SLAM 點雲圖鑑Point clouds for construction
M-01 · 2015

OKVIS

OKVIS 以非線性最佳化緊耦合（tightly-coupled）融合相機重投影誤差與 IMU 慣性誤差，並只保留有限數量的關鍵影格，透過邊際化維持即時運算。關鍵影格可相隔任意時間，仍以線性化慣性項連結。作者以自製、硬體同步的雙目慣性裝置收集資料，並與 MSCKF 濾波器比較，也示範線上外參校正。

里程計＋局部建圖經典已讀全文同儕審查已出版視覺、視覺慣性與 RGB-D 稠密 SLAM

閱讀版本
Accepted manuscript deposited in Imperial College Spiral (26 pages); SAGE version of record not compared (journals.sagepub.com returned HTTP 403 to curl)
全文取得途徑
作者自存版本
本頁內容

OKVIS fuses reprojection and inertial errors in a keyframe-based sliding-window nonlinear optimisation with marginalisation, evaluated on custom hardware-synchronised stereo-inertial data against an MSCKF filter.

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stereo、monocular camera、IMU
原文測試平台handheld (Vicon Loops; outdoor loop around ETH Main Building; indoor multi-floor ETH main building demonstration)、helmet-mounted (Bicycle Trajectory)
狀態估計Nonlinear least squares (Google Ceres) over reprojection errors (keypoint std 0.8 px) and IMU error terms in a window of M keyframes plus the S most recent frames (M = 7, S = 3 in all experiments); frames leaving the window are marginalised by Schur complement with first-estimate Jacobians, dropping non-keyframe landmark observations and marginalising landmarks seen only in the oldest keyframes so that sparsity is kept; optional online camera-IMU extrinsics estimation
資料關聯Customised multi-scale SSE-optimised Harris corners with BRISK descriptors oriented along the projected gravity direction; brute-force 3D-2D matching against landmarks predicted visible, outliers removed by a Mahalanobis test on the IMU-propagated pose and an OpenGV absolute-pose RANSAC; then brute-force 2D-2D matching with stereo and temporal triangulation (only points with low depth uncertainty initialised) and a relative RANSAC against the newest keyframe. In the comparison all algorithms were fed the same correspondences produced by the stereo pipeline
時間表示Discrete states at image times (position, orientation quaternion, velocity, gyro and accelerometer biases); each IMU error term integrates all IMU readings between successive camera frames with the classical Runge-Kutta method, gyro bias modelled as random walk and accelerometer bias as bounded random walk; kept keyframes may be arbitrarily far apart in time
去畸變不適用
迴圈閉合none (odometry)
全域最佳化none
地圖表示sparse landmarks in a bounded keyframe window
先驗資訊No prior map. Intrinsics and camera-IMU extrinsics pre-calibrated with the method of Furgale et al. (2013); IMU noise from the ADIS16448 datasheet made slightly more conservative; weak zero-mean priors on speed (3 m/s) and biases (0.1 rad/s gyro, 0.2 m/s^2 accelerometer) for robust initialisation; the online-extrinsics option uses weak priors (10 mm, 0.6 degrees)
可輸出幾何time series of poses, velocities and IMU biases plus a sparse landmark map (conclusion)
計算需求Real time with bounded complexity (the dense part grows with O(M^3) in the number of keyframes); computationally more demanding than the MSCKF baseline; no host CPU or timing figures are reported; the sensor's FPGA can perform keypoint detection to save CPU

