Degeneracy factor / solution remapping
本文把退化定義為解對約束擾動的剛度,並證明線性化系統的退化因子 D 等於 AᵀA 最小特徵值加一,對應的特徵向量即為最退化的方向。方法以門檻判定退化方向(門檻取自一組同時含良好與退化場景的樣本資料,設在兩群間隔的中點),再以解重映射(solution remapping)在退化方向保留預測值、只在條件良好方向更新,可作為 Levenberg-Marquardt 等求解器的外掛步驟,額外複雜度為 O(kn² + n³)。作者以 uEye 相機與馬達旋轉的 Hokuyo UTM-30LX 組成的手持視覺光達系統,在走廊、平坦地面以及 538 m 室內外路線測試,終點位置誤差為行進距離的 0.71%。
本頁內容
Defines a degeneracy factor D = lambda_min + 1 of A^T A (stiffness of the solution under a constraint disturbance), thresholds eigenvalues to find degenerate directions, and applies solution remapping that keeps the prediction in degenerate directions and updates only well-conditioned ones; demonstrated on a handheld camera and rotating-lidar system with 0.71% end error over 538 m.
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 (uEye monochrome, 60 Hz, 752 x 480, 76 deg horizontal FOV)、custom 3D lidar (Hokuyo UTM-30LX rotated by a motor, 0.25 deg encoder) |
|---|---|
| 原文測試平台 | handheld (sensor pack carried by a person walking at 0.5 m/s) |
| 狀態估計 | Plug-in to linear or nonlinear least-squares solvers: eigen-decomposition of A^T A at the first nonlinear iteration, eigenvalues below an empirical threshold mark degenerate directions, and solution remapping updates only well-conditioned directions (Algorithm 1); demonstrated inside three Levenberg-Marquardt modules (frame-to-frame visual odometry, sweep-to-sweep refinement, sweep-to-map registration) |
| 資料關聯 | 不適用 (generic to optimization-based estimation; demonstrated with scan matching and visual constraints) |
| 時間表示 | 不適用 to the degeneracy method itself; the host vision-lidar system models visual odometry drift with constant velocity within each 1 s lidar sweep |
| 去畸變 | 不適用 to the degeneracy method itself; the host system removes lidar distortion caused by visual odometry drift with a linear motion model within a sweep |
| 迴圈閉合 | 原文未報告 |
| 全域最佳化 | none |
| 地圖表示 | 不適用 |
| 先驗資訊 | A prediction x_p is required and kept in the degenerate directions; in Test 1 it came from a constant velocity model |
| 可輸出幾何 | 不適用 |
| 計算需求 | Adds O(kn^2 + n^3) time to the original solver (Theorem 1), which reduces to O(k) for small fixed n such as 6-DOF; directions are computed at the first nonlinear iteration only; host system runs visual odometry at 60 Hz and scan matching at 1 Hz; processor not 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 |
|---|---|---|---|---|---|
| LiDAR | Hokuyo UTM-30LX | 方法輸入 | 未標示 | 180 deg field of view, 0.25 deg resolution, 40 lines/s; rotated by a motor to form a custom 3D lidar | (Zhang et al., 2016, Sec. V-A; Fig. 5) |
| 相機 | uEye monochrome camera | 方法輸入 | 未標示 | 60 Hz frame rate, 752 x 480 pixels, 76 deg horizontal field of view | (Zhang et al., 2016, Sec. V-A; Fig. 5) |
| 載具平台 | handheld custom-built camera and lidar sensor pack | 方法輸入 | 未標示 | carried by a person walking at 0.5 m/s | (Zhang et al., 2016, Sec. VI; Fig. 5) |
| 其他 | motor (model not reported) | 方法輸入 | 未標示 | rotates the laser scanner back and forth between -90 and 90 deg at 180 deg/s average; one 180 deg sweep lasts 1 s | (Zhang et al., 2016, Sec. V-A; Sec. V-B2) |
| 其他 | encoder (model not reported) | 方法輸入 | 未標示 | measures motor rotation angle with 0.25 deg resolution | (Zhang et al., 2016, Sec. V-A) |
作者報告的優勢與限制
優勢
- improves estimation in environmentally degenerate cases (abstract)
- In the corridor test the D/ICN ratio drops at the two feature-poor corners, which the authors read as D being more effective than ICN; IMCE is noisy because it also depends on the residual sum and constraint count (Sec. VI Test 1, Fig. 7)
- On flat ground, the compared ICN and D curves show an obvious drop between 35 and 70 s and D/ICN decreases slightly, while IMCE is noisy or barely decreases because it also depends on R and the number of constraints (Sec. VI Test 2, Fig. 8)
- Identified the physically expected degenerate directions: lateral or vertical translation for visual odometry (Test 1); forward and lateral translation plus yaw on flat ground (Test 2)
- Test 3 (538 m indoor and outdoor loop): 0.71% end position error, whereas a constant motion prior gave about three times larger end drift (Sec. VI, Fig. 10)
- Negligible add-in cost as a plug-in to common solvers (Sec. I, Theorem 1)
限制
- Threshold is set empirically at the midpoint between well-conditioned and degenerate lambda_min groups from one sample dataset (Sec. IV-B, Fig. 4)
