On-manifold IMU preintegration
本文把兩個關鍵影格之間的大量 IMU 量測預先積分成單一相對運動約束,並提出正確處理旋轉群 SO(3) 流形結構的預積分理論,推導旋轉雜訊的性質、MAP 估計式,以及殘差、雜訊傳播與偏差事後修正的解析 Jacobian。作者將預積分 IMU 因子放入因子圖,以 iSAM2 做增量平滑,並用 structureless 視覺模型避免將三維點納入最佳化。論文亦指出其理論延伸自 Lupton 與 Sukkarieh 的原始預積分概念。
本頁內容
Develops IMU preintegration on the SO(3) manifold with analytic Jacobians and bias correction, and embeds it as factors in an iSAM2-based visual-inertial factor graph with structureless vision factors.
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 (left camera of a forward-looking VI-Sensor, 20 Hz)、IMU (ADIS16448 MEMS inside the VI-Sensor, 800 Hz) |
|---|---|
| 原文測試平台 | simulation、handheld (VI-Sensor; outdoor runs with the VI-Sensor attached to a Google Tango device and carried while walking) |
| 狀態估計 | factor-graph MAP estimation solved by iSAM2 incremental smoothing (full smoothing), preintegrated IMU factors on SO(3) with a-posteriori bias correction, structureless vision factors |
| 資料關聯 | sparse visual features tracked by the SVO front-end (sparse image alignment) |
| 時間表示 | discrete keyframe states; IMU measurements preintegrated between keyframes |
| 去畸變 | 不適用 |
| 迴圈閉合 | none (VIO); past states are not marginalized, so loop closures could be added (Sec. VIII-B1) |
| 全域最佳化 | incremental smoothing over all keyframes via iSAM2 |
| 地圖表示 | 3D landmarks eliminated by structureless model (no dense map) |
| 先驗資訊 | none |
| 可輸出幾何 | keyframe trajectory, velocities and IMU biases |
| 計算需求 | real-time; average iSAM2 back-end update about 10 ms (10 optimization iterations) and SVO front-end about 3 ms per frame on a laptop with Intel i7 2.4 GHz (Sec. VIII-B, Sec. IX); full smoothing at 100 Hz claimed (Sec. I) |
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 | 方法輸入 | 未標示 | MEMS IMU in the VI-Sensor, 800 Hz | (Forster et al., 2017a, Sec. VIII-B2) |
| 相機 | VI-Sensor | 方法輸入 | 未標示 | forward-looking; two embedded WVGA monochrome cameras, only the left camera used; 20 Hz | (Forster et al., 2017a, Sec. VIII-B2) |
| 運算硬體 | standard laptop (Intel i7, 2.4 GHz) | 執行運算平台 | 未標示 | Intel i7, 2.4 GHz | (Forster et al., 2017a, Sec. VIII-B2) |
| 其他 | Vicon system | 參考或真值量測 | 未標示 | mounted in the room; hand-eye calibration to the camera by least squares | (Forster et al., 2017a, Sec. VIII-B2) |
| 其他 | Google Tango Peanut sensor (mapper version 3.15) | 比較對象設備 | 未標示 | engineered VIO device; VI-Sensor rigidly attached to it for the outdoor runs | (Forster et al., 2017a, Sec. VIII-B3) |
作者報告的優勢與限制
優勢
- Avoids committing to a linearization point during integration compared with global-frame integration approaches (Sec. IX)
- Analytic Jacobians and a-posteriori bias correction (abstract; Appendix)
- iSAM2 accuracy is practically the same as batch optimization at roughly constant update time in simulation (Sec. VIII-A1)
- The estimator is consistent: the average NEES over 50 Monte Carlo runs stays below the chi-square upper bound (Sec. VIII-A2)
- On the 430 m indoor run, average drift over 360 m is 0.3 m versus 0.7 m for OKVIS and MSCKF (Sec. VIII-B2)
- On the 160 m multi-floor run, end-to-end error is 0.5 m versus 1.4 m for Google Tango (Sec. VIII-B3)
限制
- Authors note fixed-lag smoothing window length is hard to set for guaranteed performance, motivating full smoothing (Sec. I)
- Integration assumes constant orientation between IMU samples; higher-order integration is suggested for slower IMUs (Sec. V)
- Biases are assumed constant between keyframes (Sec. VI, Eq. 34)
- Compared systems use different front-ends, and the Tango comparison is only qualitative because the sensors differ (Sec. VIII-B2; Sec. VIII-B3)
- (inference) Evaluated for monocular VIO; LiDAR-inertial use is reported only in other works
營建工程相關證據
原文未報告。實驗包含 Vicon 室內、辦公建物周邊與多樓層建物內的手持軌跡(既有建物,非營建工地)[Sec. VIII-B];Le Gentil 等人指出許多 LiDAR-慣性估計亦依賴預積分概念[Le Gentil et al., 2020, Sec. I],但本文本身未在營建場域驗證,也未評估點雲幾何。
原文驗證環境:模擬、受控實驗、獨立參考量測、已完工建築
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 13 個比較組,合計 98 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 9 組列在最後,並連到性能比較頁。
Eckenhoff et al., 2019 · Table II 本方法 22 筆