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
慣性量測單元（IMU）ADIS16448方法輸入未標示MEMS IMU recorded at 800 Hz; noise used: gyro 1.2e-3 rad/(s sqrt(Hz)), accelerometer 8.0e-3 m/(s^2 sqrt(Hz)), gyro bias 2.0e-5, accelerometer bias 5.5e-5 (Table I)(Leutenegger et al., 2015, Sec. VII-A1, VII-A2, Table I)
GNSS 接收器Leica Viva GS14參考或真值量測未標示post-processed DGPS 3D ground truth at 1 Hz; measurements with position uncertainty above 1 m discarded(Leutenegger et al., 2015, Sec. VII-A3, VII-B2; Table II)
雙目相機two embedded WVGA monochrome cameras (model not stated)方法輸入未標示11 cm baseline, 20 Hz in the datasets (hardware up to 60 Hz), rigidly mounted on an aluminium frame with the IMU(Leutenegger et al., 2015, Sec. VII-A1; Fig. 11)
載具平台helmet-mounted sensor and GNSS recorder (bicycle ride)方法輸入未標示7940 m in 23 min, up to 13.1 m/s(Leutenegger et al., 2015, Sec. VII-B2; Table II; Fig. 14)
載具平台hand-held sensor方法輸入未標示Vicon Loops 1200 m; ETH Main Building outdoor loop 620 m; qualitative 470 m indoor walk over three floors(Leutenegger et al., 2015, Sec. VII-B1, VII-B3; Fig. 1)
運算硬體host computer (model not stated)執行運算平台未標示receives sensor data via Gigabit Ethernet; no specification or timing reported(Leutenegger et al., 2015, Sec. VII-A1)
其他FPGA board of the custom visual-inertial sensor (Nikolic et al. 2014)方法輸入未標示hardware synchronisation of imagery and IMU including camera pre-triggering; optional keypoint detection; Gigabit Ethernet to the host(Leutenegger et al., 2015, Sec. VII-A1)
其他Vicon motion tracking system參考或真值量測未標示6D ground truth at 200 Hz (Vicon Loops)(Leutenegger et al., 2015, Sec. VII-A3, VII-B1; Table II)

作者報告的優勢與限制

優勢

  • Consistently outperforms the MSCKF reference filter on the same inputs (conclusion)
  • Online extrinsics calibration shown; slight miscalibration manifests as scale error in stereo (conclusion)

限制

  • Computationally more demanding than the filter baseline (conclusion)
  • Odometry without loop closure; drift is not globally corrected (inference from system scope)
  • Yaw drift is clearly present on Vicon Loops, though smaller than for the MSCKF (Sec. VII-B1)
  • The stereo version was slightly worse than the monocular one on the ETH Main Building loop, attributed to a slight stereo calibration mismatch; online extrinsics calibration removed the scale mismatch (Sec. VII-B3, VII-C1)

營建工程相關證據

論文未報告營建工地測試。量化評估資料為 Vicon 室內手持繞圈（Vicon 6D 真值）、頭盔架設的自行車軌跡（7.9 km，DGPS 真值），以及繞行 ETH 主建築外部的手持戶外迴圈（620 m，DGPS 真值）；ETH 主建築室內 470 m 跨樓層行走只以圖 1 定性展示，未量化誤差。作者指出雙目版本的外參輕微誤差會表現為尺度誤差，這與工程點雲的尺度可信度相關（推論延伸）。

原文驗證環境：受控實驗、獨立參考量測

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 16 個比較組，合計 165 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 12 組列在最後，並連到性能比較頁。

von Stumberg & Cremers, 2022 · Table II 本方法 29 筆

表格設定（擷取紀錄原文）：TUM-VI (handheld); RMSE ATE in m; other methods from the TUM-VI paper, DM-VIO median of 5 runs with SE(3) alignment; X = failure; sequence length in brackets (von Stumberg & Cremers, 2022, Table II)

RMSE ATE，TUM-VI · corridor1 (305 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 von Stumberg & Cremers, 2022 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：handheld, TUM-VI (large-scale indoor and outdoor scenes; the scene of each sequence is not described in this paper)

資料來源作者報告值(von Stumberg & Cremers, 2022, Table II)

數值與出處
方法（原文寫法）報告值出處
ROVIO (stereo)0.47 m(von Stumberg & Cremers, 2022, Table II)
VINS (VINS-Mono, mono)0.63 m(von Stumberg & Cremers, 2022, Table II)
OKVIS (stereo)本方法0.33 m(von Stumberg & Cremers, 2022, Table II)
BASALT (stereo)0.34 m(von Stumberg & Cremers, 2022, Table II)

Eckenhoff et al., 2019 · Table II 本方法 22 筆

表格設定（擷取紀錄原文）：Indirect stereo VIO; absolute RMSE averaged over 10 runs; ground-truth initialisation (Eckenhoff et al., 2019, Table II)

position RMSE，EuRoC MAV · V1 01 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。

這些是 Eckenhoff 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：均方根誤差（RMSE）；對齊方式：未對齊；單位：m；場景：public benchmark (EuRoC MAV)

