Closed-form preintegration
本文推導 IMU 預積分方程的閉式解,而非以離散取樣近似量測動態,並提出兩種慣性模型:分段常數量測,以及分段常數的局部真實加速度。作者以 Monte Carlo 模擬分析模型選擇對估計的影響,並把此預積分分別用於緊耦合滑動視窗最佳化與鬆耦合直接影像對齊兩種視覺慣性系統。
本頁內容
Derives closed-form IMU preintegration under two inertial models and validates it in sliding-window and direct visual-inertial systems.
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 | IMU、stereo camera (both real-world systems) |
|---|---|
| 原文測試平台 | simulation、UAV (EuRoC MAV dataset)、handheld (VI-Sensor in two University of Delaware buildings) |
| 狀態估計 | indirect system: tightly coupled sliding-window optimization with marginalization using closed-form preintegration; direct system: loosely coupled direct stereo image alignment whose relative-pose factors and preintegration factors are optimized with iSAM2 (GTSAM) without marginalization |
| 資料關聯 | indirect VIO: FAST features on a uniform grid tracked with KLT, stereo correspondences by KLT from left to right image, 8-point RANSAC outlier rejection and Cauchy loss (Sec. VII-A); direct VINS: photometric alignment of high-gradient pixels against keyframes with a Huber cost, keyframe depth from OpenCV StereoSGBM (Sec. V-B; Sec. VII-B) |
| 時間表示 | discrete keyframes with analytically integrated IMU models |
| 去畸變 | 不適用 |
| 迴圈閉合 | direct VINS keeps all states (no marginalization) to allow loop closures and outperforms the indirect VIO on loop-rich EuRoC sequences (Sec. VII-B) |
| 全域最佳化 | indirect VIO: sliding-window bundle adjustment (inertial window 6 and pose window 8 on EuRoC, up to 300 features) solved with Ceres using the Schur complement, with marginalization; direct VINS: iSAM2 incremental smoothing in GTSAM over all states without marginalization, allowing loop-closure relative-pose factors; simulation: GTSAM fixed-lag smoother (Sec. VI; Sec. VII) |
| 地圖表示 | no persistent map: sparse features in inverse-depth parameterization inside the sliding window (features marginalized after 3 s in simulation); the direct VINS keeps keyframes with stereo depth maps for alignment (Sec. V-A; Sec. VI; Sec. VII-B) |
| 先驗資訊 | none |
| 可輸出幾何 | IMU trajectory (orientation, position, velocity, biases); IMU-camera extrinsics and camera intrinsics estimated online in the UD experiments; no dense map or point cloud (Sec. VII-A2) |
| 計算需求 | real-time claimed; EuRoC window sizes chosen for real-time operation with minimal dropped frames; direct image alignment implemented as a CUDA kernel for GPU acceleration (GPU model not reported); no runtime values reported (Sec. VII) |
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 | (Eckenhoff et al., 2019, Sec. VII-A1) |
| 雙目相機 | 原文未報告 | 資料集感測器 | EuRoC MAV | stereo pairs at 20 Hz | (Eckenhoff et al., 2019, Sec. VII-A1) |
| 雙目相機 | VI-Sensor | 方法輸入 | 未標示 | hand-held; IMU at 400 Hz; stereo images used by the stereo VIO; camera intrinsics and IMU-camera extrinsics estimated online | (Eckenhoff et al., 2019, Sec. VII-A2) |
| 運算硬體 | 原文未報告 (GPU running the CUDA direct-alignment kernel) | 執行運算平台 | 未標示 | 原文未報告 | (Eckenhoff et al., 2019, Sec. VII-B) |
作者報告的優勢與限制
優勢
- Closed-form solutions improve estimation accuracy compared with discrete sampling (abstract)
- In 50 Monte Carlo Gazebo MAV runs, Model 2 reached RMSE 0.093 m and 0.300 deg versus 0.107 m and 0.328 deg for discrete preintegration, without extra computational overhead (Sec. VI)
