OpenVINS
OpenVINS 是以研究平台定位的開源視覺慣性估測程式庫，核心為流形上的滑動視窗 EKF（MSCKF），採用首次估計 Jacobian（FEJ）維持一致性，並可把部分特徵作為 SLAM 地標保留在狀態中。系統支援相機內參、相機與 IMU 外參及時間偏移的線上校正，以型別化索引系統自動管理狀態與共變異數，另附以 SE(3) B-spline 產生量測的模擬器與軌跡評估工具。作者在模擬與 EuRoC 資料上與多個開源 VIO 比較，顯示其精度具競爭力。
本頁內容
OpenVINS is an open, documented research platform whose core is an on-manifold FEJ-MSCKF sliding-window filter with optional SLAM landmarks and online intrinsic, extrinsic and time-offset calibration, plus a B-spline visual-inertial simulator and evaluation toolbox.
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 or stereo camera (arbitrary number of cameras supported)、IMU |
|---|---|
| 原文測試平台 | simulation (B-spline trajectories, 10 Hz camera, 400 Hz IMU)、UAV (EuRoC MAV Vicon room sequences) |
| 狀態估計 | modular on-manifold EKF over a sliding window of stochastic IMU pose clones (MSCKF) with First-Estimates Jacobians; optional SLAM landmarks kept in the state and initialised by QR (Givens) splitting of the linearised system; type-based index system manages state and covariance (Secs. II, III-A, III-B, V-A) |
| 資料關聯 | sparse visual feature tracking with an OpenCV-based front end; features within the window used in nullspace-projected MSCKF updates; SLAM landmarks in several parameterisations (global 3D, inverse MSCKF, full inverse depth, anchored 3D); errors on raw pixels to allow intrinsic calibration (Secs. II, III-C, V-B) |
| 時間表示 | discrete IMU propagation; single camera-IMU time offset estimated online (Secs. II-A, III-D) |
| 去畸變 | 不適用 |
| 迴圈閉合 | none (VIO only; authors note a pose-graph optimiser could be appended, Sec. V-B) |
| 全域最佳化 | none |
| 地圖表示 | sliding window of IMU pose clones plus a bounded set of SLAM landmarks (up to 50 in the simulation setup) as 3D points in the state |
| 先驗資訊 | none; calibration initial values can be poor and are refined online (Sec. V-A) |
| 可輸出幾何 | IMU pose trajectory with covariance, online camera intrinsics, camera-IMU extrinsics and time offset, and sparse landmark positions; no dense map |
| 計算需求 | single-threaded Intel Xeon E3-1505M v6 at 3.00 GHz: 2.7x and 4.3x real time for monocular SLAM and VIO, 1.2x and 1.9x for stereo SLAM and VIO on the first EuRoC sequence; per-frame timing not rigorously evaluated and limited by the OpenCV front end and SLAM feature updates (Sec. V-B) |
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 | 資料集感測器 | EuRoC MAV | MEMS IMU, 200 Hz | (Geneva et al., 2020, Sec. V-B) |
| 雙目相機 | EuRoC MAV stereo camera (model not reported in this paper) | 資料集感測器 | EuRoC MAV | 20 Hz stereo images | (Geneva et al., 2020, Sec. V-B) |
| 運算硬體 | Intel Xeon E3-1505M v6 @ 3.00GHz | 執行運算平台 | 未標示 | single-threaded execution | (Geneva et al., 2020, Sec. V-B) |
作者報告的優勢與限制
優勢
- With online calibration the estimator stays consistent from bad initial calibration (ATE 0.139 m, position NEES 2.007), whereas disabling calibration with a bad guess gives 508.719 m ATE and often diverges (Table I; Sec. V-A)
- Monocular OpenVINS with SLAM landmarks had the lowest average position ATE (0.079 m) among monocular methods on five EuRoC Vicon-room sequences (Table II)
- Adding SLAM landmarks greatly reduces monocular drift (Sec. V-B; Table II)
- Monocular variant best in RPE among compared open-source codes; stereo variant second to Basalt (Table III; Sec. V-B)
- Documentation, simulator and evaluation toolbox are treated as main contributions (Secs. III-E, IV)
限制
- Per-frame runtime not rigorously evaluated; Basalt was faster, and OpenVINS was limited by its OpenCV front end and SLAM feature update (Sec. V-B)
- SLAM landmarks bring a smaller gain in the stereo case (Sec. V-B)
- Only the VIO part is evaluated; no loop closure or mapping back end (Sec. V-B)
- Real-world evaluation limited to EuRoC Vicon-room sequences, with V2_03 excluded because some methods could not run on it (Tables II-III)
營建工程相關證據
論文未報告營建工地測試；實測只用 EuRoC 室內 Vicon 房間序列，並有模擬實驗。其線上內外參與時間偏移校正、開源程式與模擬器，對施工現場以低成本相機與 IMU 組裝感測器並評估 VIO 一致性有參考價值；系統只輸出軌跡與稀疏地標，需另接建圖模組才能產生點雲（推論）。
原文驗證環境：模擬、公開基準
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
Geneva et al., 2020
（2020）OpenVINS: A Research Platform for Visual-Inertial Estimation2020 IEEE International Conference on Robotics and Automation (ICRA), pp. 4666-4672
DOI 10.1109/icra40945.2020.9196524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
- 程式碼釋出：rpng/open_vins (ov_core, ov_eval, ov_msckf) with documentation at docs.openvins.com https://github.com/rpng/open_vins
程式碼：https://github.com/rpng/open_vins（授權：GPL-3.0 (LICENSE file)）。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。