- Assumes A is appropriately noise-weighted and the problem is full rank (Sec. III)
- Relies on the prediction in degenerate directions, which the authors treat as the only usable estimate there (Sec. IV-B, Sec. VII; dependency, not an author-stated limitation)
- single eigenvalue threshold needs heuristic tuning because rotation and translation eigenvalue scales differ (Tuna et al., 2024 Sec. VII-D/E, secondary)
營建工程相關證據
原文未在工地驗證;後續 X-ICP 在瑞士 Rümlang 工地的比較中以本法為基線並指出其單一門檻的問題(見 Tuna et al., 2024)。
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 45 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。
Tuna et al., 2024 · Table I 本方法 9 筆
資料集與序列Seemuhle underground mine (authors' data) · VLP-16 run; first 15 m alignment
表格設定(擷取紀錄原文,這些數值分屬表中不同部分):(Tuna et al., 2024, Table I)
- Seemuhle underground mine, ANYmal with VLP-16, 521.8 m; APE via EVO against Leica RTC 360 ground truth, mu (sigma); 'first 15 m' = trajectory aligned on the first 15 m (about 200 poses), 'origin' = aligned at the first pose; plus last-position error
- Same setting as other Table I rows
APE Translation mu(sigma) [m],Seemuhle underground mine (authors' data) · VLP-16 run; first 15 m alignment
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 Tuna et al., 2024 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(Tuna et al., 2024, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| X-ICP (Proposed)原文提出 | 2.05 m | (Tuna et al., 2024, Table I) |
| Zhang et al. [12]本方法 | 3.36 m | (Tuna et al., 2024, Table I) |
| Hinduja et al. [17] | 5.79 m | (Tuna et al., 2024, Table I) |
Hatleskog & Alexis, 2024 · Table II 本方法 8 筆
資料集與序列Seemühle Mine · full trajectory
表格設定(擷取紀錄原文):Seemühle mine, mean (SD) APE and RPE (Hatleskog & Alexis, 2024, Table II)
RPE [m], mean (SD 0.01),Seemühle Mine · full 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 Hatleskog & Alexis, 2024 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(Hatleskog & Alexis, 2024, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Zhang [14]本方法 | 0.01 m | (Hatleskog & Alexis, 2024, Table II) |
| Hinduja [15] | 0.01 m | (Hatleskog & Alexis, 2024, Table II) |
| Ours原文提出 | 0.01 m | (Hatleskog & Alexis, 2024, Table II) |
| Lee [19] (Switch-SLAM) | 0.01 m | (Hatleskog & Alexis, 2024, Table II) |
Lee et al., 2025a · Table V 本方法 4 筆
表格設定(擷取紀錄原文):Ground-Challenge corridors; translation APE and RPE via EVO; only RMSE columns kept (mean, max, std omitted); Zhang et al. and X-ICP reimplemented by the authors on the same LiDAR-only framework (Lee et al., 2025a, Table V)
absolute pose error, translation, RMSE [m],Ground-Challenge · Corridor1 (zigzag)
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 Lee et al., 2025a 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(Lee et al., 2025a, Table V)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| KISS-ICP [5] | 2.17 m | (Lee et al., 2025a, Table V) |
| CT-ICP [6] | 0.54 m | (Lee et al., 2025a, Table V) |
| DLO [14] | 0.45 m | (Lee et al., 2025a, Table V) |
| Zhang et al. [18]本方法 | 0.28 m | (Lee et al., 2025a, Table V) |
| X-ICP [24] | 2.05 m | (Lee et al., 2025a, Table V) |
| Ours (GenZ-ICP)原文提出 | 0.24 m | (Lee et al., 2025a, Table V) |
Hatleskog & Alexis, 2024 · Table I 本方法 4 筆
指標degeneracy-induced drift (qualitative)
表格設定(擷取紀錄原文,這些數值分屬表中不同部分):(Hatleskog & Alexis, 2024, Table I)
- Qualitative presence or absence of degeneracy-induced drift in partial maps; no numeric values
- Qualitative drift check
degeneracy-induced drift (qualitative),Rümlang Construction Site · exp. 1 (VLP-16, FOV cut to 180 deg)
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。
- 未報告(沒有數值,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 Hatleskog & Alexis, 2024 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(Hatleskog & Alexis, 2024, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Ours原文提出 | 無數值未報告註記(擷取紀錄):no degeneracy-induced drift | (Hatleskog & Alexis, 2024, Table I, Fig. 2) |
| Zhang [14]本方法 | 無數值未報告註記(擷取紀錄):degeneracy-induced drift | (Hatleskog & Alexis, 2024, Table I, Fig. 2) |
| Hinduja [15] | 無數值未報告註記(擷取紀錄):degeneracy-induced drift | (Hatleskog & Alexis, 2024, Table I, Fig. 2) |
| Lee [19] (Switch-SLAM) | 無數值未報告註記(擷取紀錄):degeneracy-induced drift | (Hatleskog & Alexis, 2024, Table I, Fig. 2) |
其他比較組
來源
Zhang et al., 2016
(2016)On degeneracy of optimization-based state estimation problems2016 IEEE International Conference on Robotics and Automation (ICRA), pp. 809-816
同儕審查已出版已讀全文經典查證後修正