Elastic LiDAR Fusion
Elastic LiDAR Fusion 把連續時間（continuous-time）SLAM 與 ElasticFusion 的「以地圖為中心」（map-centric）概念結合：局部仍以滑動視窗的連續時間軌跡處理手持旋轉 LiDAR 的運動畸變，但全域一致性不靠整條軌跡的批次最佳化，而是在迴圈發生時對整張面元地圖做非剛性變形（deformation graph）。因此迴圈閉合的計算量取決於迴圈前探索的空間大小，而非運作時間。多次觀測以機率式面元融合（surfel fusion）合併，論文報告可降低重建表面的雜訊。
本頁內容
Elastic LiDAR Fusion keeps a local continuous-time trajectory for de-skewing but achieves global consistency by deforming a probabilistically fused surfel map on loop closure instead of batch trajectory optimization.
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 | rotating 2D LiDAR (Hokuyo UTM-30LX spinning, with encoder)、IMU (Microstrain 3DM-GX3)、Grasshopper3 2.8 MP colour camera with fisheye lens used only for colourisation、Optris PI 450 thermal-infrared camera (382 x 288 pixels) on the device but not used |
|---|---|
| 原文測試平台 | handheld |
| 狀態估計 | local sliding-window continuous-time trajectory optimisation solved by iterative nonlinear least squares over subsampled trajectory elements Q, IMU biases and an additional state d that the paper does not define (x = [Q, b_omega, b_alpha, d]), combining surfel-to-surfel, surfel-to-map-prior and IMU acceleration and angular-velocity constraints (Eq. 3-8); global consistency by Gauss-Newton optimisation of an ElasticFusion-style deformation graph with loop, pinning and regularisation terms (Eq. 16-19) |
| 資料關聯 | multi-resolution 3D ellipsoidal sparse surfels (from Bosse and Zlot 2009, ref. [9]) matched pairwise and to the global-map prior in their averaged normal direction (Eq. 4-5); dense 2D disk surfels associated with a sensor-noise model that searches deeper along the beam direction under a resolution threshold, then fused by Bayesian fusion (Sec. V-A) |
| 時間表示 | continuous-time trajectory in a local window using linear interpolation between subsampled poses (slerp-like rotation, linear translation), chosen over B-splines to keep high-frequency motion at low cost |
| 去畸變 | handled by the continuous-time trajectory representation (Sec. I, VIII) |
| 迴圈閉合 | two detection sources: (1) rigid ICP between active and inactive sparse-surfel maps detects moderate misalignment on the fly (Algorithm 1); (2) for large misalignment, 3D point-cloud place recognition by keypoint voting (ref. [17], Bosse and Zlot 2013) with descriptors computed every frame and compared with stored scene keys; loop constraints are applied as map deformation, not trajectory optimisation |
| 全域最佳化 | deformation graph applied to the whole map (ElasticFusion-inspired) with surfel uncertainty propagation (Sec. III) |
| 地圖表示 | sparse multi-resolution ellipsoid surfel map plus dense 2D disk surfel map with probabilistic surfel fusion (Sec. III) |
| 先驗資訊 | no external prior; the initial map prior is built from a short period of stationary scanning at the start |
| 可輸出幾何 | dense fused surfel map (e.g., 20 mm surfels at 10 mm resolution in Fig. 1); no global trajectory is maintained (Sec. VII-B) |
| 計算需求 | per-frame runtime and hardware not reported anywhere in the paper; only global loop-closure optimisation cost is given (192 states and 0.12 s for the proposed method versus 3396 states and 195.40 s for batch CT-SLAM on the Fig. 1 map, Table I) |
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 laser spun on a hand-held device to give 3D scans | (Park et al., 2018, Sec. VII; Fig. 2a) |
| 行動掃描設備 | experimental handheld 3D spinning LiDAR (as written; builder not stated) | 方法輸入 | 未標示 | integrates 2D laser, encoder, IMU, colour camera and thermal camera | (Park et al., 2018, Fig. 2a; Sec. VII) |
| 慣性量測單元（IMU） | Microstrain 3DM-GX3 | 方法輸入 | 未標示 | 原文未報告 | (Park et al., 2018, Sec. VII) |
| 相機 | Grasshopper3 2.8 MP color camera | 方法輸入 | 未標示 | 2.8 MP; fisheye lens (Sec. V-A); used only for colourising the dense surfel map | (Park et al., 2018, Sec. V-A; Sec. VII) |
| 熱像儀 | Optris PI 450 thermal-infrared camera | 資料集感測器 | authors' hand-held spinning LiDAR data | 382 x 288 pixels; mounted on the device; Fig. 2a caption states the thermal camera is not used in the paper | (Park et al., 2018, Sec. VII; Fig. 2a caption) |
| 其他 | encoder (spinning mechanism) | 方法輸入 | 未標示 | part of the hand-held spinning LiDAR | (Park et al., 2018, Sec. VII; Fig. 2a caption) |
作者報告的優勢與限制
優勢
- Loop-closure optimisation state dimension and time much smaller than batch CT-SLAM: 192 states and 0.12 s versus 3396 states and 195.40 s on the Fig. 1 office map (Sec. VII-A, Table I; hardware not reported)
- Floor-patch projective-distance noise 5.79-13.07 mm vs 15.78-19.40 mm for the CT-SLAM point cloud (four 0.7 m patches) (Table III)
- Uniform surfel density after fusion regardless of raw point count (Sec. VII-C)
限制
