Zebedee
Zebedee 把 Hokuyo UTM-30LX 2D 雷射掃描儀與 MicroStrain 3DM-GX2 IMU 裝在同一感測頭，再以彈簧連接手把或載具，利用手持晃動或載具振動讓掃描面不規則擺動而取得三維覆蓋。配套 SLAM 改自作者先前的旋轉 2D 雷射方法：在滑動時間視窗內，以多解析度體素中時空相近的點群建立面元，在位置與法向量構成的 6 維空間做互為最近鄰的配對，並以面元匹配誤差、IMU 角速度與加速度偏差及視窗銜接條件，求解以固定間隔取樣、其間內插的連續時間軌跡修正量，同時估計雷射與 IMU 的時間延遲及 IMU 偏差；另保留少量「固定視角」面元以抑制漂移。開迴路結果再作為初值，對整段軌跡做一次批次全域配準，得到閉迴路軌跡與點雲。摘要所稱「即時」是指處理時間短於資料擷取時間，論文實驗實際是在收集後離線處理。
本頁內容
Spring-mounted 2D lidar (Hokuyo UTM-30LX) plus IMU (MicroStrain 3DM-GX2) sensor with sliding-window continuous-time surfel SLAM that also estimates laser-IMU latency and IMU biases, limits drift with fixed views, and refines the whole trajectory by batch global registration; processing ran faster than acquisition but offline in the experiments.
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 | 2D time-of-flight laser Hokuyo UTM-30LX (270 deg FoV, 30 m maximum range, 40 Hz) (Sec. II)、Industrial-grade MEMS IMU MicroStrain 3DM-GX2 at 100 Hz with a rotational rate range of at least 600 deg/s; the second-generation device uses a MicroStrain 3DM-GX3 (Sec. II, IV-D) |
|---|---|
| 原文測試平台 | handheld (first generation tethered to a pushcart for power and logging; later with electronics in a backpack) (Sec. IV, Fig. 2)、hands-free backpack-mounted configuration (Sec. IV-D, Fig. 2(e))、dual-spring vehicle mount on a John Deere Gator TE electric vehicle, used for spring characterization only; no vehicle SLAM results are reported (Sec. II-A, Fig. 1) |
| 狀態估計 | Sliding-window continuous-time trajectory correction: stacked 6-DoF corrections sampled at regular intervals (linear interpolation assumed for the Jacobians), linearized and solved as Ax=b by iteratively reweighted least squares with a Lorentzian M-estimator and a decreasing outlier threshold; terms are surfel match errors, IMU acceleration and rotational-rate deviations and initial-condition constraints; the state is augmented with laser-IMU latency and IMU bias corrections (Sec. III-B to III-D) |
| 資料關聯 | Surfels from spatially and temporally proximal point clusters in a multiresolution voxel grid (resolution doubling per level, two grids offset by half a cell), planarity-filtered; clusters whose rotational velocity normal to the scan plane is below about 15 deg/s are discarded; approximate kNN in the 6-D position and normal space via a kd-tree, reciprocal matches only, with time separation above half the nominal sweep period; correspondences recomputed every iteration (Sec. III-A) |
| 時間表示 | continuous-time: trajectory corrections sampled at regular intervals and interpolated in between; each window advances by a fraction of its length and the first three correction samples are constrained for continuity (Sec. III, III-B) |
| 去畸變 | implicit: laser points are projected with the continuous-time trajectory estimate, and laser-IMU latency is estimated online in each window (Sec. III, III-C) |
| 迴圈閉合 | no explicit loop detection or place recognition; loops are closed implicitly by batch global registration of surfel correspondences over the whole trajectory, which needs a good open-loop initial guess (Sec. III-F) |
| 全域最佳化 | batch global registration over the entire trajectory in one window, minimizing surfel match errors, deviations from the open-loop velocities and gravity deviations; latency and IMU biases are kept from the open-loop solution (Sec. III-F) |
