Temporal basis functions (continuous-time batch)
作者指出離散時間估計在 IMU、捲簾快門相機或掃描式雷射等高頻感測器下，需為每個量測時間加入位姿變數，使狀態維度過大。本文把完整的 MAP 估計移到連續時間，在高斯假設下推導目標函數，並以少量時間基底函數的係數作為待估狀態，以批次 Gauss-Newton 法求解，再由系統矩陣的逆取得共變異數。論文以三次 B 樣條的矩陣形式實作，旋轉以 Cayley-Gibbs-Rodrigues 參數表示，程式碼已併入 Kalibr 校正工具箱。實驗包含相機與 IMU 外參校正（模擬 1000 次，以及以 Vicon 為參考的實測資料）與捲簾快門相機對圓點圖板的定位，作者並指出節點數量與配置仍是未解問題。
本頁內容
Formulates full MAP estimation in continuous time, representing the trajectory by coefficients of temporal basis functions (cubic B-splines with Cayley-Gibbs-Rodrigues rotations, released in Kalibr), solved by batch Gauss-Newton and validated on camera-IMU calibration and rolling-shutter localization against Vicon.
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 (Point Grey Research Bumblebee XB3, 24 cm baseline) for camera-IMU calibration、IMU (MicroStrain 3DM-GX2)、rolling-shutter camera (Matrix Vision BlueCougar-X102d)、global-shutter camera (Matrix Vision BlueCougar-X012b) used for comparison |
|---|---|
| 原文測試平台 | simulation、handheld (sensor head moved by hand over point landmarks; camera rig waved in front of a dot-pattern board) |
| 狀態估計 | batch MAP estimation in continuous time under Gaussian assumptions; state represented by coefficients of cubic B-spline basis functions (Cayley-Gibbs-Rodrigues rotation parameters) and solved by Gauss-Newton; covariance recovered from the inverse of the Gauss-Newton system matrix; white-noise Gaussian-process motion prior acts as regularizer (Secs. 3-6) |
| 資料關聯 | known correspondences: point landmarks observed by the camera, with initial camera poses from OpenCV solvePnP (Sec. 6.2); known circle-grid positions on the pattern board (Sec. 7.1); no data-association search is described |
| 時間表示 | continuous-time state (cubic B-spline, uniform knots) with discrete-time measurements, each evaluated at its own timestamp; rolling-shutter measurement time inferred from the image row and the line delay (Eq. 79) |
| 去畸變 | no LiDAR deskew; rolling-shutter distortion handled by evaluating the spline at the capture time of each image row (Sec. 7.1); sweeping lasers discussed only as related work (Sec. 2) |
| 迴圈閉合 | none (batch calibration and localization experiments) |
| 全域最佳化 | full batch Gauss-Newton over the whole dataset (up to about 2 min 20 s; limited to four iterations in the real-data timing test) (Sec. 6.3.2) |
| 地圖表示 | sparse point landmarks estimated in calibration (priors on three landmarks); known dot-grid positions in rolling-shutter localization; no dense map |
| 先驗資訊 | priors on the positions of three landmarks for observability; a-priori guesses for gravity, landmark positions and the camera-IMU transform (Secs. 6.1-6.2); known pattern-board geometry (Sec. 7.1) |
| 可輸出幾何 | continuous IMU or camera pose trajectory (spline) with covariance, camera-IMU transform, gravity, IMU bias splines and landmark positions (Secs. 6.1, 6.2, 7.1) |
| 計算需求 | offline batch; simulation: three or four Gauss-Newton iterations took about 12 s per trial; real data (1639 stereo images, 14,211 IMU measurements) converged in about 26 s with 300 knots on a MacBook Pro (2.66 GHz Core 2 Duo, 4 GB RAM), dominated by the linear-system solve (Sec. 6.3) |
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） | MicroStrain 3DM-GX2 | 方法輸入 | 未標示 | 原文未報告 | (Furgale et al., 2015, Fig. 4 caption) |
| 相機 | Matrix Vision BlueCougar-X102d | 方法輸入 | 未標示 | CMOS rolling-shutter camera, 1280 x 960 grayscale at 15 frames per second | (Furgale et al., 2015, Fig. 9 caption; Sec. 7.1) |
| 相機 | Matrix Vision BlueCougar-X012b | 比較對象設備 | 未標示 | CMOS global-shutter camera, 1280 x 960 grayscale at 15 frames per second, synchronized with the rolling-shutter camera | (Furgale et al., 2015, Fig. 9 caption; Sec. 7.1) |
