MINS
MINS 以 IMU 為核心,在一個 MSCKF 形式的擴展卡爾曼濾波器中緊耦合相機、輪速計、LiDAR 與 GNSS:每種感測器都有專屬的量測更新,並能線上校正所有感測器的外參、時間偏移與內參。面對非同步量測,系統以高階流形上多項式內插取得任一時刻的位姿並建立內插誤差模型,且依運動動態調整複製狀態的頻率,以兼顧精度與計算量。LiDAR 採用直接點到平面更新並維護 ikd 樹局部地圖。
本頁內容
An IMU-centric MSCKF-style filter that tightly fuses cameras, wheel odometry, LiDAR and GNSS with online spatiotemporal and intrinsic calibration, using high-order on-manifold interpolation with an error model and dynamic cloning to handle asynchronous sensors efficiently.
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 | IMU、cameras (one or more)、wheel encoders、LiDAR (one or more)、GNSS |
|---|---|
| 原文測試平台 | car (KAIST Urban dataset)、Clearpath Husky UGV (UD Husky dataset)、simulation |
| 狀態估計 | MSCKF-style EKF with the IMU as the backbone; camera updates with nullspace projection and measurement compression; integrated 2D wheel odometry updates with wheel intrinsics; direct LiDAR point-on-plane updates against an ikd-tree local map anchored to a cloned pose; GNSS position updates after 4-DoF alignment, then estimation in the global frame; online spatiotemporal and intrinsic calibration of all sensors (Sec. 3) |
| 資料關聯 | visual feature tracks for MSCKF updates; LiDAR points matched to planes in a dense local map (FAST-LIO2-style point-on-plane); wheel and GNSS measurements used directly (Sec. 3) |
| 時間表示 | asynchronous measurements handled by high-order on-manifold polynomial interpolation between stochastic clones (default third order, 20 Hz cloning) with an interpolation error model; dynamic cloning adapts the clone rate to motion (Sec. 4) |
| 去畸變 | not described; each point cloud is treated as a measurement at a single time t_k whose pose is obtained through the on-manifold interpolation (Secs. 3.4, 4) |
| 迴圈閉合 | no |
| 全域最佳化 | none; GNSS provides global correction when available |
| 地圖表示 | dense LiDAR local map (ikd-tree) anchored to a clone; no global map |
| 先驗資訊 | initial calibration guesses; IMU-wheel dynamic initialization (Sec. 5) |
| 可輸出幾何 | IMU pose, velocity and calibration states at the IMU rate |
| 計算需求 | timing on a ThinkPad P17 laptop with an Intel i7 and 32 GB RAM (Table 8); simulation run times single-threaded on an Intel i7 (Table 5) |
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 | 16-channel Velodyne (two units) | 方法輸入 | KAIST Urban | 10 Hz each; merged into a synthetic 20 Hz LiDAR for single-LiDAR baselines | (Lee et al., 2025b, Sec. 7) |
| LiDAR | 64-channel Ouster | 方法輸入 | UD Husky dataset | 10 Hz | (Lee et al., 2025b, Sec. 7) |
| 慣性量測單元(IMU) | MTi-300歸入:Xsens MTi-300 | 方法輸入 | KAIST Urban | 100 Hz | (Lee et al., 2025b, Sec. 7) |
| GNSS 接收器 | EVK-7P | 方法輸入 | KAIST Urban | 10 Hz | (Lee et al., 2025b, Sec. 7) |
| GNSS 接收器 | GPS 18x | 方法輸入 | UD Husky dataset | 1 Hz | (Lee et al., 2025b, Sec. 7) |
| GNSS 接收器 | Reach M+ (RTK) | 參考或真值量測 | UD Husky dataset | 5 Hz; outdoor ground truth | (Lee et al., 2025b, Sec. 7) |
| 雙目相機 | Flea3 stereo | 方法輸入 | KAIST Urban | 10 Hz | (Lee et al., 2025b, Sec. 7) |
| 雙目相機 | RealSense T265歸入:Intel RealSense T265 | 方法輸入 | UD Husky dataset | two 30 Hz cameras and a 200 Hz IMU | (Lee et al., 2025b, Sec. 7) |
| 輪式或腿式里程計 | LM13 | 方法輸入 | KAIST Urban | 100 Hz | (Lee et al., 2025b, Sec. 7) |
| 輪式或腿式里程計 | Husky wheel encoders | 方法輸入 | UD Husky dataset | 10 Hz | (Lee et al., 2025b, Sec. 7) |
| 載具平台 | Husky | 方法輸入 | UD Husky dataset | UGV for the UD Husky dataset | (Lee et al., 2025b, Sec. 7; Figure 15) |
| 運算硬體 | ThinkPad P17 | 執行運算平台 | KAIST Urban | Intel i7, 32 GB RAM | (Lee et al., 2025b, Table 8) |
| 其他 | OptiTrack | 參考或真值量測 | UD Husky dataset | indoor ground truth | (Lee et al., 2025b, Sec. 7) |
作者報告的優勢與限制
優勢
