ORB-SLAM
ORB-SLAM 以同一組 ORB 特徵同時支援追蹤、局部建圖、重定位（relocalization）與迴圈閉合（loop closure），分成三個平行執行緒。系統以共視圖（covisibility graph）限定局部 BA 範圍，偵測到迴圈後估計相似變換 Sim(3) 以校正單眼尺度漂移，再於稀疏的 Essential Graph 上做位姿圖最佳化。寬鬆建立、嚴格剔除關鍵影格與地圖點的策略使地圖只在場景內容改變時成長。
本頁內容
ORB-SLAM uses one ORB feature type for tracking, mapping, relocalization and loop closing, with covisibility-limited local BA and Sim(3) Essential-Graph optimisation to correct monocular scale drift.
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 |
|---|---|
| 原文測試平台 | handheld、wheeled UGV、vehicle |
| 狀態估計 | motion-only BA in tracking, local BA in mapping, pose-graph optimisation over Sim(3) constraints on the Essential Graph; Levenberg-Marquardt in g2o (Sec. III-B) |
| 資料關聯 | ORB (oriented multi-scale FAST with 256-bit descriptor): FAST corners on 8 scale levels (factor 1.2), 1000 corners for 512x384 to 752x480 images and 2000 for KITTI 1241x376, spread by a per-level grid; constant-velocity prediction with guided search of last-frame points, then projection of a covisibility-based local map (keyframes sharing points plus their neighbours) with viewing-angle and scale-range checks; DBoW2 bag-of-words place recognition with an offline ORB vocabulary, covisibility-grouped scores and all matches above 75% of the best score |
| 時間表示 | discrete poses (keyframes) |
| 去畸變 | 不適用 |
| 迴圈閉合 | Per keyframe, BoW candidates scoring above the lowest score among covisible neighbours (theta_min 30) are kept and a loop is accepted only after three consecutive consistent candidates; a Sim(3) is estimated from 3D-3D ORB matches with Horn's method inside RANSAC and refined with guided matching, serving as geometric validation; duplicated points are fused and covisibility edges added; then Sim(3) pose-graph optimisation on the Essential Graph (spanning tree, covisibility edges with theta_min 100, loop edges; 10 LM iterations in the experiments) |
| 全域最佳化 | Sim(3) pose-graph optimisation over the Essential Graph after each loop; the running system performs no full BA. In an offline test, 20 LM iterations of full BA at the end of each sequence slightly improved loopy KITTI trajectories (KITTI 00: 6.68 to 5.33 m RMSE) with negligible effect on open ones; on KITTI 09, full BA alone converged poorly (48.77 m before loop closing, 18.82 m after 100 iterations) whereas Essential Graph optimisation reached 8.36 to 8.95 m |
| 地圖表示 | sparse map points plus keyframes with covisibility graph |
| 先驗資訊 | No prior map; assumes calibrated camera intrinsics (keypoints undistorted when a distortion model is given) and an ORB visual vocabulary trained offline from a large image set |
| 可輸出幾何 | keyframe trajectory and sparse 3D map points (up to scale in monocular mode) |
| 計算需求 | Intel Core i7-4700MQ (4 cores at 2.40 GHz) with 8 GB RAM, no GPU; the three threads run alongside ROS, so medians over several runs are reported. On NewCollege (512x382, 20 fps) the tracking thread takes a median 30.57 ms per frame (about 25 to 30 Hz) and local mapping a median 383.59 ms per keyframe, dominated by local BA (median 296.08 ms); each of the six loop corrections takes 0.51 to 4.69 s in total |
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 |
|---|---|---|---|---|---|
| LiDAR | Velodyne laser scanner (KITTI ground truth, model not stated) | 參考或真值量測 | KITTI odometry | used with GPS for the ground truth | (Mur-Artal et al., 2015, Sec. VIII-E) |
| GNSS 接收器 | GPS (KITTI ground truth, model not stated) | 參考或真值量測 | KITTI odometry | ground truth from GPS and a Velodyne laser scanner | (Mur-Artal et al., 2015, Sec. VIII-E) |
| 相機 | KITTI camera (model not stated) | 資料集感測器 | KITTI odometry | 10 fps, 1241x376 | (Mur-Artal et al., 2015, Sec. V-A, VIII-E) |