- Gains over discrete preintegration are larger at lower IMU rates, relevant for low-cost MEMS sensors (Sec. VI; Table I)
- On the 230 m UD Smith Hall run, Model 1 ending error was 0.632 m versus 1.699 m for OKVIS (Sec. VII-A2)
限制
- Proposed models improve on discrete preintegration in most but not all cases and are competitive with, not uniformly better than, OKVIS on EuRoC (Sec. VII-A; Tables II-III)
- Direct VINS is sensitive to tuning parameters and drifts more when loop closures are unavailable; some V2 03 diff runs accepted incorrect loop closures (Sec. VII-B)
- Improvement over discrete preintegration is small in the MAV simulation and shrinks at higher IMU rates (Sec. VI; Table I)
- UD indoor runs have no ground truth; only return-to-start ending errors are reported (Sec. VII-A2)
營建工程相關證據
未在營建場域驗證;真實資料為 EuRoC 與 Delaware 大學 Gore Hall(228 m,跨三層樓)及 Smith Hall(230 m,跨兩層樓)的手持 VI-Sensor 資料。後兩者沒有獨立參考真值,以回到起點的終點誤差評估,並以樓層平面圖投影呈現軌跡,未評估點雲幾何。
原文驗證環境:模擬、公開基準、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 4 個比較組,合計 92 筆紀錄。
Eckenhoff et al., 2019 · Table II 本方法 44 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本方法原文提出 | 0.2522 m | (Eckenhoff et al., 2019, Table II) |
| MODEL-2本方法原文提出 | 0.216 m | (Eckenhoff et al., 2019, Table II) |
| DISCRETE | 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 44 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本方法原文提出 | 0.2445 m | (Eckenhoff et al., 2019, Table IV) |
| MODEL-2本方法原文提出 | 0.2482 m | (Eckenhoff et al., 2019, Table IV) |
| DISCRETE | 0.253 m | (Eckenhoff et al., 2019, Table IV) |
Eckenhoff et al., 2019 · Text Sec.VII-A2 (Gore Hall) 本方法 2 筆
指標ending error (0.33% of the path)
資料集與序列UD Gore Hall
表格設定(擷取紀錄原文):228 m hand-held loop from the first floor up the staircase to the third floor and back, one loop per level; ending error at the return to the start; each preintegration model run ten times and averaged (Eckenhoff et al., 2019, Text Sec.VII-A2 (Gore Hall))
ending error (0.33% of the path),UD Gore Hall
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, Text Sec.VII-A2 (Gore Hall))
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Model 1本方法原文提出 | 0.763 m | (Eckenhoff et al., 2019, Sec. VII-A2) |
| Model 2本方法原文提出 | 0.747 m | (Eckenhoff et al., 2019, Sec. VII-A2) |
Eckenhoff et al., 2019 · Text Sec.VII-A2 (Smith Hall) 本方法 2 筆
資料集與序列UD Smith Hall
表格設定(擷取紀錄原文):230 m hand-held run over the second and first floors returning to the start, with people walking, varying lighting and feature-poor areas; ending error at the return to the start (Eckenhoff et al., 2019, Text Sec.VII-A2 (Smith Hall))
ending error (0.28% of the path),UD Smith Hall
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 Eckenhoff et al., 2019 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 |
|---|---|---|
| Model 1本方法原文提出 | 0.632 m | (Eckenhoff et al., 2019, Sec. VII-A2) |
來源
Eckenhoff et al., 2019
(2019)Closed-form preintegration methods for graph-based visual–inertial navigationThe International Journal of Robotics Research, 38(5):563-586
DOI 10.1177/0278364919835021arXiv 1805.02774
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:arXiv:1805.02774 (v1 2018-05-07; v2 2019-03-20 final) https://arxiv.org/abs/1805.02774