- On UD Husky indoor I1 and I2, MINS(I,L) ATE 0.07 and 0.11 m, lower than FAST-LIO2 (0.18, 0.16) and LIW-OAM (0.12, 0.16) (Table 7)
- With all sensors, MINS(I,C,L,W,G) ATE 0.96 to 3.03 m on outdoor and unstructured sequences where GNSS alone gives 2.17 to 15.71 m and several baselines exceed 5 m (Table 7)
- Total time per update 47 ms with all sensors versus 146 ms for Lvio-Fusion and 484 ms for VINS-Fusion with GNSS on KAIST Urban 38 (Table 8)
- In simulation the full combination gives the lowest RMSE (0.261 deg / 0.050 m) with NEES under 4 for all combinations (Table 5)
- Dynamic cloning with coefficient 1 cuts run time to 61.9 % of fixed 30 Hz cloning with a small accuracy loss (Table 4)
限制
- Fusing all sensors is not always the most accurate choice; optimal sensor weighting is future work (Sec. 8)
- No loop closure or global map
- KAIST single-LiDAR baselines were fed a synthetic 20 Hz LiDAR merged from two 16-channel units (Sec. 7)
營建工程相關證據
MINS 可同時融合輪速計、LiDAR、相機與 GNSS,並線上校正時間與外參,這與施工現場巡檢載具常見的多感測器配置相符(推論)。UD Husky 資料包含室內結構化與室外非結構化環境,並以動作捕捉與 RTK 為真值;不過論文只評估軌跡,沒有點雲地圖精度。
原文驗證環境:公開基準、模擬、獨立參考量測、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 6 個比較組,合計 97 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 2 組列在最後,並連到性能比較頁。
Lee et al., 2025b · Table 7 本方法 38 筆
指標Average (5 runs) position ATE
表格設定(擷取紀錄原文):UD Husky dataset (Clearpath Husky; indoor structured I1-I2 with OptiTrack GT, outdoor O1-O2 and unstructured T1-T4 with Emlid RTK GT); average (5 runs) position ATE in metres; '-' = error above 5 m not reported (Lee et al., 2025b, Table 7)
Average (5 runs) position ATE,UD Husky dataset · O1
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)
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 Lee et al., 2025b 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(Lee et al., 2025b, Table 7)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| VINS-Fusion(V) | 無數值未報告註記(擷取紀錄):error above 5 m, not reported ('-') | (Lee et al., 2025b, Table 7; Sec. 7) |
| VINS-Fusion(L) | 無數值未報告註記(擷取紀錄):error above 5 m, not reported ('-') | (Lee et al., 2025b, Table 7; Sec. 7) |
| ORB-SLAM3 | 4.51 m | (Lee et al., 2025b, Table 7; Sec. 7) |
| MINS(I,C)本方法原文提出 | 2.93 m | (Lee et al., 2025b, Table 7; Sec. 7) |
| MINS(I,C,W)本方法原文提出 | 1.39 m | (Lee et al., 2025b, Table 7; Sec. 7) |
| FAST-LIO2 | 2.34 m | (Lee et al., 2025b, Table 7; Sec. 7) |
| LIW-OAM | 1.01 m | (Lee et al., 2025b, Table 7; Sec. 7) |
| MINS(I,L)本方法原文提出 | 1.73 m | (Lee et al., 2025b, Table 7; Sec. 7) |
| MINS(I,L,W)本方法原文提出 | 1.35 m | (Lee et al., 2025b, Table 7; Sec. 7) |
| GNSS | 3.17 m | (Lee et al., 2025b, Table 7; Sec. 7) |
| VINS-Fusion(G) | 無數值未報告註記(擷取紀錄):error above 5 m, not reported ('-') | (Lee et al., 2025b, Table 7; Sec. 7) |
| Lvio-Fusion | 3.14 m | (Lee et al., 2025b, Table 7; Sec. 7) |
| MINS(I,C,L,W,G)本方法原文提出 | 0.96 m | (Lee et al., 2025b, Table 7; Sec. 7) |
Lee et al., 2025b · Table 5 本方法 24 筆
資料集與序列MINS simulation · simulated trajectory
表格設定(擷取紀錄原文):Simulation with different sensor combinations (I IMU, C camera, G GNSS, W wheel, L LiDAR); orientation and position RMSE (mean over 10 runs, +- std in the table) and run time; single-threaded Intel i7 (Lee et al., 2025b, Table 5)
Orientation RMSE,MINS simulation · simulated trajectory
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 Lee et al., 2025b 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(Lee et al., 2025b, Table 5)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| MINS(I,C)本方法原文提出 | 0.505 deg | (Lee et al., 2025b, Table 5; Sec. 6) |
| MINS(I,G)本方法原文提出 | 1.244 deg | (Lee et al., 2025b, Table 5; Sec. 6) |
| MINS(I,W)本方法原文提出 | 3.053 deg | (Lee et al., 2025b, Table 5; Sec. 6) |
