MSCKF 2.0
論文比較兩類以 EKF 為基礎的視覺慣性里程計（VIO）：狀態含特徵點的 EKF-SLAM，以及只保留滑動視窗位姿的 MSCKF，並以蒙地卡羅模擬顯示 MSCKF 在精度、一致性與運算量上都較佳。作者證明兩者的線性化模型都讓繞重力軸的偏航角看似可觀測，濾波器因此低估不確定度。MSCKF 2.0 以閉式 IMU 誤差狀態轉移矩陣，並在計算 Jacobian 時固定使用位置與速度的首次估計值，恢復正確的不可觀測子空間；同時把相機與 IMU 外參放入狀態線上估計。
本頁內容
MSCKF 2.0 restores the correct 4-DOF unobservable subspace of EKF-based VIO by evaluating Jacobians at first (propagated) estimates with a closed-form IMU transition matrix, and estimates camera-IMU extrinsics online, giving a consistent real-time monocular VIO that beat EKF-SLAM variants and a fixed-lag smoother.
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、IMU |
|---|---|
| 原文測試平台 | vehicle (camera and IMU on a car roof; 21.5 km urban drive in Riverside, CA)、simulation built from real vehicle datasets (5.5 km urban dataset; 29 km Cheddar Gorge dataset) |
| 狀態估計 | multi-state-constraint EKF: IMU state plus a sliding window of past IMU poses; feature errors removed by left-nullspace projection; Jacobians evaluated at first (propagated) estimates of position and velocity so the linearised model keeps a 4-DOF unobservable subspace; closed-form IMU error-state transition matrix; camera-to-IMU rotation and translation in the state (Secs. 3.3, 5, 7) |
| 資料關聯 | Shi-Tomasi corners matched by normalised cross-correlation in the real experiment (about 290 features per image); each feature used once its track ends, triangulated by Gauss-Newton; Mahalanobis gating at the 95th percentile of chi-square with 2N-3 degrees of freedom (Secs. 3.3, 9) |
| 時間表示 | discrete poses at image times; IMU propagation assumes linearly varying signals between samples, orientation by 4th-order Runge-Kutta and position and velocity by Simpson integration (Sec. 5.2) |
| 去畸變 | 不適用 |
| 迴圈閉合 | none (VIO; loop closure is left to a separate algorithm, Sec. 3.2) |
| 全域最佳化 | none |
| 地圖表示 | none; features are never kept in the state vector, only a sliding window of poses |
| 先驗資訊 | none; manually measured camera-IMU extrinsics used only as initial values for online estimation (Sec. 9) |
| 可輸出幾何 | IMU pose, velocity, biases and camera-IMU extrinsics with covariance; no map or point cloud |
| 計算需求 | 10 ms per update including image processing on a Core i7 at 2.66 GHz with single-threaded C++ in the 20 Hz real-world run (Sec. 9); estimator-only 0.93 ms per update in simulation versus 1.54 to 4.45 ms for EKF-SLAM variants (Sec. 3.4) |
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） | Xsens MTi-G歸入：Xsens MTI-G | 方法輸入 | 未標示 | 100 Hz | (Li & Mourikis, 2013, Sec. 9) |
| 慣性量測單元（IMU） | ISIS IMU | 資料集感測器 | 5.5 km urban vehicle dataset used to generate simulations | 原文未報告 | (Li & Mourikis, 2013, Sec. 3.4) |
| 慣性量測單元（IMU） | Xsens IMU (model not reported) | 資料集感測器 | Cheddar Gorge dataset | 100 Hz | (Li & Mourikis, 2013, Sec. 8.2) |
| GNSS 接收器 | GPS-INS (model not reported) | 參考或真值量測 | 未標示 | position ground truth | (Li & Mourikis, 2013, Sec. 9) |
| 雙目相機 | PointGrey Bumblebee2 | 方法輸入 | 未標示 | stereo pair, only one camera's images used, stored at 20 Hz | (Li & Mourikis, 2013, Sec. 9) |
| 載具平台 | car (roof-mounted IMU and camera) | 方法輸入 | 未標示 | about 21.5 km in 37 min | (Li & Mourikis, 2013, Sec. 9) |
| 運算硬體 | Intel Core i7 2.66 GHz | 執行運算平台 | 未標示 | single-threaded C++ implementation | (Li & Mourikis, 2013, Secs. 3.4, 9) |
作者報告的優勢與限制
優勢
- In the Cheddar Gorge simulation MSCKF 2.0 reached 97.7 m position RMSE and NEES 6.53 versus 133.4 m and 50.97 for an iterative fixed-lag smoother that used about five times more computation (Table 3; Sec. 8.2)
- Performance nearly identical to an 'ideal' MSCKF that uses true states for Jacobians (Table 2; Sec. 8.1)
- Online camera-IMU calibration gave pose accuracy almost equal to perfectly known calibration (Table 4; Sec. 8.3)
- Real 21.5 km drive: largest position error about 58 m (0.28% of distance) versus about 230 m (MSCKF) and 202 m (FLS) (Sec. 9)
限制
- Odometry only: loop closure and re-observation of revisited areas are not handled (Sec. 3.2)
- Position and yaw remain unobservable, so their uncertainty (and drift) grows over time even for MSCKF 2.0 (Sec. 8.2)
- Discrete-time IMU integration needs assumptions about signal behaviour between samples, so some approximation is unavoidable (Sec. 5.2)
- Evaluated on vehicle motion only; no handheld, indoor or aerial tests (inference from Secs. 3.4, 8, 9)
營建工程相關證據
論文未報告營建工地或室內測試；真實實驗為車頂相機與 IMU 在市區行駛 21.5 km，並以 GPS-INS 為參考，其餘結果為依真實資料建構的蒙地卡羅模擬。其價值在於說明濾波式 VIO 的一致性與相機 IMU 外參線上估計，對現場以 VIO 提供位姿先驗時如何解讀共變異數有參考意義（推論）。
原文驗證環境：模擬、獨立參考量測
報告的性能數據
性能數據仍在分批查證，目前尚未收錄此方法的報告值。
來源
Li & Mourikis, 2013
（2013）High-precision, consistent EKF-based visual-inertial odometryThe International Journal of Robotics Research, 32(6):690-711
同儕審查已出版已讀全文經典查證後修正
相關版本
- 會議版：Improving the accuracy of EKF-based visual-inertial odometry (ICRA 2012, Li and Mourikis). The IJRR paper says its observability result first appeared in an earlier conference version, Li and Mourikis (2012a); the reference-list entry (ICRA, St Paul, pp. 828-835) matches this Crossref record 10.1109/ICRA.2012.6225229
- 前身方法：A Multi-State Constraint Kalman Filter for Vision-aided Inertial Navigation (ICRA 2007) 10.1109/ROBOT.2007.364024