle (car-mounted HDL-32 and KITTI)、wearable (helmet-mounted HDL-32, indoor and outdoor walking)
狀態估計two parts: per-sweep robust nonlinear least squares for a 6-DoF bias on an IMU-predicted trajectory over two sweeps (Tukey loss, MAD-based residual scale, Ceres), and an online factor graph with pose prior, laser odometry and IMU preintegration factors solved in Ceres every five sweeps (Sec. 3.2, 4)
資料關聯Poisson-disk-like query points (minimum spacing about 20 cm, then about 500 random points) matched point-to-plane to a local map within radius epsilon; plane normal by eigen analysis of neighbours (Sec. 3.1-3.2)
時間表示discrete factor-graph states at sweep starts; within two consecutive sweeps, an IMU-integrated (or linearly extrapolated) trajectory corrected by a bias growing linearly in time in the log map (Eq. 1)
去畸變two-scan motion compensation: the trajectory over sweeps k and k+1 is estimated jointly with registration, then only sweep k is compensated and inserted into the local map (Sec. 3.2)
迴圈閉合not in the reported system; authors state their implementation can close loops but the chapter focuses on odometry (footnote 1)
全域最佳化online factor graph over all sweep poses with IMU velocity and bias nodes (one dynamic node per five pose nodes); IMU factors re-integrated when the base state changes (Sec. 4)
地圖表示local map of all compensated points from the last about 100 sweeps, inserted only if farther than about 5 cm from stored points (approximate Poisson disk), stored in a uniform grid ordered by time (Sec. 3.2)
先驗資訊yaw-only constrained prior pose at the origin; gravity alignment from IMU; for KITTI an approximate intrinsic correction by Deschaud (Sec. 4.1, 5)
可輸出幾何odometry trajectory and motion-compensated sweeps accumulated into a point cloud map (Fig. 1)
計算需求real time on an Intel i7-3700K (as written): about 32 ms per sweep for motion compensation averaged over 1000 frames of campus run 1; pose-graph optimization about 108.6 ms every five sweeps in a separate thread (Sec. 5, Runtime)

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
LiDARVelodyne HDL 32歸入：Velodyne HDL-32E方法輸入MC2SLAM own datasets (campus run 1-2, campus drive, field)multi-beam LiDAR with built-in IMU; per-point timestamps from datasheet(Neuhaus et al., 2019, Sec. 5 Own Datasets)
LiDARVelodyne HDL 64歸入：Velodyne HDL-64E資料集感測器KITTI odometryscans pre-compensated in the KITTI odometry release(Neuhaus et al., 2019, Sec. 5 KITTI Dataset)
慣性量測單元（IMU）Velodyne HDL 32 built-in IMU歸入：Velodyne HDL-32 built-in IMU方法輸入MC2SLAM own datasets (campus run 1-2, campus drive, field)原文未報告(Neuhaus et al., 2019, Sec. 5 Own Datasets)
GNSS 接收器commercial GPS/INS (model not stated)參考或真值量測KITTI odometryground truth for the training sequences(Neuhaus et al., 2019, Sec. 5 KITTI Dataset)
載具平台helmet (head-mounted sensor)方法輸入MC2SLAM own datasets (campus run 1-2, campus drive, field)person walking indoors and outdoors (campus run)(Neuhaus et al., 2019, Sec. 5; Fig. 1)
載具平台car方法輸入MC2SLAM own datasets (campus run 1-2, campus drive, field)car-mounted sensor (campus drive, field)(Neuhaus et al., 2019, Sec. 5)
運算硬體Intel i7-3700K (as written)歸入：Intel i7-3700K執行運算平台未標示原文未報告(Neuhaus et al., 2019, Sec. 5 Runtime)

作者報告的優勢與限制

優勢

  • Lower KITTI training translation drift than IMLS on 6 of the 11 listed sequences and lower than LOAM on all 11 (Table 1)
  • KITTI test set translational error 0.69%, shared 4th/5th place among LiDAR methods at the time, with better rotation error and runtime than the equally ranked IMLS (Sec. 5)
  • IMU integration is essential for erratic head-mounted motion: campus run drift 0.41% and 0.54% with IMU versus 8.96% and 15.65% without (Table 2)
  • Local map strongly reduces long-term drift, e.g. field 0.42% with versus 7.50% without (Table 2, Fig. 3)
  • About 32 ms per sweep for motion compensation, real time on a desktop CPU (Sec. 5)

限制

  • KITTI scans had to be re-distorted with the previous relative pose and approximately recalibrated, which may introduce errors; the IMU part could not be evaluated on KITTI (Sec. 5)
  • Own-dataset accuracy is measured only by drift at manually registered loop checkpoints, not against an independent 6-DoF ground truth (Sec. 5)
  • IMU factors are re-integrated whenever the base state changes, which may be an issue for very long trajectories (Sec. 4.1)
  • Loop closing is not part of the evaluated system (footnote 1)
  • Using 1500 instead of 500 query points slightly improves accuracy but more than doubles bias-estimation time (17.6 to 37.8 ms) (Sec. 5)

