Quatro++
Quatro++ 針對 LiDAR SLAM 迴圈閉合中的全域配準，處理機械旋轉式 LiDAR 點雲稀疏、以及離群剔除後剩下不足三個內點造成退化兩個問題。方法先以地面分割移除幾何資訊少的地面點，再做特徵匹配與最大團內點選擇，並假設地面載具以偏航旋轉為主，以 GNC 估計準 SO(3) 旋轉與分量式平移；滾轉與俯仰可由 INS 補償。作者在 KITTI、NAVER LABS、MulRan 與手持式 HiltiOxford 資料上評估，並接入 SLAM 迴圈模組。
本頁內容
Quatro++ uses ground segmentation and a yaw-dominant quasi-SE(3) robust solver to improve global registration success for LiDAR loop closing under sparsity and degeneracy.
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 | 3D LiDAR (Velodyne HDL-64E, VLP-16, Ouster OS1-64, HESAI XT32 across datasets)、INS optional for roll/pitch (Sec. 5.5) |
|---|---|
| 原文測試平台 | vehicle (KITTI, MulRan)、handheld (HiltiOxford, qualitative feasibility only)、indoor NAVER LABS localization data (platform not stated) |
| 狀態估計 | decoupled estimation on translation-invariant measurements: quasi-SO(3) (yaw-only) rotation by GNC truncated least squares with alternating weight updates (noise bound 0.3, at most 50 iterations, kappa = 1.4), then component-wise translation estimation (COTE); the c2f variant adds G-ICP fine alignment |
| 資料關聯 | Patchwork ground segmentation removes ground points; voxel sampling; FPFH descriptors with sensor-specific radii (nu < r_normal < r_FPFH, Table 1); reciprocal-test matching; MCIS-heuristic pruning that keeps the maximal clique found within a time threshold |
| 時間表示 | 不適用 |
| 去畸變 | 原文未報告 |
| 迴圈閉合 | coarse alignment in the loop-closing module of LeGO-LOAM; QSC-LeGO-LOAM combines ScanContext loop detection, Quatro++ and local registration, with MSE-based false-loop rejection |
| 全域最佳化 | pose graph optimization over odometry and loop constraints (Eq. 14) in LeGO-LOAM |
| 地圖表示 | LiDAR scans |
| 先驗資訊 | ground-contact assumption; optional INS roll/pitch |
| 可輸出幾何 | relative pose (quasi-SE(3)) for loop constraints |
| 計算需求 | whole Quatro++ (preprocessing, correspondence estimation and Quatro) under 1 s per registration; Quatro optimization averages 5.0 ms (KITTI) and 6.4 ms (NAVER LABS) on an Intel Core i9-9900KF; preprocessing and matching times on Intel Core i7-7700K and i9-13900 are shown only as plots |
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 HDL-64E | 資料集感測器 | KITTI | 原文未報告 | (Lim et al., 2024, Sec. 6.1) |
| LiDAR | Velodyne VLP-16 | 資料集感測器 | NAVER LABS localization dataset | 原文未報告 | (Lim et al., 2024, Sec. 6.1) |
| LiDAR | Ouster OS1-64 | 資料集感測器 | MulRan | 原文未報告 | (Lim et al., 2024, Sec. 6.1) |
| LiDAR | HESAI XT32歸入：Hesai XT-32 | 資料集感測器 | Hilti-Oxford (HiltiOxford) | hand-held sensor configuration | (Lim et al., 2024, Sec. 6.1; Sec. 7.5) |
| 慣性量測單元（IMU） | INS (model not stated) | 方法輸入 | KITTI Seq. 06 (Table 5); HiltiOxford (Fig. 18) | raw roll and pitch used to compensate the source cloud before quasi-SO(3) estimation | (Lim et al., 2024, Sec. 5.5; Sec. 7.5; Table 5) |
| 運算硬體 | Intel Core i7-7700K | 執行運算平台 | 未標示 | preprocessing and correspondence timing | (Lim et al., 2024, Fig. 16) |
| 運算硬體 | Intel Core i9-13900 | 執行運算平台 | 未標示 | preprocessing and correspondence timing | (Lim et al., 2024, Fig. 16) |
| 運算硬體 | Intel Core i9-9900KF | 執行運算平台 | 未標示 | optimization timing (Quatro 5.0 ms KITTI, 6.4 ms NAVER LABS) | (Lim et al., 2024, Fig. 17) |
作者報告的優勢與限制
優勢
- higher success rate than state-of-the-art global registration under sparsity and degeneracy (abstract)
- ground segmentation significantly increases success for ground vehicles (abstract)
- improved loop-constraint quality and mapping precision (abstract, Sec. 8)
- QSC-LeGO-LOAM gives the lowest APE mean on DCC01, KAIST02 and Riverside01 (5.65, 3.88, 20.19) (Table 7)
- Quatro++-c2f reaches trel 0.50% and rrel 0.20 deg/100m on KITTI Seq. 00 at Delta = 5 (Table 6)
- handheld registration succeeds on HiltiOxford after INS roll-pitch compensation (Sec. 7.5, Fig. 18)
限制
- sacrifices relative roll/pitch estimation, requiring INS to recover (Sec. 5.5)
- MSE-based false-loop rejection can cause false negatives and false positives (Sec. 5.3)
- relies on ground contact and yaw-dominant motion assumptions (Sec. 4.1)
- some failure cases with low MSE occur in corridor-like scenes, making false positive loops hard to reject (Sec. 5.3)