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

數值與出處
方法（原文寫法）報告值出處
MODEL-1原文提出0.2522 m(Eckenhoff et al., 2019, Table II)
MODEL-2原文提出0.216 m(Eckenhoff et al., 2019, Table II)
DISCRETE0.2547 m(Eckenhoff et al., 2019, Table II)
OKVIS本方法0.2356 m(Eckenhoff et al., 2019, Table II)

Usenko et al., 2020 · Table I 本方法 20 筆

指標RMS ATE of the estimated trajectory

表格設定（擷取紀錄原文）：EuRoC MAV; RMS ATE (m) after alignment with ground truth; upper part VIO methods (pose per frame), lower part mapping methods on keyframes (KF); X = failure; V2_03 excluded (Usenko et al., 2020, Table I)

RMS ATE of the estimated trajectory，EuRoC MAV · MH_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。

這些是 Usenko et al., 2020 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：indoor MAV (machine hall and Vicon room)

資料來源作者報告值(Usenko et al., 2020, Table I)

數值與出處
方法（原文寫法）報告值出處
VI DSO, mono0.06 m(Usenko et al., 2020, Table I)
OKVIS mono本方法0.34 m(Usenko et al., 2020, Table I)
OKVIS stereo本方法0.23 m(Usenko et al., 2020, Table I)
VINS FUSION mono0.18 m(Usenko et al., 2020, Table I)
VINS FUSION stereo0.24 m(Usenko et al., 2020, Table I)
IS VIO stereo0.06 m(Usenko et al., 2020, Table I)
Proposed VIO, stereo原文提出0.07 m(Usenko et al., 2020, Table I)
VI SLAM (Kasyanov et al.) mono, KF0.25 m(Usenko et al., 2020, Table I)
VI SLAM (Kasyanov et al.) stereo, KF0.11 m(Usenko et al., 2020, Table I)
VI ORB-SLAM mono, KF0.07 m(Usenko et al., 2020, Table I)
Pure BA, stereo, KF (ablation)0.09 m(Usenko et al., 2020, Table I)
BA + Identity Factors, stereo, KF (ablation)0.08 m(Usenko et al., 2020, Table I)
Proposed VI Mapping, stereo, KF原文提出0.08 m(Usenko et al., 2020, Table I)

Geneva et al., 2020 · Table II 本方法 20 筆

表格設定（擷取紀錄原文）：EuRoC MAV Vicon-room sequences, mean ATE over ten runs per method (orientation deg / position m); VIO outputs only; V2_03 excluded; alignment method not stated (Geneva et al., 2020, Table II)

ATE position (m), mean of ten runs，EuRoC MAV · V1_01_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。

這些是 Geneva et al., 2020 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：原文未報告；單位：m；場景：indoor Vicon room, MAV

資料來源作者報告值(Geneva et al., 2020, Table II)

數值與出處
方法（原文寫法）報告值出處
mono ov slam原文提出0.058 m(Geneva et al., 2020, Table II)
mono ov vio原文提出0.076 m(Geneva et al., 2020, Table II)
mono okvis本方法0.09 m(Geneva et al., 2020, Table II)
mono rovioli (ROVIO in maplab)0.153 m(Geneva et al., 2020, Table II)
mono rvio (R-VIO)0.094 m(Geneva et al., 2020, Table II)
mono vinsfusion vio0.064 m(Geneva et al., 2020, Table II)
stereo ov slam原文提出0.061 m(Geneva et al., 2020, Table II)
stereo ov vio原文提出0.061 m(Geneva et al., 2020, Table II)
stereo basalt (VIO)0.035 m(Geneva et al., 2020, Table II)
stereo iceba (ICE-BA)0.059 m(Geneva et al., 2020, Table II)
stereo okvis本方法0.039 m(Geneva et al., 2020, Table II)
stereo smsckf (S-MSCKF)0.086 m(Geneva et al., 2020, Table II)
stereo vinsfusion vio0.054 m(Geneva et al., 2020, Table II)

其他比較組

列出其餘 12 個比較組

來源

  • Leutenegger et al., 2015

    Stefan Leutenegger, Simon Lynen, Michael Bosse, Roland Siegwart, Paul Furgale（2015）Keyframe-based visual–inertial odometry using nonlinear optimizationThe International Journal of Robotics Research, 34(3):314-334

    DOI 10.1177/0278364914554813程式碼

    同儕審查已出版已讀全文經典查證後修正

程式碼：https://github.com/ethz-asl/okvis（授權：BSD-style 3-clause (LICENSE header)）。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