| 雙目相機 | NewCollege stereo camera (model not stated) | 資料集感測器 | NewCollege | 20 fps, 512x382; processed as monocular input | (Mur-Artal et al., 2015, Sec. VIII-A) |
| RGB-D 相機 | TUM RGB-D benchmark camera (model not stated) | 資料集感測器 | TUM RGB-D | hand-held indoor sequences; images used as monocular input | (Mur-Artal et al., 2015, Sec. VIII, VIII-B) |
| 載具平台 | robot traversing a campus and adjacent parks (NewCollege) | 資料集感測器 | NewCollege | 2.2 km sequence with several loops and fast rotations | (Mur-Artal et al., 2015, Sec. VIII-A) |
| 載具平台 | car (KITTI) | 資料集感測器 | KITTI odometry | driven around a residential area; 11 sequences | (Mur-Artal et al., 2015, Sec. VIII-E) |
| 運算硬體 | Intel Core i7-4700MQ | 執行運算平台 | 未標示 | 4 cores at 2.40 GHz, 8 GB RAM; no GPU | (Mur-Artal et al., 2015, Sec. VIII) |
| 其他 | external motion capture system (model not stated) | 參考或真值量測 | TUM RGB-D | provides TUM RGB-D ground truth | (Mur-Artal et al., 2015, Sec. VIII-B) |
作者報告的優勢與限制
優勢
- Evaluated on 27 sequences from popular datasets (abstract)
- Wide-baseline relocalization and loop closing with automatic initialisation (abstract)
- Real-time on CPU without GPU (Sec. IX-A)
限制
- Monocular map is sparse; authors list denser reconstruction as future work (Sec. IX-C)
- Accuracy reported after aligning scale with ground truth, i.e. monocular scale is not metric (Sec. IX-A)
- Authors acknowledge direct methods are more robust to blur and low texture (Sec. IX-B)
- Refuses to initialise on fr3_nstr_tex_far, a planar scene with twofold ambiguity (Sec. VIII-B, Table III)
- Fails on KITTI 01, a highway with few trackable close objects (Sec. VIII-E, Table V)
- Without loops (KITTI 08) scale drift is not corrected; error about 5% of the trajectory dimension (Sec. VIII-E, Fig. 12)
- The large NewCollege loop traversed in opposite directions was not detected by place recognition and does not align (Sec. VIII-A, Fig. 6)
- Points at infinity (without sufficient parallax) are not used in tracking (Sec. IX-C)
營建工程相關證據
論文未報告營建工地測試；資料集涵蓋室內（TUM RGB-D）、戶外車載（KITTI）與校園機器人（NewCollege）。單眼尺度需另行對齊，對工程量測不足以單獨提供公制幾何（推論）。
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 19 個比較組，合計 150 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 15 組列在最後，並連到性能比較頁。
Mur-Artal et al., 2015 · Table V 本方法 33 筆
表格設定（擷取紀錄原文）：KITTI odometry, keyframe trajectory RMSE (m), median of 5 executions, Sim(3) alignment; right columns after 20 LM iterations of full BA at the end of the sequence; X = sequence 01 not processed (Mur-Artal et al., 2015, Table V)
RMSE (m), ORB-SLAM keyframe trajectory，KITTI odometry · 00
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 Mur-Artal et al., 2015 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。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| ORB-SLAM本方法原文提出 | 6.68 m | (Mur-Artal et al., 2015, Table V) |
Forster et al., 2017b · Table I 本方法 22 筆
指標absolute translation error (RMSE)
表格設定（擷取紀錄原文）：EuRoC; absolute translation error RMSE of keyframe positions after least-squares translation and scale alignment, averaged over five runs; loop closure deactivated for ORB-SLAM and LSD-SLAM; ORB-SLAM and DSO values taken from the DSO paper [42] with and without enforced real-time execution; x = tracking failed (version of record Table I) (Forster et al., 2017b, Table I)
absolute translation error (RMSE)，EuRoC · Machine Hall 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。
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 Forster et al., 2017b 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。
資料來源作者報告值(Forster et al., 2017b, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| SVO (stereo)原文提出 | 0.08 m | (Forster et al., 2017b, Table I) |
| SVO (stereo, edgelets)原文提出 | 0.08 m | (Forster et al., 2017b, Table I) |
| SVO (stereo, edgelets + prior)原文提出 | 0.04 m | (Forster et al., 2017b, Table I) |
