Continuous 3D scan-matching (Bosse and Zlot)
本文處理移動中以旋轉 2D 雷射取得三維點雲時的運動畸變：每半圈（sweep）約需 1 秒，車輛在期間移動會使點雲局部變形。作者不停車，也不依賴里程計或 IMU，而是把 ICP 改成「掃描對掃描」的連續時間配準：先把點分入 0.5 至 8 m 的多解析度體素金字塔，以每個體素的一、二階矩求出橢球及其平面度與圓柱度，再在結合位置與形狀的 9 維描述空間中找對應；第二步不是求單一剛體轉換，而是每 0.2 秒取樣一個軌跡修正量，以匹配、平滑（加速度）與初始條件約束組成線性系統，配合 Cauchy 型穩健權重反覆求解；匹配約束在相鄰樣本間線性內插，最後以三次樣條重建連續軌跡。方法本身沒有迴圈閉合或全域最佳化，屬開迴路里程計。作者在工業園區與輕度林地以滑移轉向裝載機測試，報告局部精確的點雲與六自由度軌跡，但 MATLAB 實作約比即時慢五倍。
本頁內容
Sweep-to-sweep continuous-time ICP for a spinning 2D lidar on a moving skid-steer loader, without odometric or inertial sensors: multi-resolution voxel ellipsoids with plane and cylinder shape parameters are matched in a 9D position-shape space, and trajectory corrections sampled every 0.2 s are solved under match, smoothness and initial-condition constraints (linear interpolation between samples, cubic-spline reconstruction), giving an open-loop 6-DoF trajectory and unwarped point cloud; no loop closure, about five times slower than real time in MATLAB.
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 | encoder on the spinning mount, described as accurate, giving each 2D scan's pose relative to the vehicle |
|---|---|
| 原文測試平台 | vehicle: Bobcat S185 skid-steer loader, spinning laser above the cab (Sec. II, Fig. 1) |
| 狀態估計 | ICP variant for sweep-to-sweep matching: small trajectory corrections sampled every 0.2 s (six samples per sweep) solved from a stacked linear system of match, smoothness (acceleration) and initial-condition constraints after first-order linearization; robust M-estimator with Lorentzian/Cauchy weights; outer loop (correspondences) usually converges within about five iterations, inner loop (reweighted solve) drops from about seven to one or two iterations (Sec. II-B, Eq. 6-16) |
| 資料關聯 | voxel (not point) correspondences: points binned into a pyramid of 3D grids (0.5 m to 8 m cells, several offsets), each voxel timestamped with the mean time of its points; per-voxel first and second moments define an ellipsoid with cylinder-likeness c (Eq. 3) and plane-likeness p (Eq. 4); nearest neighbours searched in a 9D descriptor [alpha*mu; p*v1; c*v3] (Eq. 5) with alpha set to ten times the grid resolution and eigenvector signs fixed (v1 towards the sensor, v3 towards +z); planar matches constrain centroid offset along the normal and cylindrical matches perpendicular to the axis (Eq. 10-12), scaled by inverse square-root eigenvalues; normal-angle constraints derived but omitted in the implementation (Sec. II-A, II-B) |
| 時間表示 | continuous-time trajectory T(tau) per sweep; corrections estimated at samples every 0.2 s, match constraints linearly interpolated between the two nearest samples (Eq. 13), and a cubic spline used to reconstitute a smooth continuous trajectory (Sec. II-B). This resolves the citing-paper discrepancy: Park et al., 2018 describes the linear interpolation of constraints, Ramezani et al., 2022 groups the paper with spline methods. |
| 去畸變 | central aim: the recovered sweep trajectory unwarps motion-distorted sweeps; points measured much later than the earliest point in a voxel are excluded to limit distortion (Sec. II-A, Fig. 4) |
| 迴圈閉合 | none in the method; the authors mention appearance-based loop-closure investigations with preliminary results (Fig. 10) and place globally consistent mapping outside the paper's scope (Sec. I, III, IV) |
| 全域最佳化 | none in the method; the off-road experiment compares the open-loop trajectory with a globally consistent closed-loop trajectory from a method 'not described in this paper' (Sec. III) |
| 地圖表示 | multi-resolution voxel ellipsoids (moments) for matching; output is the concatenated unwarped point cloud (Sec. II-A, Fig. 2, Fig. 6) |
| 先驗資訊 | none from external sensors; prior trajectories from a motion model (sweep a: previous sweep motion; sweep b: deceleration to zero at sweep end); IMU or odometry seeding possible but not used (Sec. II-B) |
| 可輸出幾何 | open-loop 6-DoF sensor trajectory and locally consistent unwarped 3D point cloud; no covariance or map-accuracy metric reported (Sec. III-IV) |
| 計算需求 | about 5 s of computation per 1 s of data in MATLAB on a 3.2 GHz Pentium 4 (about five times slower than real time); authors foresee real time after porting to C++ (Sec. IV) |
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 | SICK LMS291 | 方法輸入 | 未標示 | commercial 2D laser range finder spun about its centre scan line at 0.5 Hz; 75 Hz scan rate, 1 deg angular resolution; 150 scans per revolution; a half-revolution 'sweep' has 13,500 points; hemispherical field of view | (Bosse & Zlot, 2009, Sec. II; Fig. 1) |
