SLAM 點雲圖鑑Point clouds for construction
M-01 · 2020

Kimera

Kimera 是模組化的開源度量語意（metric-semantic）視覺慣性 SLAM 函式庫，包含以 GTSAM iSAM2 固定延遲平滑器實作的 VIO、以 PCM 剔除錯誤迴圈的強健位姿圖最佳化、低延遲 3D 網格生成器，以及用雙目稠密匹配（SGM）與 Voxblox TSDF 產生全域語意網格的模組。各模組可獨立或組合執行，並在 CPU 上即時運作。作者以 EuRoC 地面真值點雲評估網格的精度與完整度，但評估前先以 ICP 將估計點雲對齊至真值。

具全域修正的完整 SLAM近十年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
arXiv 1910.02490v3 (2020-03-04, 8 pages, marked accepted for ICRA 2020); ICRA 2020 IEEE version of record not compared
全文取得途徑
arXiv 預印本
本頁內容

Kimera combines a GTSAM-based VIO, robust PCM-filtered pose-graph optimisation, a fast mesher and a Voxblox TSDF semantic mesh built from dense stereo, evaluating mesh accuracy and completeness against EuRoC ground-truth clouds after ICP alignment.

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monocular camera、stereo、IMU
原文測試平台UAV、simulation
狀態估計keyframe-based MAP visual-inertial estimator run as full or fixed-lag smoothing (fixed-lag typically used to bound estimation time), with on-manifold IMU preintegration and structureless vision factors solved by iSAM2 in GTSAM and marginalisation of states leaving the horizon; Kimera-RPGO keeps odometry and loop edges separately, selects the largest consistent loop set with a modified incremental PCM (odometry chi-squared check plus pairwise consistency, fast maximum clique) and optimises the pose graph with Gauss-Newton in GTSAM
資料關聯Shi-Tomasi corners tracked by Lucas-Kanade, left-right stereo matching, 5-point mono and 3-point stereo RANSAC verification (optional 2-point and 1-point variants using IMU rotation) at keyframes; structureless vision factors triangulated by DLT with degenerate and high-reprojection-error points removed; DBoW2 bag-of-words loop candidates verified with the same mono and stereo checks
時間表示discrete keyframe states; IMU-rate estimates (Sec. II)
去畸變不適用
迴圈閉合DBoW2 putative loops, geometric verification, outlier rejection with a modified PCM before GTSAM PGO (Sec. II-B)
全域最佳化robust pose-graph optimisation (Kimera-RPGO) (Sec. II-B2)
地圖表示per-frame mesh from 2D Delaunay triangulation of tracked features back-projected with VIO landmark estimates; multi-frame mesh over the VIO horizon, coupled back to VIO through regularity factors when planar surfaces are detected in the mesh; global Voxblox TSDF built at keyframes from semi-global-matching dense stereo with bundled raycasting (fast option), meshed by marching cubes, with Bayesian per-voxel semantic label updates within the truncation distance
先驗資訊2D semantic segmentation of images for labelling (Sec. II-D2)
可輸出幾何trajectory, low-latency local mesh, and global semantically annotated mesh from TSDF (Sec. II)
計算需求CPU only (model not reported), four threads; IMU preintegration about 40 us (IMU-rate estimates above 200 Hz), feature tracking 4.5 ms per frame, keyframe front-end 45 ms, per-frame mesh under 5 ms, multi-frame mesh 15 ms, VIO back-end under 40 ms, Kimera-RPGO 55 ms on average on EuRoC, Kimera-Semantics about 0.1 s per keyframe for a 720x480 depth image