- front-end feature extraction and matching are not improved; left to future work (Sec. 8)
- at small frame intervals Quatro without ground segmentation had lower errors than Quatro++ (Sec. 7.3, Table 6)
- ground segmentation occasionally lowers the success rate slightly (Sec. 7.2, Table 4)
營建工程相關證據
HiltiOxford 手持資料只用於定性可行性展示（Fig. 18），且需先以 INS 補償滾轉與俯仰；論文沒有描述該資料的場域是否為工地。其餘評估為 KITTI 與 MulRan 車載資料以及 NAVER LABS 室內資料。地面接觸與偏航主導的假設能否用於多樓層或斜坡工地屬推論。
原文驗證環境：公開基準
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 3 個比較組，合計 25 筆紀錄。
Lim et al., 2024 · Table 6 本方法 12 筆
表格設定（擷取紀錄原文）：KITTI Seq. 00 odometry test with frame interval Delta (source i+Delta, target i); trel [%] and rrel [deg/100m] by RPG evaluation tools; c2f = global registration then local registration (G-ICP); deep-learning rows copied by the authors from the original papers; † = Seq. 00 used for training (Lim et al., 2024, Table 6)
trel，KITTI · Seq. 00, Delta = 1
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 Lim 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(Lim et al., 2024, Table 6)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| ICP | 6.88% | (Lim et al., 2024, Table 6) |
| G-ICP | 1.26% | (Lim et al., 2024, Table 6) |
| VGICP | 1.03% | (Lim et al., 2024, Table 6) |
| FGR | 2.73% | (Lim et al., 2024, Table 6) |
| TEASER++ | 2.11% | (Lim et al., 2024, Table 6) |
| Quatro (Ours)原文提出 | 1.45% | (Lim et al., 2024, Table 6) |
| Quatro++ (Ours)本方法原文提出 | 1.9% | (Lim et al., 2024, Table 6) |
| LO-Net | 1.47% | (Lim et al., 2024, Table 6) |
| LO-Net+M | 0.78% | (Lim et al., 2024, Table 6) |
| DMLO† | 0.83% | (Lim et al., 2024, Table 6) |
| DMLO+M† | 0.73% | (Lim et al., 2024, Table 6) |
| A-LOAM + StickyPillars† | 0.65% | (Lim et al., 2024, Table 6) |
| SuMa | 0.68% | (Lim et al., 2024, Table 6) |
| A-LOAM | 0.7% | (Lim et al., 2024, Table 6) |
| Quatro-c2f (Ours)原文提出 | 0.65% | (Lim et al., 2024, Table 6) |
| Quatro++-c2f (Ours)本方法原文提出 | 0.68% | (Lim et al., 2024, Table 6) |
Lim et al., 2024 · Table 7 本方法 12 筆
表格設定（擷取紀錄原文）：absolute pose errors of full SLAM results on MulRan (Ouster OS1-64, vehicle); LeGO-LOAM variants: SC = ScanContext loop detection, TSC = TEASER++ + ScanContext, QSC = Quatro++ + ScanContext; unit not stated (presumably m); values as printed (for TSC on DCC01 RMSE < mean, and SC on Riverside01 std = mean) (Lim et al., 2024, Table 7)
Absolute pose error, Mean，MulRan · DCC01
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 Lim 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(Lim et al., 2024, Table 7)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LeGO-LOAM | 28.89 | (Lim et al., 2024, Table 7) |
| SC-LeGO-LOAM | 6.13 | (Lim et al., 2024, Table 7) |
| TSC-LeGO-LOAM | 6.1 | (Lim et al., 2024, Table 7) |
| QSC-LeGO-LOAM (Ours)本方法原文提出 | 5.65 | (Lim et al., 2024, Table 7) |
Lim et al., 2024 · Text Sec. 7.4 本方法 1 筆
指標total time of Quatro++ takes less than one second
資料集與序列not stated
表格設定（擷取紀錄原文）：total time of Quatro++ (preprocessing, correspondence estimation and Quatro) stated as an upper bound; hardware not named for the total (Lim et al., 2024, Text Sec. 7.4)
total time of Quatro++ takes less than one second，not stated
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 Lim et al., 2024 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 |
|---|---|---|
| Quatro++ (whole pipeline)本方法原文提出 | 1 s僅報告範圍註記（擷取紀錄）：upper bound (less than 1 s per registration) | (Lim et al., 2024, Sec. 7.4) |
來源
Lim et al., 2024
（2024）Quatro++: Robust global registration exploiting ground segmentation for loop closing in LiDAR SLAMThe International Journal of Robotics Research, 43(5):685-715
DOI 10.1177/02783649231207654arXiv 2311.00928程式碼
同儕審查已出版已讀全文近十年查證後修正
相關版本
- 會議版：A Single Correspondence Is Enough: Robust Global Registration to Avoid Degeneracy in Urban Environments (ICRA 2022, pp. 8010-8017; arXiv:2203.06612) 10.1109/ICRA46639.2022.9812018
- 程式碼釋出：Quatro (re-implementation; also integrated in TEASER++ per README) https://github.com/url-kaist/Quatro
程式碼：https://github.com/url-kaist/Quatro（授權：CC BY-NC-SA 4.0 (declared in README 'License' section; no LICENSE file; non-commercial terms)）。有公開程式碼不等於已被重現，也不代表目前版本與論文版本相同。