Map-centric dense 3D LiDAR SLAM (ElasticLiDAR++)
本文是 Elastic LiDAR Fusion 的期刊延伸，正式版將系統命名為 ElasticLiDAR++，把以地圖為中心的變形式 SLAM 推廣到旋轉單線與多線 3D LiDAR，並融合 IMU 與相機。局部以 100 Hz 線性內插的連續時間軌跡處理運動畸變，再以 B 樣條控制點在 SE(3) 上估計修正量，同時線上估計 LiDAR 與相機的時間延遲；全域則以變形圖讓整張面元地圖產生彈性變形來閉合迴圈，不保存整條軌跡。面元以常態逆 Wishart 模型遞迴融合，搭配保持表面解析度的匹配規則，使多次掃描融合成不重複的稠密地圖。迴圈的錯位估計結合面元點對面約束與 3D 特徵點對點約束，並在多個位置序列式融合，直到不確定度低於門檻。作者以 CT-SLAM 的全域最佳化軌跡為參考，軌跡差異為 0.173 至 0.552 m，平面補丁雜訊最多約降為三分之一；實作需約 2.1 秒處理 1 秒資料，尚非即時。
本頁內容
Journal extension of Elastic LiDAR Fusion, named ElasticLiDAR++ in the version of record: local continuous-time LiDAR-inertial-visual trajectory optimisation (linear interpolation at 100 Hz, B-spline SE(3) corrections, online time-lag estimation), normal-inverse-Wishart surfel fusion with resolution-preserving matching, and map deformation with sequential 3D-feature-plus-surfel metric loop closure; not real time (2.1 s per 1 s of data).
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 with encoder, rotor at 1 rotation/s, hand-held)、3D LiDAR (Velodyne VLP-16, hand-held and robot-mounted)、IMU (Microstrain 3DM-GX3 in hand-held payloads、model not stated for the robot payload)、RGB camera on the single-beam device、independent GoPro without common clock on the multi-beam hand-held device、camera on robot payload (model not stated) |
|---|---|
| 原文測試平台 | handheld (single-beam spinning LiDAR)、handheld (multi-beam VLP-16)、legged robot、wheeled ground robot、simulation |
| 狀態估計 | local composition-type continuous-time trajectory optimisation: discrete poses at 100 Hz corrected by cubic B-spline control points with SE(3) update, minimising surfel-to-surfel, surfel-to-map-prior and IMU residuals over control points, IMU biases and two time lags (LiDAR and camera) (Eq. 3-9); Gauss-Newton deformation-graph optimisation for loop closure (Eq. 21-24); sequential SE(3) pose fusion for metric localisation (Appendix B) |
| 資料關聯 | surfel-to-surfel and surfel-to-map-prior constraints; surface-resolution-preservative surfel matching for non-pinhole sensors (Sec. I, III) |
| 時間表示 | continuous-time within a local window: linear interpolation on se(3) between discrete poses generated at 100 Hz for residuals and undistortion, cubic B-spline correction trajectory for the update; LiDAR and camera time lags estimated online |
| 去畸變 | continuous-time trajectory representation (abstract) |
| 迴圈閉合 | detection: the overview describes 2D visual features compared with stored key frames (Sec. III); Sec. VII-D calls the trigger a 'visual place voting method' but cites [52], which is Bosse and Zlot's 3D LiDAR keypoint voting paper; Sec. VII-A states that experiments used a combined 3D and 2D detector [12], [55]; active-inactive sparse-surfel ICP detects moderate misalignment (Sec. V-D, VI-A4). Misalignment is estimated from LiDAR only by tightly combining surfel point-to-plane and 3D sparse-feature (for example FPFH) point-to-point constraints, sequentially fused at several places until the covariance meets a threshold; residual monitoring rejects false positives (Sec. VI-A5, Appendix B) |
| 全域最佳化 | non-rigid map deformation with surfel uncertainty propagation (Sec. III) |
| 地圖表示 | multi-resolution sparse ellipsoid surfels plus fixed-size hexagonal dense surfels with Wishart-based fusion (Sec. III) |
| 先驗資訊 | none |
| 可輸出幾何 | dense fused surfel map (Sec. III, Fig. 1) |
| 計算需求 | 2.1 s average to process 1 s of VLP-16, IMU and camera data on an i7-6700K CPU with 24 GB RAM and a GTX970 GPU; surfel fusion implemented in single-thread MATLAB; authors expect real time after a parallel C++ port (not demonstrated) |
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 | spinning single-beam laser, rotor 1 rotation/s; datasets moved at 0.9 m/s and 0.7 rad/s | (Park et al., 2022, VoR Sec. VII-A; Fig. 9a) |
| LiDAR | Velodyne VLP-16 | 方法輸入 | 未標示 | multi-beam; hand-held, legged-robot and wheeled-robot payloads | (Park et al., 2022, VoR Sec. VII-A; Fig. 9b-d) |
| 行動掃描設備 | hand-held single-beam 3D spinning LiDAR device | 方法輸入 | 未標示 | Hokuyo UTM-30LX, encoder, Microstrain 3DM-GX3, RGB camera | (Park et al., 2022, VoR Fig. 9a; Sec. VII-A) |
| 行動掃描設備 | hand-held multi-beam LiDAR device | 方法輸入 | 未標示 | Velodyne VLP-16, Microstrain 3DM-GX3, GoPro | (Park et al., 2022, VoR Fig. 9b; Sec. VII-A) |
| 慣性量測單元（IMU） | Microstrain 3DM-GX3 | 方法輸入 | 未標示 | used in both hand-held payloads | (Park et al., 2022, VoR Sec. VII-A) |
