ROVIO
ROVIO 是單目視覺慣性里程計，把影像塊的像素強度誤差直接當作 EKF 更新的創新項，而非使用特徵點重投影誤差。整個濾波狀態採機器人中心（robocentric）表示，地標以方位向量加上反距離參數化，並以最小維度的流形差分表示不確定度，因此特徵在第二次觀測時即可使用，系統不需額外初始化程序。多層影像塊在影像金字塔上追蹤，強度誤差經 QR 分解降維以維持即時運算，同時線上估計 IMU 偏差與相機外參。
本頁內容
ROVIO is a monocular VIO that feeds photometric errors of multilevel image patches directly into a robocentric EKF, with landmarks parameterised by bearing vector and inverse distance, enabling undelayed feature use, power-up-and-go operation and online extrinsic calibration in real time on one CPU core.
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 | handheld (VI-Sensor; slow and fast motion with motion-capture ground truth)、UAV (multirotor with forward-looking visual-inertial sensor; filter output used for flight control) |
| 狀態估計 | EKF with a fully robocentric state (position, velocity, attitude, IMU biases, camera-IMU extrinsics, and per feature a bearing vector plus inverse-distance parameter) using minimal boxplus differences on SO(3) and S2; linearisation point of uncertain features improved by a patch search; an iterated EKF is named as an alternative (Secs. II-A, II-C) |
| 資料關聯 | direct: multilevel 8 x 8 pixel patches on a 4-level image pyramid tracked inside the filter; mean-subtracted intensity errors reduced by QR decomposition to a 2D innovation per feature; affine patch warping from two extra bearing vectors; FAST candidates ranked by a multi-level Shi-Tomasi score with bucketing; Mahalanobis outlier rejection (Secs. II-C, III) |
| 時間表示 | discrete; continuous IMU-driven dynamics discretised with Euler forward integration; one update per image (IMU 200 Hz, images 20 Hz in the experiments) (Secs. II-B, IV-A) |
| 去畸變 | 不適用 |
| 迴圈閉合 | none |
| 全域最佳化 | none |
| 地圖表示 | up to 50 robocentric landmarks (bearing vector plus inverse distance) with multilevel patches held in the filter state (Sec. IV-A) |
| 先驗資訊 | none; camera intrinsics assumed known (factory calibration); IMU-camera extrinsics roughly guessed with translation set to zero and estimated online (Sec. IV-A) |
| 可輸出幾何 | IMU pose, robocentric velocity, IMU biases and camera-IMU extrinsics with covariance; sparse landmark estimates; no dense map |
| 計算需求 | single core of an Intel i7-2760QM: 6.65 ms per image with 10 features up to 29.72 ms with 50 features (Table I); run on board a UAV with 50 features at 20 Hz (Sec. V) |
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） | ADIS 16448歸入：ADIS16448 | 方法輸入 | 未標示 | industrial grade, angular random walk 0.66 deg/sqrt(Hz), velocity random walk 0.11 m/s/sqrt(Hz), 200 Hz | (Bloesch et al., 2015, Sec. IV-A) |
| 雙目相機 | VI-Sensor (two global-shutter wide-VGA 1/3 inch imagers) | 方法輸入 | 未標示 | fronto-parallel stereo, lenses with 120 deg diagonal field of view, factory calibrated pinhole plus radial-tangential model, hardware time-synchronised to the IMU with mid-exposure triggering, 20 Hz; only one camera used | (Bloesch et al., 2015, Sec. IV-A) |
| 載具平台 | multirotor UAV (model not reported) | 方法輸入 | 未標示 | forward-oriented visual-inertial sensor; filter output used for feedback control | (Bloesch et al., 2015, Sec. IV-D) |
| 運算硬體 | Intel i7-2760QM | 執行運算平台 | 未標示 | single core | (Bloesch et al., 2015, Sec. IV-B; Table I) |
| 其他 | external motion capture system (model not reported) | 參考或真值量測 | 未標示 | pose ground truth; some bad tracking filtered out in the fast-motion run | (Bloesch et al., 2015, Secs. IV-A, IV-C) |
作者報告的優勢與限制
優勢
- No initialisation procedure needed because of the robocentric inverse-distance landmark parametrisation (abstract; Sec. II-A)
- Tracks hand-held motion with mean rotation rate about 3.5 rad/s and peaks up to 8 rad/s against motion capture (Sec. IV-C)
- Relative position error in a 1 min slow-motion run tends to be similar to, and often slightly better than, a batch optimisation framework along the lines of OKVIS (Sec. IV-B, Fig. 3; values only plotted)
- Accuracy nearly independent of the number of tracked features above 20 (Sec. IV-B)
- Output used directly in the control loop of a multirotor UAV from take-off to landing with calibration converging during take-off (Sec. IV-D)
限制
- A divergence mode exists when the velocity estimate diverges (missing motion or many outliers): feature distances are pushed to infinity, which removes the corrective effect (Sec. IV-C)
- Yaw is unobservable and drifts slowly (Sec. IV-C)
- The IMU-camera translation needs much rotational motion to converge and the accelerometer bias converges slowly (Sec. IV-C)
- A position offset against motion capture was observed on the UAV and attributed mainly to online calibration (Sec. IV-D)
- Evaluation limited to short motion-capture sequences; quantitative comparison only in plots (Sec. IV; inference)
營建工程相關證據
論文未報告營建工地測試；手持實驗與 UAV 飛行都在動作捕捉系統範圍內進行並以其為參考。直接法影像塊與強度誤差回饋在低紋理或快速運動下的穩健性，對施工現場手持或無人機巡檢的位姿估計有參考價值，但系統只輸出稀疏地標，無法直接產生工程點雲（推論）。
原文驗證環境：受控實驗、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
Bloesch et al., 2015
（2015）Robust visual inertial odometry using a direct EKF-based approach2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 298-304
DOI 10.1109/iros.2015.7353389程式碼
同儕審查已出版已讀全文經典查證後修正
相關版本
- 期刊延伸版：Iterated extended Kalman filter based visual-inertial odometry using direct photometric feedback (IJRR 36(10):1053-1072, 2017; Bloesch, Burri, Omari, Hutter, Siegwart); metadata checked in Crossref, content not read 10.1177/0278364917728574
- 程式碼釋出：ethz-asl/rovio (open-source C++ implementation referenced as [2] in the paper) https://github.com/ethz-asl/rovio
程式碼：https://github.com/ethz-asl/rovio（授權：BSD-style (LICENSE file: Copyright (c) 2014, Autonomous Systems Lab, redistribution permitted with conditions)）。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。