| 地圖表示 | view-based: the trajectory is the full state and raw points are projected when needed; multiresolution surfels for matching plus a small buffer of fixed views (surfels from recent finalized windows) (Sec. III, III-A, III-E) |
| 先驗資訊 | none |
| 可輸出幾何 | 6-DoF sensor-head trajectory and a 3D point cloud projected with the closed-loop trajectory (Sec. III-F, IV-B; Figs. 7, 8, 16) |
| 計算需求 | MATLAB with C++ MEX on a 3.2 GHz Intel Xeon CPU; open-loop processing took about 62 %, 71 % and 73 % of acquisition time with 0, 2 and 5 fixed views; global optimization took under 1 min (3.5 min office) and under 2 min (6.5 min courtyard); experiments were processed offline after collection, and a real-time C++ version was under development (Sec. IV-B, V-B) |
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 | 2D time-of-flight, 270 deg FoV, 30 m maximum range, 40 Hz scan rate, 60 x 60 x 85 mm, 210 g; manufacturer range accuracy 3 to 5 cm as cited by the authors | (Bosse et al., 2012, Sec. II; Sec. IV-A) |
| LiDAR | SICK LMS291 (spinning) | 參考或真值量測 | 未標示 | rotated at 1 Hz about its middle scan ray, mounted at 750 mm on a pushcart, 13 500 points per second, hemispherical FoV facing behind the cart | (Bosse et al., 2012, Sec. IV; Fig. 2(b)) |
| 行動掃描設備 | Zebedee handheld, first generation | 方法輸入 | 未標示 | laser and IMU in a 150 g 3D-printed housing on a single spring (50 to 150 mm, 5 to 20 g); total mass well under 0.5 kg; tethered to a pushcart for power and logging in early tests | (Bosse et al., 2012, Sec. II; Sec. IV; Fig. 2(a), 2(b)) |
| 行動掃描設備 | Zebedee handheld, second generation | 方法輸入 | 未標示 | MicroStrain 3DM-GX3 IMU mounted on the back of the laser and a spring with slightly different physical characteristics; design developed with assistance from CMD Product Design and Innovation; used for the stairwell oscillation-stoppage experiment | (Bosse et al., 2012, Sec. IV-D; Fig. 2(d); Acknowledgment) |
| 慣性量測單元（IMU） | MicroStrain 3DM-GX2 | 方法輸入 | 未標示 | industrial-grade, triaxial MEMS gyros and accelerometers, 100 Hz output, rotational rate range of at least 600 deg/s (nonstandard option), 41 x 63 x 32 mm, 50 g | (Bosse et al., 2012, Sec. II) |
| 慣性量測單元（IMU） | MicroStrain 3DM-GX3歸入：Microstrain 3DM-GX3 | 方法輸入 | 未標示 | used in the second-generation handheld Zebedee, mounted on the back of the laser | (Bosse et al., 2012, Sec. IV-D) |
| 慣性量測單元（IMU） | second IMU on the sensor base (model not stated) | 參考或真值量測 | 未標示 | measures base motion as input for spring system identification | (Bosse et al., 2012, Sec. II-A) |
| 載具平台 | John Deere Gator TE electric vehicle | 方法輸入 | 未標示 | automated electric vehicle carrying the dual-spring vehicle-mounted Zebedee; driven off-road to provide inputs for spring system identification; no vehicle SLAM results are reported | (Bosse et al., 2012, Sec. II-A; Fig. 1(b)) |
| 載具平台 | wheeled pushcart | 方法輸入 | 未標示 | carries the spinning SICK reference and power and logging for the first-generation handheld | (Bosse et al., 2012, Sec. IV; Fig. 2(b)) |
| 運算硬體 | 3.2 GHz Intel Xeon CPU | 執行運算平台 | 未標示 | MATLAB with C++ MEX; open-loop at about 62 to 73 % of acquisition time | (Bosse et al., 2012, Sec. IV-B) |
| 其他 | Vicon motion capture system (14 cameras) | 參考或真值量測 | 未標示 | millimeter position precision within about 2 x 2 m; about 1 deg tag orientation precision; 10 x 8 m room | (Bosse et al., 2012, Sec. IV-C) |
