ICP covariance (Censi)
作者以 ICP 最小化的誤差函數為對象,利用隱函數定理推導估計值對量測的一階敏感度,得到封閉形式共變異數,並考慮同一量測被多個對應重複使用與量測彼此相關的情形。論文只處理二維平面(x、y、θ)的定位與掃描匹配,分析對象為點對線段 ICP;以 52 條射線、雜訊標準差 0.03 m 的模擬測距儀,在正方形、走廊與圓形環境各做 300 次蒙地卡羅模擬;在走廊與圓形等約束不足情形,改以 Fisher 資訊矩陣找出不可觀測方向,只比較可觀測子空間上的誤差。
本頁內容
Closed-form first-order ICP covariance via the implicit function theorem (uses d2J/dz dx, not only the Hessian), derived and tested for 2D (x, y, theta) localization and point-to-segment scan matching in Monte Carlo simulation, with Fisher-information analysis for corridor and circular under-constrained cases.
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 | simulated 2D range finder: 52 rays over 360 deg, zero-mean Gaussian range noise with 0.03 m standard deviation |
|---|---|
| 原文測試平台 | simulation |
| 狀態估計 | first-order covariance of the minimizer via the implicit function theorem: cov(x_hat) = (d2J/dx2)^-1 (d2J/dz dx) cov(z) (d2J/dz dx)^T (d2J/dx2)^-1 evaluated at the estimate; uses only the error function J, not the ICP algorithm internals; requires matrix products and a 3x3 inversion |
| 資料關聯 | analysed ICP variant is 2D point-to-segment ('vanilla' ICP): each correspondence uses one point of the current scan and the two reference-scan points that form the closest polyline segment |
| 時間表示 | 不適用 |
| 去畸變 | 不適用 |
| 迴圈閉合 | none |
| 全域最佳化 | none |
| 地圖表示 | 不適用 |
| 先驗資訊 | measurement covariance cov(z), which may be a full (correlated) matrix; for localization the map is treated as noise-free (z = y_t); an optional odometry term in a MAP cost (Eq. 10) bounds the covariance by cov(u) |
| 可輸出幾何 | covariance of the ICP pose estimate |
| 計算需求 | evaluating Eq. 6 needs matrix multiplications and inversion of a 3x3 matrix; described as negligible relative to ICP itself; no timing figures or hardware reported |
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 |
|---|---|---|---|---|---|
| 其他 | simulated 2D range finder (52 rays distributed on 360 deg) | 方法輸入 | 未標示 | 52 rays distributed on 360 deg; zero-mean Gaussian range noise with 0.03 m standard deviation | (Censi, 2007, Sec. V) |
作者報告的優勢與限制
優勢
- square environment, scan matching: predicted std 7.7 mm, 7.7 mm, 0.060 deg vs Monte Carlo 7.6 mm, 7.8 mm, 0.058 deg, while the Hessian method gives 20.0 mm, 20.3 mm, 0.171 deg (Fig. 4 table)
- localization: prediction close to the Cramer-Rao bound and to the sample covariance (Fig. 4 table)
- accounts for measurements shared by several correspondences and for correlated measurement noise (Sec. IV)
限制
- only planar 2D localization and scan matching (x, y, theta) are derived and tested; no 3D formulation is given (Sec. I, IV, V)
- validation limited to Monte Carlo simulation of three synthetic environments; no real sensor data (Sec. V)
- assumes ICP converged inside the basin of the true solution; wrong convergence is not modelled (Sec. I-B)
- models only the error due to sensor noise; the Cramer-Rao bound approximates localization well but is optimistic for scan matching (Sec. I-B)
- moderately optimistic in the circular environment on the observable manifold (Fig. 4 summary table)
- closed-form derivatives are not given in the paper (Sec. IV)
營建工程相關證據
原文未報告;論文指出走廊是實務上最常見的約束不足情形(Sec. V),與長廊、隧道等營建場景的關聯及對工程點雲不確定性傳遞的意義均屬推論
原文驗證環境:模擬
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 6 個比較組,合計 31 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 2 組列在最後,並連到性能比較頁。
Brossard et al., 2020 · Table 1 本方法 8 筆
資料集與序列Challenging data sets for point cloud registration (Pomerleau et al. 2012) · average of 8 sequences
表格設定(擷取紀錄原文):ICP covariance consistency averaged over 8 sequences, 1020 registrations x 1000 initializations; Q_ini 0.1 m and 10 deg; noise and bias SD 5 cm; NNE target 1; starred = robust statistics after removing extreme quantiles; baselines printed as orders of magnitude (Brossard et al., 2020, Table 1)