| SVO (stereo, bundle adjustment)原文提出 | 0.04 m | (Forster et al., 2017b, Table I) |
| SVO (monocular)原文提出 | 0.17 m | (Forster et al., 2017b, Table I) |
| SVO (monocular, edgelets)原文提出 | 0.17 m | (Forster et al., 2017b, Table I) |
| SVO (monocular, edgelets + prior)原文提出 | 0.1 m | (Forster et al., 2017b, Table I) |
| SVO (monocular, bundle adjustment)原文提出 | 0.06 m | (Forster et al., 2017b, Table I) |
| ORB-SLAM (monocular, no loop-closure)本方法 | 0.02 m | (Forster et al., 2017b, Table I) |
| ORB-SLAM (monocular, no loop, real-time)本方法 | 0.61 m | (Forster et al., 2017b, Table I) |
| DSO (monocular) | 0.05 m | (Forster et al., 2017b, Table I) |
| DSO (monocular, real-time) | 0.05 m | (Forster et al., 2017b, Table I) |
| LSD-SLAM (monocular, no loop-closure) | 0.18 m | (Forster et al., 2017b, Table I) |
Tateno et al., 2017 · Table 1 本方法 20 筆
表格設定（擷取紀錄原文）：Percentage of key-frame depth values within 10% of ground-truth depth (accuracy and density together) (Tateno et al., 2017, Table 1)
Perc. Correct Depth (error < 10%)，ICL-NUIM · office0
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。
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 Tateno et al., 2017 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。
資料來源作者報告值(Tateno et al., 2017, Table 1)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| CNN-SLAM (Our Method)原文提出 | 19.41% | (Tateno et al., 2017, Table 1) |
| LSD-BS [4] (LSD-SLAM bootstrapped with ground-truth depth) | 0.603% | (Tateno et al., 2017, Table 1) |
| LSD [4] (LSD-SLAM) | 0.335% | (Tateno et al., 2017, Table 1) |
| ORB [20] (ORB-SLAM)本方法 | 0.018% | (Tateno et al., 2017, Table 1) |
| Laina [16] (CNN depth fed to point-based fusion) | 17.194% | (Tateno et al., 2017, Table 1) |
| Remode [23] (REMODE, poses from LSD-BS) | 4.479% | (Tateno et al., 2017, Table 1) |
Mur-Artal et al., 2015 · Table III 本方法 16 筆
指標Absolute KeyFrame Trajectory RMSE
表格設定（擷取紀錄原文）：TUM RGB-D keyframe ATE RMSE (cm), median over 5 executions; ORB-SLAM, PTAM (two manually chosen initial frames) and LSD-SLAM (first 10 keyframes discarded) aligned with Sim(3); RGBD-SLAM trajectories from the benchmark website aligned with SE(3), Sim(3) value in brackets; X = tracking lost (Mur-Artal et al., 2015, Table III)
Absolute KeyFrame Trajectory RMSE，TUM RGB-D · fr1_xyz
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。
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 Mur-Artal et al., 2015 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。
資料來源作者報告值(Mur-Artal et al., 2015, Table III)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| ORB-SLAM本方法原文提出 | 0.9 cm | (Mur-Artal et al., 2015, Table III) |
| PTAM | 1.15 cm | (Mur-Artal et al., 2015, Table III) |
| LSD-SLAM | 9 cm | (Mur-Artal et al., 2015, Table III) |
其他比較組
列出其餘 15 個比較組
- Mur-Artal et al., 2015 · Table VI
- Yang et al., 2020a · Table 4
- Yang et al., 2020a · Table 6
- Mur-Artal et al., 2015 · Table II
- Deng & Gan, 2026 · Table 11
- Mur-Artal et al., 2015 · Table I
- Mur-Artal et al., 2015 · Table IV
- Asadi et al., 2018 · Table 2
- Forster et al., 2017b · Table II
- Lipson et al., 2024 · Table 3
- Lipson et al., 2024 · Table 4
- Teed & Deng, 2021 · Table 1
- Teed & Deng, 2021 · Table 3
- Ghadimzadeh Alamdari et al., 2025 · Table 2
- Campos et al., 2021 · Table II
來源
Mur-Artal et al., 2015
（2015）ORB-SLAM: A Versatile and Accurate Monocular SLAM SystemIEEE Transactions on Robotics, 31(5):1147-1163
DOI 10.1109/tro.2015.2463671arXiv 1502.00956程式碼
同儕審查已出版已讀全文經典
相關版本
- 預印本：ORB-SLAM: a Versatile and Accurate Monocular SLAM System (arXiv v1-v2) https://arxiv.org/abs/1502.00956
- 程式碼釋出：ORB_SLAM https://github.com/raulmur/ORB_SLAM
程式碼：https://github.com/raulmur/ORB_SLAM（授權：GPLv3 (LICENSE.txt 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。