作者報告的優勢與限制
優勢
- Closed-loop positional RMS error vs Vicon of 0.88, 0.72 and 0.69 cm with 0, 2 and 5 fixed views (Sec. IV-C, Fig. 12)
- Mobile mapping point-cloud error std of 3.9 cm (office) and 4.1 cm (courtyard) relative to a spinning SICK LMS291 reference cloud (Sec. IV-B)
- Pseudostationary error std of 2 to 2.5 cm in hallway and stairwell and around 5 cm in the remaining datasets; the authors state that the manufacturer's 3 to 5 cm range accuracy agrees with the minimum error std among the environments (Sec. IV-A)
- Fixed views cut open-loop drift from about 5 to 10 cm/min to about 1.5 and 0.5 cm/min (2 and 5 fixed views) and keep the solution robust to oscillation stoppages of several seconds (Sec. IV-C, IV-D)
- Both open-loop and global steps ran in less than the data acquisition time (Sec. IV-B)
- Light and mechanically simple: total mass well under 0.5 kg, hardware cost essentially a 2D laser plus an IMU (Sec. I, II)
- Hands-free backpack mounting gives enough excitation for reliable trajectory estimation (Sec. IV-D)
限制
- Degenerate environments such as long smooth tunnel-like spaces or large featureless open areas, and scenes dominated by moving objects (Sec. V-A)
- Needs fairly continual excitation of the sensor head; may not suit electric ground vehicles on smooth terrain (Sec. V-A)
- The estimated trajectory is that of the oscillating sensor head, not the carrying platform (Sec. V-A)
- No uncertainty (covariance) estimate (Sec. V-A)
- Very sensitive to laser-IMU timing: about 10 cm point error at 10 m per millisecond of latency error at 600 deg/s, with latency drift of about 1 ms/min observed (Sec. II-B, IV-B)
- Global registration needs a good initial guess; with zero fixed views a 6 s oscillation stoppage on stairs produced about 5 m vertical error (Sec. III-F, IV-D)
- Range bias of about 1 to 2 cm toward the sensor, attributed to the Hokuyo laser (Sec. IV-A)
- Vicon orientation precision (about 1 deg) allowed only position errors to be evaluated, within about a 2 x 2 m region (Sec. IV-C)
- (inference) Stairwell and backpack accuracy were measured against the method's own closed-loop solutions rather than an independent reference (Sec. IV-D)
營建工程相關證據
論文動機提到測量、行動建圖與受限空間中的基礎設施檢測，並表示已部署於大型組裝廠房與室內空間，但未提供這些部署的量化結果（Sec. I、III、V）；實驗場景為辦公室、走廊、三層樓梯間、戶外中庭、高草地與道路，未在營建工地驗證。（推論）輕量手持與背負式配置適合工地巡檢式掃描，但須確保雷射與 IMU 計時同步及足夠的感測頭擺動。
原文驗證環境：受控實驗、獨立參考量測、已完工建築、跨場域
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 20 個比較組，合計 120 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 16 組列在最後，並連到性能比較頁。
Park et al., 2022 · Table V 本方法 24 筆
表格設定（擷取紀錄原文）：known planar patches; position error = projective distance to patch mean plane (mm), normal error in rad; CT-SLAM [3] cloud is undistorted by its globally optimised trajectory but unfused; no ground truth. Values read from the VoR Table V by the second checker (Park et al., 2022, Table V)
Position Err. (projective distance)，authors' real datasets · patch a
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 Park et al., 2022 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(Park et al., 2022, Table V)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| CT-SLAM [3] (raw points)本方法 | 8.2 mm | (Park et al., 2022, Table V (VoR, p. 991; identical to arXiv v1 Table IV)) |
