Gaussian Process Preintegration (GPM)
本文以高斯過程(GP)連續表示慣性量測,並對 GP 核函數施加線性運算子,推導出「高斯預積分量測」(GPM)。旋轉僅繞單一軸時,旋轉與速度、位置增量都可解析積分;若含三維旋轉,旋轉增量仍需先以 GP 上取樣再數值積分,速度與位置增量則由重投影到起始 IMU 座標後的加速度計訊號以 GP 推論。此法不需明確運動模型,也沒有離散積分雜訊,適合非同步的慣性輔助估計;作者並推導偏差與感測器間時間偏移的一階修正 Jacobian。模擬比較中 GPM 與 UPM 的誤差約比離散預積分低一個數量級,並整合進 LiDAR 慣性定位與建圖框架 IN2LAAMA 驗證可用性。
本頁內容
Gaussian-process preintegration models IMU signals as GPs and applies linear operators to the kernels, giving analytic preintegration over arbitrary intervals when rotation is about a single axis (3D rotations still need GP upsampling and numerical integration of Delta R), with first-order bias and inter-sensor time-shift corrections; benchmarked in simulation and validated inside the IN2LAAMA lidar-inertial framework.
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、3D LiDAR (validation within IN2LAAMA) |
|---|---|
| 原文測試平台 | simulation、handheld (Velodyne VLP-16 3D lidar + MTi3 Xsens IMU, as written) |
| 狀態估計 | preintegrated measurements from GP models (square exponential kernel) of IMU signals with linear operators on kernels: analytic integral inference for single-axis rotation, numerical integration after GP upsampling for 3D rotation; velocity and position inferred from GP models of accelerometer data reprojected into the IMU frame at t1; first-order post-integration bias and time-shift Jacobians |
| 資料關聯 | 不適用 |
| 時間表示 | continuous representation of inertial signals |
| 去畸變 | not described in this paper; the authors state that GPMs can be used with asynchronous platforms and rolling-shutter-like sensors, citing spinning lidars as such sensors (Sec. I); how IN2LAAMA [23] handles lidar motion distortion is not described here |
| 迴圈閉合 | 不適用 |
| 全域最佳化 | 不適用 |
| 地圖表示 | 不適用 |
| 先驗資訊 | none |
| 可輸出幾何 | preintegrated measurements (Delta R, Delta v, Delta p with covariance) over any time interval; within IN2LAAMA a lidar point-cloud map and trajectory are produced (Fig. 4) |
| 計算需求 | benchmarks in single-thread, non-optimised Matlab on a laptop with an Intel i5-6300U at 2.40 GHz and 24 GiB RAM; GPM inference 1.3 to 212 ms (single-axis) and 10.2 to 817 ms (3D rotation) for 0.05 to 5 s intervals, plus about 0.28 to 7.1 s of hyper-parameter training when not reused (Table III); cubic GP complexity, close to real time (Sec. VI) |
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 | Velodyne VLP-16 | 方法輸入 | 未標示 | 3D lidar | (Le Gentil et al., 2020, Sec. V-B) |
| 慣性量測單元(IMU) | MTi3 Xsens IMU歸入:Xsens MTi-3 | 方法輸入 | 未標示 | 原文未報告 | (Le Gentil et al., 2020, Sec. V-B) |
| 慣性量測單元(IMU) | simulated IMU | 資料集感測器 | simulated sinusoidal trajectories | 100 Hz; accelerometer noise sd 0.02 m/s2; gyroscope noise sd 0.002 rad/s | (Le Gentil et al., 2020, Table I caption; Table III caption) |
| 載具平台 | hand-held sensor suite | 方法輸入 | 未標示 | moved up and down while walking in the UTS lab; 6.2 m trajectory, maximum estimated velocity 1.7 m/s | (Le Gentil et al., 2020, Sec. V-B; Fig. 4) |
| 運算硬體 | Intel i5-6300U | 執行運算平台 | 未標示 | laptop; 2.40 GHz; 24 GiB RAM; single-thread Matlab code, not optimised for high performance | (Le Gentil et al., 2020, Sec. V-A3) |
作者報告的優勢與限制
優勢
- Analytical integration over any time interval; no explicit motion model and no numerical integration noise (abstract)
- Outperformed upsampled and on-manifold discrete preintegration in simulated accuracy tests (Sec. V-A; Sec. VI)
限制
- Cubic computational complexity of GPs; close-to-real-time only (Sec. VI)