NNE trans.,Challenging data sets for point cloud registration (Pomerleau et al. 2012) · average of 8 sequences
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 Brossard 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(Brossard et al., 2020, Table 1)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Q_censi (closed-form, Censi 2007)本方法 | 1000有附註註記(擷取紀錄):printed only as order of magnitude 10^3 | (Brossard et al., 2020, Table 1) |
| Q_monte carlo (65 Monte Carlo ICP samples) | 1000有附註註記(擷取紀錄):printed only as order of magnitude 10^3 | (Brossard et al., 2020, Table 1) |
| proposed原文提出 | 4.2 | (Brossard et al., 2020, Table 1) |
Landry et al., 2019 · Table II 本方法 7 筆
指標Avg. KL divergence
表格設定(擷取紀錄原文,這些數值分屬表中不同部分):(Landry et al., 2019, Table II)
- Average KL divergence between sampled and predicted ICP covariance per test group; trained on the named other group
- Average KL divergence per test group
Avg. KL divergence,Challenging data sets (ETH) · Apartment (trained on Haupt. and Stairs; 1190 pairs)
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 Landry 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(Landry et al., 2019, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Baseline (mean training covariance) | 34.1 | (Landry et al., 2019, Table II) |
| Ours (CELLO-3D)原文提出 | 26.6 | (Landry et al., 2019, Table II) |
| Censi本方法 | 91900000 | (Landry et al., 2019, Table II) |
Censi, 2007 · Fig. 4 table (circle, scan matching) 本方法 5 筆
資料集與序列simulation (Monte Carlo, 300 runs) · circular environment, scan matching
表格設定(擷取紀錄原文):Std of scan-matching error: Monte Carlo sample (true) vs predicted; sigma(w1), sigma(w2) are errors projected on the observable manifold O (Censi, 2007, Fig. 4 table (circle, scan matching))
sigma(x) of ICP error,simulation (Monte Carlo, 300 runs) · circular environment, scan matching
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 Censi, 2007 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(Censi, 2007, Fig. 4 table (circle, scan matching))
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| true (Monte Carlo sample) | 284 mm | (Censi, 2007, Fig. 4 (embedded table, circle)) |
| Hessian [1] | 13 mm | (Censi, 2007, Fig. 4 (embedded table, circle)) |
| Offline [1] | 378 mm | (Censi, 2007, Fig. 4 (embedded table, circle)) |
| proposed本方法原文提出 | 13 mm | (Censi, 2007, Fig. 4 (embedded table, circle)) |
Censi, 2007 · Fig. 4 table (corridor, scan matching) 本方法 5 筆
資料集與序列simulation (Monte Carlo, 300 runs) · corridor environment, scan matching
表格設定(擷取紀錄原文):Std of scan-matching error: Monte Carlo sample (true) vs predicted; sigma(w1), sigma(w2) are errors projected on the observable manifold O (Censi, 2007, Fig. 4 table (corridor, scan matching))
sigma(x) of ICP error,simulation (Monte Carlo, 300 runs) · corridor environment, scan matching
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 Censi, 2007 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(Censi, 2007, Fig. 4 table (corridor, scan matching))
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| true (Monte Carlo sample) | 38 mm | (Censi, 2007, Fig. 4 (embedded table, corridor)) |
| Hessian [1] | 33 mm | (Censi, 2007, Fig. 4 (embedded table, corridor)) |
| Offline [1] | 50 mm | (Censi, 2007, Fig. 4 (embedded table, corridor)) |
| proposed本方法原文提出 | 40 mm | (Censi, 2007, Fig. 4 (embedded table, corridor)) |
其他比較組
來源
Censi, 2007
(2007)An accurate closed-form estimate of ICP's covarianceProceedings 2007 IEEE International Conference on Robotics and Automation, pp. 3167-3172
同儕審查已出版已讀全文經典