- Trajectory accuracy (RMSE 0.041-0.076 m) is measured against the batch CT-SLAM trajectory of ref. [3] (Bosse et al. 2012, Zebedee), not an independent ground truth (Sec. VII-B, Table II)
- Surface evaluation uses planarity of floor patches (projective distance to each patch's mean plane) because ground truth is not available (Sec. VII-C)
- Journal extension reports map distortion reaching 10 cm on a 60 m scale map and a partial-observation problem (Park et al., 2022, Sec. IX, arXiv version)
營建工程相關證據
原文未報告（測試於辦公室、會議室、多樓層結構與室內外混合環境等既有建物；報告的平面補丁雜訊指標與工程表面品質相關（推論），但未做工地或工程任務驗證）。
原文驗證環境：已完工建築
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 4 個比較組，合計 22 筆紀錄。
Park et al., 2022 · Table VI 本方法 12 筆
表格設定（擷取紀錄原文）：loop-closure misalignment estimation on mixed indoor and outdoor data; ground truth from the globally optimised trajectory; 10 locations x 50 random initial guesses (500 triggers) per level; Easy sigma_theta_z 10 deg, sigma_theta_xy 1 deg, sigma_t 0.5 m; Medium 50, 5, 5; Hard 100, 20, 50; text calls the values RMSE, caption calls them error norms with std in parentheses. Values read from the VoR Table VI by the second checker (Park et al., 2022, Table VI)
e_t translation error，authors' mixed indoor and outdoor point clouds · Easy initial guess
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 VI)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| (a) Sparse surfel ICP (configuration of previous work [2])本方法 | 0.04 m | (Park et al., 2022, Table VI (VoR, p. 993; identical to arXiv v1 Table V)) |
| (b) Open3D global registration (FPFH + RANSAC) [60] | 0.3 m | (Park et al., 2022, Table VI (VoR, p. 993; identical to arXiv v1 Table V)) |
| (c) SHOT initialisation + point-to-plane ICP [61] | 1.49 m | (Park et al., 2022, Table VI (VoR, p. 993; identical to arXiv v1 Table V)) |
| (d) Proposed sequential metric localisation原文提出 | 0.03 m | (Park et al., 2022, Table VI (VoR, p. 993; identical to arXiv v1 Table V)) |
Park et al., 2018 · Table II 本方法 4 筆
指標Traj Error (m), absolute trajectory RMSE vs CT-SLAM [3]
表格設定（擷取紀錄原文）：absolute trajectory RMSE between the deformed trajectory of the proposed method and the globally optimised CT-SLAM [3] trajectory (reference is another estimate, not ground truth); trajectories stored only for this comparison (Park et al., 2018, Table II)
Traj Error (m), absolute trajectory RMSE vs CT-SLAM [3]，authors' hand-held spinning LiDAR data · Fig. 1 (length 330 m, 14.6 min, size 20x20 m)
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 Park et al., 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 |
|---|---|---|
| Proposed (Elastic LiDAR Fusion)本方法原文提出 | 0.047 m | (Park et al., 2018, Table II) |
Park et al., 2018 · Table III 本方法 4 筆
指標projective distance error to patch mean plane
表格設定（擷取紀錄原文）：floor patches of 0.7 m radius (Fig. 8b); error = mean projective distance of points or surfels to the mean plane of each patch (relative noise, no ground truth); CT-SLAM cloud is the raw point cloud of [3]; patch point and surfel counts in sequence field (Park et al., 2018, Table III)
projective distance error to patch mean plane，authors' hand-held spinning LiDAR data · patch a (47.8x10^4 points, 3.7x10^3 surfels)
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., 2018 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., 2018, Table III)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| CT-SLAM [3] (raw point cloud) | 16.08 mm | (Park et al., 2018, Table III) |
| Proposed (Elastic LiDAR Fusion, fused surfels)本方法原文提出 | 7.72 mm | (Park et al., 2018, Table III) |
Park et al., 2018 · Table I 本方法 2 筆
資料集與序列authors' hand-held spinning LiDAR data · Fig. 1 office map
表格設定（擷取紀錄原文）：global loop-closure optimisation cost for the map of Fig. 1 (office); proposed closes the loop at Fig. 5 (i), CT-SLAM batch-optimises the whole subsampled trajectory at the end (Fig. 5 (ii)); each state is 6-DoF; hardware not reported (Park et al., 2018, Table I)
No. State (optimisation state dimension)，authors' hand-held spinning LiDAR data · Fig. 1 office map
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., 2018 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., 2018, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Proposed (Elastic LiDAR Fusion, deformation graph)本方法原文提出 | 192 count | (Park et al., 2018, Table I) |
| CT-SLAM [3] (global batch trajectory optimisation) | 3396 count | (Park et al., 2018, Table I) |
來源
Park et al., 2018
（2018）Elastic LiDAR Fusion: Dense Map-Centric Continuous-Time SLAM2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 1206-1213
DOI 10.1109/icra.2018.8462915arXiv 1711.01691
同儕審查已出版已讀全文近十年查證後修正
相關版本
- 預印本：arXiv 1711.01691 (v1 2017-11-06, v3 2018-03-05) https://arxiv.org/abs/1711.01691
- 期刊延伸版：Elasticity Meets Continuous-Time: Map-Centric Dense 3D LiDAR SLAM 10.1109/TRO.2021.3096650