- Assumes constant bias during preintegration; releasing this is future work (Sec. VI)
- GPMs are much slower for motions with 3D rotations because Delta R is computed numerically (Sec. V-A3, Table III)
- With 3D rotations the accuracy gain over UPMs is smaller because both share the numerical rotation step (Sec. V-A2)
- Query rates above about 2 Hz do not allow hyper-parameter training from scratch for each interval (Sec. VI)
- (inference) Real-data validation is a single 6.2 m hand-held run and no ground-truth reference is reported (Sec. V-B)
營建工程相關證據
未在營建場域驗證。真實資料僅為 UTS 實驗室(既有建物)內 6.2 m 的手持 VLP-16 軌跡;以人工分割平牆的平均點到平面距離比較(離散預積分 9.0 mm,GPM 6.3 mm)[Sec. V-B]。這是本群集少數直接量測點雲幾何品質的結果,但規模極小、無獨立參考,其工程意義有限(推論)。
原文驗證環境:模擬、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 5 個比較組,合計 58 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 1 組列在最後,並連到性能比較頁。
Le Gentil et al., 2020 · Table III 本方法 32 筆
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 hyper-parameter training time per preintegrated measurement,simulated IMU trajectories · 1D rotations, interval 0.05 s
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 III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| UPM (upsampled preintegration [17])硬體:laptop with Intel i5-6300U CPU at 2.40 GHz and 24 GiB RAM; single-thread Matlab code, not optimised | 294 ms | (Le Gentil et al., 2020, Table III) |
| GPM本方法原文提出硬體:laptop with Intel i5-6300U CPU at 2.40 GHz and 24 GiB RAM; single-thread Matlab code, not optimised | 356 ms | (Le Gentil et al., 2020, Table III) |
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]) | 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本方法原文提出 | 0.02 mm | (Le Gentil et al., 2020, Table II) |
Le Gentil et al., 2020 · Table I 本方法 8 筆
表格設定(擷取紀錄原文):Simulated random sinusoidal trajectories, integration interval 1 to 5 s, 100 trials; average relative error with respect to travelled linear or angular distance; IMU 100 Hz, accelerometer noise sd 0.02 m/s2, gyroscope noise sd 0.002 rad/s, UPM upsampled to 1 kHz (Le Gentil et al., 2020, Table I)
average relative rotation error of preintegrated measurement,simulated IMU trajectories · 1D rotations, slow motion
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 I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| PM (on-manifold preintegration [14]) | 0.253% | (Le Gentil et al., 2020, Table I) |
| UPM (upsampled preintegration [17]) | 0.038% | (Le Gentil et al., 2020, Table I) |
| GPM本方法原文提出 | 0.028% | (Le Gentil et al., 2020, Table I) |
Le Gentil et al., 2020 · Text Sec.V-B real 本方法 1 筆
指標average point-to-plane distance on a flat wall
資料集與序列UTS lab hand-held sequence · 6.2 m trajectory
表格設定(擷取紀錄原文):Real hand-held run (Velodyne VLP-16 and MTi3 Xsens IMU) in the UTS lab, 6.2 m trajectory, maximum estimated speed 1.7 m/s; about 150k points on a manually segmented flat wall, plane fitted by PCA in each map; no independent reference (Le Gentil et al., 2020, Text Sec.V-B real)
average point-to-plane distance on a flat wall,UTS lab hand-held sequence · 6.2 m trajectory
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, Text Sec.V-B real)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| IN2LAAMA with PM | 9 mm | (Le Gentil et al., 2020, Sec. V-B; Fig. 4) |
| IN2LAAMA with GPM本方法原文提出 | 6.3 mm | (Le Gentil et al., 2020, Sec. V-B; Fig. 4) |
其他比較組
列出其餘 1 個比較組
來源
Le Gentil et al., 2020
(2020)Gaussian Process Preintegration for Inertial-Aided State EstimationIEEE Robotics and Automation Letters, 5(2):2108-2114
同儕審查已出版已讀全文近十年查證後修正