| Proposed (fused surfels)原文提出 | 3.4 mm | (Park et al., 2022, Table V (VoR, p. 991; identical to arXiv v1 Table IV)) |
Sammartano & Spanò, 2018 · Table 5 本方法 10 筆
表格設定（擷取紀錄原文）：Tower (A), ZEB1 surfaces vs CRP reference model (about 1 cm accuracy), cloud-to-cloud best-fitting alignment. (Sammartano & Spanò, 2018, Table 5)
Mean，Valperga castle · Tower (A), full raw roundtrip
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 Sammartano & Spanò, 2018 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 |
|---|---|---|
| ZEB1本方法 | 0.025 m | (Sammartano & Spanò, 2018, Table 5) |
Bosse et al., 2012 · Text Sec.IV-B 本方法 9 筆
資料集與序列authors' mobile mapping experiments · office and courtyard
表格設定（擷取紀錄原文）：Handheld Zebedee tethered to a pushcart carrying a spinning SICK LMS291; looped path traversed twice in each environment; the spinning-laser data give a globally registered reference trajectory and cloud; the Zebedee closed-loop cloud is compared with the reference cloud in the same manner as the pseudostationary tests (the two clouds registered to each other by surfel-based optimization, then point-to-surfel errors); runtime figures are for MATLAB with C++ MEX on a 3.2 GHz Intel Xeon (Bosse et al., 2012, Text Sec.IV-B)
open-loop processing time as share of acquisition time，authors' mobile mapping experiments · office and courtyard
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 Bosse et al., 2012 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(Bosse et al., 2012, Text Sec.IV-B)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Zebedee SLAM open-loop (0 fixed views)本方法原文提出硬體：3.2 GHz Intel Xeon CPU; MATLAB with C++ MEX | 62%有附註註記（擷取紀錄）：approximate | (Bosse et al., 2012, Sec. IV-B) |
| Zebedee SLAM open-loop (2 fixed views)本方法原文提出硬體：3.2 GHz Intel Xeon CPU; MATLAB with C++ MEX | 71%有附註註記（擷取紀錄）：approximate | (Bosse et al., 2012, Sec. IV-B) |
| Zebedee SLAM open-loop (5 fixed views)本方法原文提出硬體：3.2 GHz Intel Xeon CPU; MATLAB with C++ MEX | 73%有附註註記（擷取紀錄）：approximate | (Bosse et al., 2012, Sec. IV-B) |
Sammartano & Spanò, 2018 · Table 7 本方法 8 筆
表格設定（擷取紀錄原文）：Ice house (B), share of outward vs return deviation errors per range; raw vs optimised. (Sammartano & Spanò, 2018, Table 7)
0.00 < error < 0.02 m (blue)，Valperga castle · Ice house (B), O&R, raw
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 Sammartano & Spanò, 2018 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 |
|---|---|---|
| ZEB1本方法 | 78.5% | (Sammartano & Spanò, 2018, Table 7) |
其他比較組
列出其餘 16 個比較組
- Thomson et al., 2013 · Table 3
- Thomson et al., 2013 · Table 4
- Bosse et al., 2012 · Text Sec.IV-C
- Sammartano & Spanò, 2018 · Table 12
- Sammartano & Spanò, 2018 · Text Sec.Metric validation
- Thomson et al., 2013 · Table 2
- Park et al., 2018 · Table III
- Sammartano & Spanò, 2018 · Table 13
- Sammartano & Spanò, 2018 · Table 4
- Sammartano & Spanò, 2018 · Table 6
- Bosse et al., 2012 · Text Sec.IV-D
- Makkonen et al., 2017 · Text Sec.4
- Bosse et al., 2012 · Text Sec.IV-A
- Park et al., 2022 · Table II
- Park et al., 2022 · Text Sec. IV-E
- Park et al., 2018 · Table I
來源
Bosse et al., 2012
（2012）Zebedee: Design of a Spring-Mounted 3-D Range Sensor with Application to Mobile MappingIEEE Transactions on Robotics, 28(5): 1104-1119
同儕審查已出版已讀全文經典查證後修正