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）EuRoC MAV IMU (model not reported in this paper)資料集感測器EuRoC MAVhigh-rate inertial measurements; state estimates output at IMU rate (above 200 Hz)(Rosinol et al., 2020, Sec. II; Sec. III-D)
雙目相機EuRoC MAV stereo camera (model not reported in this paper)資料集感測器EuRoC MAVstereo frames used as Kimera input (monocular mode also supported)(Rosinol et al., 2020, Sec. II; Sec. III-A)
運算硬體CPU (model not reported)執行運算平台未標示real-time CPU execution with four threads(Rosinol et al., 2020, Abstract; Sec. II; Sec. III-D)
其他EuRoC V1 and V2 ground-truth point cloud (acquisition device not stated in this paper)參考或真值量測EuRoC MAVused for mesh accuracy and completeness after ICP registration(Rosinol et al., 2020, Sec. III-B)
其他Unity-based photo-realistic simulator provided by MIT Lincoln Lab資料集感測器photo-realistic simulator (MIT Lincoln Lab)ROS sensor streams with ground-truth geometry, semantics, depth and poses; 720x480 dense depth images(Rosinol et al., 2020, Sec. III-C; Sec. III-D; Acknowledgments)

作者報告的優勢與限制

優勢

  • Modules can run in isolation or together, falling back to VIO or full SLAM (abstract)
  • On EuRoC V1 and V2 the global TSDF mesh had lower error than the multi-frame mesh on 4 of 6 sequences (slightly higher on V1_02 and V2_01), while the multi-frame mesher requires two orders of magnitude less time (Sec. III-B; Table IV)
  • With PCM, V1_01 ATE stays between 0.045 and 0.05 m for loop thresholds from 10 to 0.001, while PGO without PCM reaches up to 1.74 m (Table III)
  • VIO drift below 0.2% (4 cm over a 32 m simulated trajectory) (Sec. III-C)

限制

  • Loop closures may still contain outliers from perceptual aliasing, e.g. identical rooms on different floors (Sec. II-B1)
  • (inference) Mesh accuracy is measured after ICP registration to ground truth, so global placement error is not included in the reported metric
  • Dense stereo has difficulty resolving texture-less walls, causing the largest drop in geometric and semantic accuracy (RMSE 0.215 m vs 0.131 m with ground-truth depth; mIoU 57.23% vs 80.03%) (Sec. III-C; Table V; Fig. 4)
  • Kimera-RPGO runtime depends on pose-graph size (Sec. III-D)
  • (observation) Table IV caption calls the metric completeness while the text calls the same values average error, so the direction of the reported distance is ambiguous

營建工程相關證據

論文未報告營建測試；評估使用 EuRoC 與照片級模擬器。作者特別提到不同樓層相同房間造成的感知混淆（perceptual aliasing），與多樓層建築掃描直接相關。其幾何評估流程（取樣網格、ICP 對齊、精度與完整度）可作為工程點雲評估參考，但 ICP 對齊會隱藏絕對位置誤差（推論）；且作者報告的全域網格平均誤差為 0.35 至 0.48 m（ICP 門檻 1.0 m；Table IV 標題稱為完整度），遠大於一般工程量測容許差（推論）。作者也指出稠密立體匹配難以處理無紋理牆面，是模擬場景中幾何與語意誤差最大的來源（Sec. III-C、Table V）。

原文驗證環境：公開基準、模擬

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 10 個比較組，合計 92 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 6 組列在最後，並連到性能比較頁。

Rosinol et al., 2020 · Table II 本方法 33 筆

指標RMSE ATE [m]

表格設定（擷取紀錄原文）：EuRoC ATE RMSE grouped as fixed-lag smoothing, full smoothing and PGO with loop closure; comparator values taken from Delmerico and Scaramuzza [77] (Sim(3) alignment per text) and VINS-Mono [24]; comparators use a monocular camera while Kimera uses stereo; Kimera aligned with SE(3); loop threshold alpha = 0.001 (Rosinol et al., 2020, Table II)

RMSE ATE [m]，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 Rosinol 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：SE(3) 剛體對齊；單位：m；場景：EuRoC MAV sequences (micro aerial vehicle dataset, ref. [19]); environments not described in this paper

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

數值與出處
方法（原文寫法）報告值出處
Kimera-VIO (fixed-lag smoothing)本方法原文提出0.11 m(Rosinol et al., 2020, Table II)
Kimera-VIO (full smoothing)本方法原文提出0.04 m(Rosinol et al., 2020, Table II)
Kimera-RPGO (loop closure)本方法原文提出0.08 m(Rosinol et al., 2020, Table II)