| LiDAR | stationary 2D laser scanner mounted horizontally on the vehicle's bucket (model not stated) | 參考或真值量測 | 未標示 | processed by the authors' 2D SLAM framework [1] assuming zero pitch, roll and vertical translation | (Bosse & Zlot, 2009, Sec. III) |
| 載具平台 | Bobcat S185 skid-steer loader | 方法輸入 | 未標示 | spinning laser mounted above the cab, spin axis facing forward; skid-steer, can turn in place | (Bosse & Zlot, 2009, Sec. II; Fig. 1) |
| 運算硬體 | 3.2 GHz Pentium 4 processor (MATLAB) | 執行運算平台 | 未標示 | about 5 s of processing per 1 s of data | (Bosse & Zlot, 2009, Sec. IV) |
| 其他 | encoder on the spinning mount | 方法輸入 | 未標示 | described as accurate; gives laser scan poses relative to the vehicle | (Bosse & Zlot, 2009, Sec. II) |
作者報告的優勢與限制
優勢
- Industrial compound (flat, paved; about 200 m trajectory): relative to a horizontal-laser 2D SLAM trajectory, translational error growth generally below 5% of distance travelled and rotational error below 0.3 deg per metre for 5-30 m windows (Sec. III, Fig. 5, Fig. 7a-b)
- Off-road lightly wooded terrain (548 m, 8 min): translational errors generally below 1% of distance travelled with a rotational bias of about 0.1 deg per metre, relative to a closed-loop trajectory from an undescribed method (Sec. III, Fig. 7c-d, Fig. 9)
- Reported operation at vehicle speeds up to 6 m/s and turn rates of 20 deg/s (Sec. III)
- Single inexpensive 2D laser, no additional odometric sensors (Sec. I)
- Undistorted bollards and containers visible in a 40 m map segment (Sec. III, Fig. 6)
限制
- Open loop: no loop closure or global optimization in the method; the Fig. 6 map is explicitly without loop closure (Sec. III); consistent with Park et al., 2018 Sec. II
- Fast spot turns are the hardest motions; yaw is momentarily unobservable when the laser senses only the ground (Sec. III)
- ICP convergence unreliable at high rotational velocities; sparse environments increase errors (Sec. III-IV)
- Not real time: about 5 s per 1 s of data in MATLAB (Sec. IV); LOAM (Sec. II) describes Bosse and Zlot's methods ([3], [6], [22]) as requiring batch processing, and Wildcat (Sec. II) calls [22] an offline system
- Weak references: the industrial reference is a 2D laser SLAM trajectory assuming zero pitch, roll and vertical translation; the off-road reference is a closed-loop trajectory from an undescribed method (Sec. III); (inference) it is probably derived from the same spinning-laser data and therefore not independent
- Only two outdoor sites; no geometric accuracy assessment of the point cloud itself (Sec. III)
營建工程相關證據
實驗平台是營建常見的 Bobcat S185 滑移轉向裝載機，但測試場域為平坦鋪面的工業園區（有護柱與貨櫃）與輕度林地越野路段（Sec. III），不是營建工地、隧道或建物。工業園區的參考軌跡來自另一台水平 2D 雷射的 2D SLAM（假設無俯仰、側傾與垂直位移），只能檢查平面運動；越野路段則以未說明方法的閉迴路軌跡作參考，所以本文沒有點雲幾何精度的獨立驗證。同一團隊後續的 Zebedee（Bosse et al., 2012）與 Northparkes 礦坑建圖（Zlot & Bosse, 2014）才延伸到手持、室內與地下環境；後者明言其非剛性配準的原始版本出自本文。若要說本文是 GeoSLAM ZEB 或 Hovermap 等商用設備的技術來源，本次仍未找到可核證的原始依據，不宜寫入正文。
原文驗證環境：跨場域、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 5 筆紀錄。
Bosse & Zlot, 2009 · Text Sec. III 本方法 4 筆
資料集與序列authors' spinning-laser data (industrial) · about 200 m traverse
表格設定（擷取紀錄原文）：industrial compound, flat paved, about 200 m; sliding windows of 5 to 30 m aligned at window start; reference = 2D SLAM trajectory of a separate horizontal laser (assumes zero pitch, roll and vertical translation); values summarise box plots in Fig. 7a-b (Bosse & Zlot, 2009, Text Sec. III)
growth of translational errors, generally expected to be less than this share of distance travelled，authors' spinning-laser data (industrial) · about 200 m traverse
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 Bosse & Zlot, 2009 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 |
|---|---|---|
| 3D sweep-matching (continuous-time ICP)本方法原文提出 | 5%僅報告範圍註記（擷取紀錄）：upper bound stated in text ('less than 5%') | (Bosse & Zlot, 2009, Sec. III; Fig. 7a-b) |
Bosse & Zlot, 2009 · Text Sec. IV 本方法 1 筆
指標processing time per second of data
資料集與序列authors' spinning-laser data · all
表格設定（擷取紀錄原文）：average processing cost of the ICP sweep-matching in MATLAB (Bosse & Zlot, 2009, Text Sec. IV)
processing time per second of data，authors' spinning-laser data · all
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 Bosse & Zlot, 2009 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 |
|---|---|---|
| 3D sweep-matching (continuous-time ICP), MATLAB implementation本方法原文提出硬體：MATLAB on a 3.2 GHz Pentium 4 processor | 5有附註註記（擷取紀錄）：approximate ('around 5 seconds') | (Bosse & Zlot, 2009, Sec. IV) |
來源
Bosse & Zlot, 2009
（2009）Continuous 3D scan-matching with a spinning 2D laser2009 IEEE International Conference on Robotics and Automation (ICRA), 4312-4319
DOI 10.1109/robot.2009.5152851
同儕審查已出版已讀全文經典查證後修正