S-MSCKF (msckf_vio)
S-MSCKF 把多狀態約束卡爾曼濾波（MSCKF）擴充到立體相機，目標是在微型飛行器的筆電等級電腦上以低運算量穩健估計位姿。前端以 FAST 角點與 KLT 光流同時做時間追蹤與左右影像匹配，並以 2 點 RANSAC 與環狀匹配剔除離群；後端以 4 維立體量測更新，不需影像校正，並採可觀測性約束 EKF（OC-EKF）維持一致性。為平均運算負載，每隔一次更新移除兩個相機狀態。
本頁內容
S-MSCKF is an open-source stereo MSCKF VIO with a KLT-based front end for temporal and stereo matching, an observability-constrained EKF, and a steady two-state removal scheme, delivering accuracy similar to OKVIS and VINS-Mono at lower CPU load for fast MAV flight.
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 | stereo camera、IMU |
|---|---|
| 原文測試平台 | UAV (3 kg FALCON quadrotor; fast flights up to 17.5 m/s over a runway; autonomous flight through woods and a warehouse)、UAV (EuRoC MAV) |
| 狀態估計 | stereo multi-state-constraint EKF: IMU state with camera-IMU extrinsics plus a window of left-camera poses; RK4 propagation; 4-D stereo measurement that does not require rectification; nullspace projection of feature errors; observability-constrained EKF (OC-EKF) for consistency; two camera states removed every other update, chosen by a two-way keyframe rule (Sec. III) |
| 資料關聯 | FAST corners tracked temporally by KLT optical flow and matched across the stereo pair also by KLT; 2-point RANSAC for temporal outliers and circular matching between consecutive stereo pairs (Sec. III-E) |
| 時間表示 | discrete; IMU at 200 Hz, cameras at 20 Hz (EuRoC) or 40 Hz (fast flight) synchronised by the IMU trigger (Sec. IV) |
| 去畸變 | 不適用 |
| 迴圈閉合 | none |
| 全域最佳化 | none |
| 地圖表示 | none in the filter; features are marginalised by nullspace projection (MSCKF) |
| 先驗資訊 | offline camera-IMU calibration supplied in the experiments; left-right stereo extrinsics assumed known (Secs. III, IV) |
| 可輸出幾何 | IMU pose and velocity; in the field test the poses were used to register a laser point cloud (laser used for mapping only) (Sec. IV-C; Fig. 6) |
| 計算需求 | filter about 10% of one core at 20 Hz, with about 80% of computation in the front end (EuRoC); CPU load measured on NUC6i7KYK (quad-core i7-6770HQ); runs onboard an Intel NUC5i7RYH on the FALCON robot (Sec. IV-A; Fig. 1) |
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 | laser scanner on FALCON (model not reported) | 資料集感測器 | autonomous flight experiment | used for mapping only; global laser point cloud registered with S-MSCKF poses | (Sun et al., 2018, Fig. 1; Sec. IV-C; Fig. 6) |
| 慣性量測單元（IMU） | VectorNav VN-100 Rugged | 方法輸入 | fast flight dataset (KumarRobotics msckf_vio wiki) | 200 Hz | (Sun et al., 2018, Sec. IV-B) |
| GNSS 接收器 | GPS (model not reported) | 參考或真值量測 | fast flight dataset (KumarRobotics msckf_vio wiki) | x-y position reference for fast-flight RMSE | (Sun et al., 2018, Sec. IV-B) |
| 雙目相機 | PointGrey CM3-U3-13Y3M-CS (two, forward-looking) | 方法輸入 | fast flight dataset (KumarRobotics msckf_vio wiki) | 40 Hz, 960 x 800, synchronised by the IMU trigger, external auto-exposure controller applying identical shutter and gain; Sec. III-E mentions a 20 cm baseline stereo configuration when discussing KLT stereo matching, without naming the rig | (Sun et al., 2018, Secs. III-E, IV-B) |
| 雙目相機 | VI sensor (EuRoC) | 資料集感測器 | EuRoC MAV | synchronised 20 Hz stereo images and 200 Hz IMU | (Sun et al., 2018, Sec. IV-A) |
| 載具平台 | FALCON quadrotor | 方法輸入 | 未標示 | 3 kg; synchronised stereo cameras and IMU, a laser scanner and a downward-facing lidar; only stereo and IMU used for estimation | (Sun et al., 2018, Fig. 1) |
| 運算硬體 | Intel NUC5i7RYH | 執行運算平台 | 未標示 | onboard computer of the FALCON robot | (Sun et al., 2018, Fig. 1) |
| 運算硬體 | NUC6i7KYK (quad-core i7-6770HQ) | 執行運算平台 | EuRoC MAV | used to measure CPU load on EuRoC | (Sun et al., 2018, Sec. IV-A) |
作者報告的優勢與限制
優勢
- Accuracy similar to OKVIS and VINS-Mono on EuRoC (except V2_03, where S-MSCKF fails), while ROVIO has larger errors on the machine-hall sequences
- filter-based methods have the lowest CPU load (Sec. IV-A
- Fig. 2, plotted)
限制
- Diverges on EuRoC V2_03 because brightness inconsistency between the stereo images breaks KLT stereo matching (Sec. IV-A)
- Global position and yaw are unobservable, so uncertainty grows and the estimate may jump or diverge once the prior uncertainty is large (Sec. V)
- Frequent removal of camera states discards some valid observations (Sec. III-D)
- KLT stereo matching is reliable for corners deeper than about 1 m with a 20 cm baseline (Sec. III-E)
- Front-end cost rises with higher image rate and resolution and short feature lifetimes during aggressive flight (Sec. IV-B)
營建工程相關證據
論文未在營建工地測試；除 EuRoC 外，飛行實驗在機場跑道、樹林與倉庫進行，並把 S-MSCKF 位姿用來拼接雷射點雲，700 m 往返終點漂移約 3 m。這顯示低運算量立體視覺 VIO 可作為無人機在室內外轉換場域（如大型倉儲或施工中建物）的位姿來源，但未報告點雲幾何精度（推論）。
原文驗證環境：公開基準、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
Sun et al., 2018
（2018）Robust Stereo Visual Inertial Odometry for Fast Autonomous FlightIEEE Robotics and Automation Letters, 3(2):965-972
DOI 10.1109/lra.2018.2793349arXiv 1712.00036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：Robust Stereo Visual Inertial Odometry for Fast Autonomous Flight (arXiv v3, RA-L accepted version) https://arxiv.org/abs/1712.00036
- 程式碼釋出：KumarRobotics/msckf_vio (also hosts the fast-flight dataset on its wiki) https://github.com/KumarRobotics/msckf_vio
程式碼：https://github.com/KumarRobotics/msckf_vio（授權：Penn Software MSCKF_VIO licence (University of Pennsylvania): use, copy and modify for non-profit research purposes only (LICENSE.txt)）。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。