Rosinol et al., 2020 · Table IV 本方法 12 筆

指標RMSE [m] (Table IV caption: completeness [78, Sec. 4.3.3])

表格設定（擷取紀錄原文）：Mesh evaluated against the EuRoC ground-truth point cloud: mesh sampled at 10^3 points per m2, registered by rigid ICP in CloudCompare (ICP threshold 1.0 m); caption calls the metric completeness while the text describes the same values as average error of the global mesh (0.35 to 0.48 m); Multi-Frame mesh computed with a large VIO horizon (full smoothing) (Rosinol et al., 2020, Table IV)

RMSE [m] (Table IV caption: completeness [78, Sec. 4.3.3])，EuRoC MAV (V1, V2 ground-truth point cloud) · V1_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 Rosinol 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：SE(3) 剛體對齊；單位：m；場景：EuRoC MAV V1 and V2 sequences with ground-truth point cloud (environment not described in this paper)

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

數值與出處
方法（原文寫法）報告值出處
Kimera-Mesher multi-frame mesh本方法原文提出0.482 m(Rosinol et al., 2020, Table IV)
Kimera-Semantics global TSDF mesh本方法原文提出0.364 m(Rosinol et al., 2020, Table IV)

Rosinol et al., 2020 · Table V 本方法 12 筆

資料集與序列photo-realistic simulator (MIT Lincoln Lab) · simulated scene (about 32 m trajectory, Sec. III-C)

表格設定（擷取紀錄原文）：Unity-based photo-realistic simulator (MIT Lincoln Lab) with ground-truth 2D semantics; mesh registered to ground truth and point RMSE computed as in Sec. III-B (distance direction not specified); simulated trajectory about 32 m (Rosinol et al., 2020, Table V)

mIoU [%]，photo-realistic simulator (MIT Lincoln Lab) · simulated scene (about 32 m trajectory, Sec. III-C)

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 Rosinol 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：未對齊；單位：%；場景：photo-realistic Unity-based simulated scene; semantic classes include wall, shelf and floor (Sec. III-C, Fig. 4)

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

數值與出處
方法（原文寫法）報告值出處
Kimera-Semantics with GT depth and GT poses本方法80.1%(Rosinol et al., 2020, Table V)
Kimera-Semantics with GT depth and Kimera-VIO poses本方法80.03%(Rosinol et al., 2020, Table V)
Kimera-Semantics with dense stereo and Kimera-VIO poses本方法原文提出57.23%(Rosinol et al., 2020, Table V)

Rosinol et al., 2020 · Table III 本方法 10 筆

指標RMSE ATE [m]

表格設定（擷取紀錄原文）：EuRoC V1_01 ATE RMSE versus DBoW2 loop-closure threshold alpha; smaller alpha gives more but less conservative loop closures (Rosinol et al., 2020, Table III)

RMSE ATE [m]，EuRoC MAV · V1_01, alpha=10

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 Rosinol 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：SE(3) 剛體對齊；單位：m；場景：EuRoC MAV V1_01 (environment not described in this paper)

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

數值與出處
方法（原文寫法）報告值出處
Kimera ablation: PGO w/o PCM本方法0.05 m(Rosinol et al., 2020, Table III)
Kimera-RPGO本方法原文提出0.05 m(Rosinol et al., 2020, Table III)

其他比較組

列出其餘 6 個比較組

來源

  • Rosinol et al., 2020

    Antoni Rosinol, Marcus Abate, Yun Chang, Luca Carlone（2020）Kimera: an Open-Source Library for Real-Time Metric-Semantic Localization and Mapping2020 IEEE International Conference on Robotics and Automation (ICRA), pp. 1689-1696

    DOI 10.1109/icra40945.2020.9196885arXiv 1910.02490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/MIT-SPARK/Kimera（授權：BSD (README states BSD License, LICENSE.BSD)）。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