| 雙目相機 | Point Grey Research Bumblebee XB3 | 方法輸入 | 未標示 | wide-baseline (24 cm) stereo cameras | (Furgale et al., 2015, Fig. 4 caption; Sec. 6.3.2) |
| 運算硬體 | MacBook Pro | 執行運算平台 | 未標示 | 2.66 GHz Core 2 Duo, 4 GB of 1067 MHz DDR3 RAM | (Furgale et al., 2015, Sec. 6.3.2) |
| 其他 | Vicon motion capture system | 參考或真值量測 | 未標示 | tracked the sensor head and landmarks; 200 Hz timestamps used for rolling-shutter evaluation | (Furgale et al., 2015, Fig. 4 caption; Sec. 6.3.2; Fig. 13 caption) |
作者報告的優勢與限制
優勢
- Keeps state size tractable for high-rate sensors: 2 min 20 s of data represented with a few hundred knots instead of about 190,200 discrete time-varying states (abstract; Sec. 6.3.2)
- Estimated calibration uncertainty matches the error histogram over 1000 simulation trials (Sec. 6.3.1)
- The continuous-time model gives a significantly better rolling-shutter motion estimate than the discrete-time model (Sec. 7.3.2)
- Derivation structured by increasingly specific assumptions to enable other continuous-time estimators (abstract)
限制
- Choice of basis and the number and placement of knots remain open; uniform knots can underfit or overfit (Sec. 7.3.1; Sec. 7.4; Sec. 8)
- Without a motion model, extra knots overfit rolling-shutter data or let the spline oscillate between global-shutter images (Sec. 7.3.1)
- Rolling-shutter estimates degrade under extreme angular motion compared with global shutter (Sec. 7.3.2)
- Spline evaluation and Jacobians add computation and produce less sparse block-banded systems than discrete time (Sec. 7.4)
- Adaptive knot selection has only been shown offline; online operation is future work (Sec. 7.4; Sec. 8)
- Knot selection for spline-based continuous-time methods remains unresolved according to Talbot et al., 2025 (Sec. V-A)
營建工程相關證據
原文未報告。實驗為模擬，以及 Toronto 大學航太研究所（UTIAS）以 Vicon 動作捕捉為參考的手持相機與 IMU 外參校正，另有捲簾快門相機對圓點圖板的定位；未使用 LiDAR，也未評估點雲幾何。作者雖以旋轉式雷射作為連續時間估計的動機，但本文沒有以雷射資料驗證。
原文驗證環境：模擬、受控實驗、獨立參考量測
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 2 個比較組，合計 3 筆紀錄。
Furgale et al., 2015 · Text Sec.6.3.1 本方法 2 筆
資料集與序列simulation
表格設定（擷取紀錄原文）：Simulated 60 s sinusoidal trajectory, 1000 trials (120 monocular images, 5950 IMU measurements, 300 pose basis functions) (Furgale et al., 2015, Text Sec.6.3.1)
time to run Gauss-Newton to convergence (three or four iterations) per trial, approximately，simulation
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 Furgale et al., 2015 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 |
|---|---|---|
| continuous-time batch estimator (B-spline)本方法原文提出 | 12 s | (Furgale et al., 2015, Sec. 6.3.1) |
Furgale et al., 2015 · Text Sec.6.3.2 本方法 1 筆
指標time to iterate to convergence with 300 knots, approximately
資料集與序列UTIAS hand-held calibration dataset
表格設定（擷取紀錄原文）：Real calibration dataset of about 2 min 20 s (1639 stereo images, 14,211 IMU measurements), 15 bias knots, pose spline with 300 knots, iterated to convergence (Furgale et al., 2015, Text Sec.6.3.2)
time to iterate to convergence with 300 knots, approximately，UTIAS hand-held calibration dataset
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 Furgale et al., 2015 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 |
|---|---|---|
| continuous-time batch estimator (B-spline)本方法原文提出硬體：MacBook Pro, 2.66 GHz Core 2 Duo, 4 GB RAM | 26 s | (Furgale et al., 2015, Sec. 6.3.2) |
來源
Furgale et al., 2015
（2015）Continuous-time batch trajectory estimation using temporal basis functionsThe International Journal of Robotics Research, 34(14):1688-1710
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：Continuous-time batch estimation using temporal basis functions (ICRA 2012, pp. 2088-2095; authors Furgale, Barfoot, Sibley) 10.1109/ICRA.2012.6225005