| MINS(I,L)本方法原文提出 | 0.474 deg | (Lee et al., 2025b, Table 5; Sec. 6) |
| MINS(I,C,G)本方法原文提出 | 0.318 deg | (Lee et al., 2025b, Table 5; Sec. 6) |
| MINS(I,C,W)本方法原文提出 | 0.505 deg | (Lee et al., 2025b, Table 5; Sec. 6) |
| MINS(I,C,L)本方法原文提出 | 0.414 deg | (Lee et al., 2025b, Table 5; Sec. 6) |
| MINS(I,C,G,W,L)本方法原文提出 | 0.261 deg | (Lee et al., 2025b, Table 5; Sec. 6) |
Lee et al., 2025b · Table 4 本方法 18 筆
資料集與序列MINS simulation · simulated trajectory
表格設定(擷取紀錄原文):Simulation: pose RMSE (deg / m) and total computation time (s) of fixed-rate 30 Hz cloning versus dynamic cloning with threshold coefficients 0.01 to 100 (Lee et al., 2025b, Table 4)
Ori. RMSE,MINS simulation · simulated trajectory
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 Lee et al., 2025b 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(Lee et al., 2025b, Table 4)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| MINS (fixed 30 Hz cloning)本方法原文提出 | 0.178 deg | (Lee et al., 2025b, Table 4; Sec. 6) |
| MINS (dynamic, coefficient 0.01 cloning)本方法原文提出 | 0.178 deg | (Lee et al., 2025b, Table 4; Sec. 6) |
| MINS (dynamic, coefficient 0.1 cloning)本方法原文提出 | 0.175 deg | (Lee et al., 2025b, Table 4; Sec. 6) |
| MINS (dynamic, coefficient 1 cloning)本方法原文提出 | 0.218 deg | (Lee et al., 2025b, Table 4; Sec. 6) |
| MINS (dynamic, coefficient 10 cloning)本方法原文提出 | 0.38 deg | (Lee et al., 2025b, Table 4; Sec. 6) |
| MINS (dynamic, coefficient 100 cloning)本方法原文提出 | 0.55 deg | (Lee et al., 2025b, Table 4; Sec. 6) |
Nubert et al., 2026 · Table V 本方法 8 筆
表格設定(擷取紀錄原文):ANYmal autonomous hikes (Forest and Mountain/Seealpsee); global ATE [m] and ARE [deg] against post-processed ground truth (offline HF batch optimization with GNSS); HF World = world frame, HF Odom = smooth odometry frame (Sec. IV-D3); LR = LiDAR registration; MINS diverged mid-way on Mountain ('div.'), 'split' excludes the divergence (Nubert et al., 2026, Table V)
ATE [m],ANYmal hike missions (authors) · Forest
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 Nubert 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。
資料來源作者報告值(Nubert et al., 2026, Table V)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| TSIF - Odom | 33.38 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| Open3D SLAM - LR | 1.46 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| MINS - World (div.)本方法 | 0.44 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| MINS - World (split)本方法 | 0.44 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| HF GNSS+IMU - World原文提出 | 0.53 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| HF GNSS+IMU - Odom原文提出 | 101.55 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| HF (LR-between) - World原文提出 | 0.52 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| HF (LR-between) - Odom原文提出 | 66 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| HF - World原文提出 | 0.42 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| HF - Odom原文提出 | 22.91 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| HF (GNSS filtered) - World原文提出 | 0.49 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
| HF (GNSS filtered) - Odom原文提出 | 15.94 m | (Nubert et al., 2026, Table V; Sec. VI-C1) |
其他比較組
來源
Lee et al., 2025b
(2025)MINS: Efficient and Robust Multisensor-Aided Inertial Navigation SystemJournal of Field Robotics, 42(7), pp. 3252-3284
DOI 10.1002/rob.22546arXiv 2309.15390程式碼
同儕審查已出版已讀全文近十年查證後修正
相關版本
- 預印本:MINS arXiv v1 (2023) https://arxiv.org/abs/2309.15390
- 程式碼釋出:rpng/MINS (GPL-3.0) https://github.com/rpng/MINS
程式碼:https://github.com/rpng/MINS(授權:GPL-3.0 (GitHub license metadata))。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。