表格設定(擷取紀錄原文):Indirect stereo VIO; absolute RMSE averaged over 10 runs; ground-truth initialisation (Eckenhoff et al., 2019, Table II)
position RMSE,EuRoC MAV · V1 01 easy
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 Eckenhoff et al., 2019 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(Eckenhoff et al., 2019, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| MODEL-1原文提出 | 0.2522 m | (Eckenhoff et al., 2019, Table II) |
| MODEL-2原文提出 | 0.216 m | (Eckenhoff et al., 2019, Table II) |
| DISCRETE本方法 | 0.2547 m | (Eckenhoff et al., 2019, Table II) |
| OKVIS | 0.2356 m | (Eckenhoff et al., 2019, Table II) |
Eckenhoff et al., 2019 · Table IV 本方法 22 筆
表格設定(擷取紀錄原文):Direct stereo VINS (iSAM2, loop closures); absolute RMSE averaged over 10 runs; ground-truth initialisation (Eckenhoff et al., 2019, Table IV)
position RMSE,EuRoC MAV · V1 01 easy
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 Eckenhoff et al., 2019 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(Eckenhoff et al., 2019, Table IV)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| MODEL-1原文提出 | 0.2445 m | (Eckenhoff et al., 2019, Table IV) |
| MODEL-2原文提出 | 0.2482 m | (Eckenhoff et al., 2019, Table IV) |
| DISCRETE本方法 | 0.253 m | (Eckenhoff et al., 2019, Table IV) |
Le Gentil et al., 2020 · Table II 本方法 16 筆
表格設定(擷取紀錄原文):Simulated fast trajectories, 100 trials, fixed query rates of 1 to 20 Hz; average absolute error of the preintegrated measurement (the paper calls it absolute pose error); units as printed: mrad and mm for 1D rotations, 'rad' and mm for 3D rotations (Le Gentil et al., 2020, Table II)
average absolute position error of preintegrated measurement,simulated IMU trajectories · 1D rotations, query rate 20 Hz
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 Le Gentil et al., 2020 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(Le Gentil et al., 2020, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| PM (on-manifold preintegration [14])本方法 | 5.61 mm | (Le Gentil et al., 2020, Table II) |
| UPM (upsampled preintegration [17]) | 0.6 mm | (Le Gentil et al., 2020, Table II) |
| GPM原文提出 | 0.02 mm | (Le Gentil et al., 2020, Table II) |
Le Gentil et al., 2020 · Table III 本方法 16 筆
指標average computation time per preintegrated measurement
表格設定(擷取紀錄原文):Average computation time over 50 trials for different integration interval lengths; IMU 100 Hz, UPM upsampled to 1 kHz; for UPM and GPM the hyper-parameter training time and the inference time are listed separately (Le Gentil et al., 2020, Table III)
average computation time per preintegrated measurement,simulated IMU trajectories · 1D rotations, interval 0.05 s
這張表在此指標與資料序列只列出本方法一筆,沒有可並列的其他方法,因此不畫圖,數值與出處見下表。這是 Le Gentil et al., 2020 在此表設定下報告的數值(author-reported results),不代表方法在其他資料或設定下的表現。
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| PM (on-manifold preintegration [14])本方法硬體:laptop with Intel i5-6300U CPU at 2.40 GHz and 24 GiB RAM; single-thread Matlab code, not optimised | 0.8 ms | (Le Gentil et al., 2020, Table III) |
其他比較組
列出其餘 9 個比較組
- Le Gentil et al., 2020 · Table I
- Yang et al., 2020a · Table 6
- Forster et al., 2017a · Text Sec.VIII-B2 (timing)
- Eckenhoff et al., 2019 · Text Sec.VII-A2 (Gore Hall)
- Eckenhoff et al., 2019 · Text Sec.VII-A2 (Smith Hall)
- Forster et al., 2017a · Text Sec.VIII-A1
- Forster et al., 2017a · Text Sec.VIII-B2 (drift)
- Forster et al., 2017a · Text Sec.VIII-B3 (multi-floor)
- Forster et al., 2017a · Text Sec.VIII-B3 (outdoor loop)
來源
Forster et al., 2017a
(2017)On-Manifold Preintegration for Real-Time Visual–Inertial OdometryIEEE Transactions on Robotics, 33(1):1-21
DOI 10.1109/tro.2016.2597321arXiv 1512.02363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
- 會議版:IMU Preintegration on Manifold for Efficient Visual-Inertial Maximum-a-Posteriori Estimation (RSS 2015) 10.15607/RSS.2015.XI.006
- 預印本:arXiv:1512.02363 (v1 2015-12-08; v3 2016-10-30) https://arxiv.org/abs/1512.02363
- 程式碼釋出:Preintegrated IMU and structureless vision factors in GTSAM 4.0 https://github.com/borglab/gtsam
程式碼:https://github.com/borglab/gtsam(授權:BSD (GTSAM))。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。