營建工程相關證據

原論文只測 KITTI、校園車載與頭戴式室內外資料，未在營建場域驗證。但以 MC2SLAM 為基礎的 Vision & Robotics 團隊參加兩屆以營建場景為主的 Hilti SLAM Challenge：2021 年在論文表列的 13 支具名隊伍中得分第三（406 分，RMSE 0.101 m）（Helmberger et al., 2022），2022 年 Hilti-Oxford 在表列隊伍中得分第二（443.8 分，平均 ATE 3.94 cm），其中施工中建物一樓、多樓層與樓梯序列分別為 1.1、2.0 與 4.8 cm （Zhang et al., 2023c）。這些參賽版本另加迴圈閉合與光束法平差，與本章所述純里程計不同，精度不能直接歸於原論文方法。

原文驗證環境：公開基準、跨場域

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 4 個比較組，合計 28 筆紀錄。

Neuhaus et al., 2019 · Table 2 本方法 12 筆

指標translation drift per meter (in %) at loop checkpoints

表格設定（擷取紀錄原文）：Own HDL-32 datasets; translation drift per metre (%) computed at manually registered loop checkpoints and averaged over 10 runs with different random seeds; settings vary local map and IMU (Neuhaus et al., 2019, Table 2)

translation drift per meter (in %) at loop checkpoints，MC2SLAM own datasets · campus run 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 Neuhaus et al., 2019 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：%；場景：head-mounted indoor and outdoor walking (campus run) and car-mounted (campus drive, field)

資料來源作者報告值(Neuhaus et al., 2019, Table 2)

數值與出處
方法（原文寫法）報告值出處
Local Map yes, IMU yes本方法原文提出0.41%(Neuhaus et al., 2019, Table 2)
Local Map no, IMU yes本方法1.9%(Neuhaus et al., 2019, Table 2)
Local Map yes, IMU no本方法8.96%(Neuhaus et al., 2019, Table 2)

Neuhaus et al., 2019 · Table 1 本方法 11 筆

指標translation drift per meter (in %)

表格設定（擷取紀錄原文）：KITTI odometry training set; translation drift per metre (%) of LOAM [24], IMLS [6] and MC2SLAM; LiDAR only (no IMU in KITTI odometry); scans re-distorted and intrinsically corrected as in IMLS (Neuhaus et al., 2019, Table 1)

translation drift per meter (in %)，KITTI odometry · 00 Urban

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 Neuhaus et al., 2019 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：不適用；單位：%；場景：vehicle, urban, highway and country roads

資料來源作者報告值(Neuhaus et al., 2019, Table 1)

數值與出處
方法（原文寫法）報告值出處
LOAM [24]0.78%(Neuhaus et al., 2019, Table 1)
IMLS [6]0.5%(Neuhaus et al., 2019, Table 1)
Ours本方法原文提出0.51%(Neuhaus et al., 2019, Table 1)

Neuhaus et al., 2019 · Text Sec.5 Runtime 本方法 4 筆

資料集與序列MC2SLAM own datasets · campus run 1

表格設定（擷取紀錄原文）：Runtime of MC2SLAM components on campus run 1: mean over 1000 frames of campus run 1; 500 query points (Neuhaus et al., 2019, Text Sec.5 Runtime)

motion compensation per sweep (total)，MC2SLAM own datasets · campus run 1

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 Neuhaus et al., 2019 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。

統計量：平均值（mean）；對齊方式：不適用；單位：ms；場景：head-mounted walking

數值與出處
方法（原文寫法）報告值出處
MC2SLAM本方法原文提出硬體：Intel i7-3700K (as written)32 ms(Neuhaus et al., 2019, Sec. 5 Runtime)

Neuhaus et al., 2019 · Text Sec.5 本方法 1 筆

指標translational error

資料集與序列KITTI odometry · 11-21 (online test set)

表格設定（擷取紀錄原文）：KITTI online test set result as listed on the KITTI website (entry MC2SLAM), shared 4th/5th place among laser-based methods at the time of writing (Neuhaus et al., 2019, Text Sec.5)

translational error，KITTI odometry · 11-21 (online test set)

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 Neuhaus et al., 2019 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。

統計量：平均值（mean）；對齊方式：不適用；單位：%；場景：vehicle, urban and country roads

數值與出處
方法（原文寫法）報告值出處
MC2SLAM本方法原文提出0.69%(Neuhaus et al., 2019, Sec. 5 KITTI Dataset)

來源

  • Neuhaus et al., 2019

    Frank Neuhaus, Tilman Koß, Robert Kohnen, Dietrich Paulus（2019）MC2SLAM: Real-Time Inertial Lidar Odometry Using Two-Scan Motion CompensationPattern Recognition (GCPR 2018), Lecture Notes in Computer Science 11269, LNCS 11269, pp. 60-72

    DOI 10.1007/978-3-030-12939-2_5

    同儕審查已出版已讀全文近十年查證後修正

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