| 慣性量測單元（IMU） | IMU of the robot payload (model not stated) | 方法輸入 | 未標示 | 原文未報告 | (Park et al., 2022, VoR Sec. VII-A) |
| 相機 | RGB camera (single-beam hand-held device, model not stated) | 方法輸入 | 未標示 | used for colourisation and visual loop detection | (Park et al., 2022, VoR Sec. VII-A) |
| 相機 | Gopro | 方法輸入 | 未標示 | independent camera without a common clock with the LiDAR | (Park et al., 2022, VoR Sec. VII-A) |
| 相機 | camera of the robot payload (model not stated) | 方法輸入 | 未標示 | 原文未報告 | (Park et al., 2022, VoR Sec. VII-A) |
| 載具平台 | four-legged robot (model not stated) | 方法輸入 | 未標示 | dataset 2.5 min, 72 x 42 m, industrial area | (Park et al., 2022, VoR Sec. VII-A; Fig. 11a) |
| 載具平台 | wheeled ground robot (model not stated) | 方法輸入 | 未標示 | dataset 7 min, 38 x 49 m, industrial area | (Park et al., 2022, VoR Sec. VII-A; Fig. 11b) |
| 運算硬體 | i7-6700K CPU, 24 GB RAM, GTX970 GPU | 執行運算平台 | 未標示 | 2.1 s per 1 s of VLP-16, IMU and camera data | (Park et al., 2022, VoR Sec. VIII-B) |
| 其他 | encoder (spinning single-beam device) | 方法輸入 | 未標示 | 原文未報告 | (Park et al., 2022, VoR Sec. VII-A) |
作者報告的優勢與限制
優勢
- Simulated local trajectory optimisation: 10.3 mm and 1.2e-3 rad final accuracy with 66 states, versus 39.0 mm and 5.0e-3 rad for the linear SO(3)+R3 composition of [3] (Table II)
- full-stack simulation: 12.3 mm mean relative trajectory error versus 35.3 mm for [3] (Sec. IV-E)
- planar-patch position error 3.2 to 4.5 mm mean versus 8.0 to 9.4 mm for unfused CT-SLAM points (up to three times less noisy, VoR Table V)
- sequential metric localisation within 0.06 m and 0.004 rad in the hard case where Open3D and SHOT baselines reach metres of error (Table VI)
- structural difference to the baseline within plus or minus 0.02 m in a redundant scan (Fig. 14)
限制
- Map distortion reached 10 cm on a 60 m scale map and is expected to grow with map size; partial-observation problem (Sec. IX); trajectory difference to CT-SLAM grows with map size (0.173 to 0.552 m RMSE), attributed to ignoring gravity when integrating local maps and in the deformation graph (Sec. VII-B, Table IV); overall 0.1 m structural offset; partial revisits without enough overlap are problematic for long-range LiDAR, so fusion range must be limited (Sec. VIII-B); repeated non-Gaussian noise such as mixed pixels is not fused away and objects smaller than the surface resolution are lost (Sec. VIII-B); metric localisation error remains larger than surface reconstruction accuracy (Sec. VIII-C); not real time, 2.1 s per 1 s of data (Sec. VIII-B); trajectory reference is another estimate, not ground truth (Sec. VII-B)
營建工程相關證據
未在營建工地測試。資料集包括小房間、多樓層建物、辦公室、室內外混合、戶外結構化與非結構化場域，以及四足與輪式機器人在工業區蒐集的資料（Sec. VII-A）。軌跡精度以 CT-SLAM 的全域最佳化軌跡為參考，不是獨立量測的地面真值；表面品質以地板或牆面補丁到平均平面的投影距離評估，平均約 3.2 至 4.5 mm；這是相對於各補丁平均平面的雜訊，不是對獨立參考量測的絕對精度。地圖在 60 m 尺度下扭曲可達 10 cm，若用於竣工量測或尺寸檢核，仍需以獨立控制點驗證（推論）。
原文驗證環境：模擬、跨場域、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 6 個比較組，合計 47 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 2 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) | 8.2 mm | (Park et al., 2022, Table V (VoR, p. 991; identical to arXiv v1 Table IV)) |
| Proposed (fused surfels)本方法原文提出 | 3.4 mm | (Park et al., 2022, Table V (VoR, p. 991; identical to arXiv v1 Table IV)) |
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]) | 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本方法原文提出 | 0.03 m | (Park et al., 2022, Table VI (VoR, p. 993; identical to arXiv v1 Table V)) |
Park et al., 2022 · Table IV 本方法 6 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); length, duration, size and beam type in sequence field (Park et al., 2022, Table IV)
Traj Error (m), absolute trajectory RMSE vs CT-SLAM [3]，authors' hand-held datasets · (a) Small room, 130 m, 6.1 min, 10x6 m, single-beam
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., 2022 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 (ElasticLiDAR++, deformed trajectory)本方法原文提出 | 0.173 m | (Park et al., 2022, Table IV (VoR, p. 990)) |
Park et al., 2022 · Table II 本方法 2 筆
資料集與序列simulation (local trajectory optimisation) · 5 s window
表格設定（擷取紀錄原文）：simulation: 5 s local window, simulated angular velocity and linear acceleration at 100 Hz with bias and Gaussian noise, 1000 random timestamped sparse surfel features; accuracy = absolute trajectory accuracy after optimisation against simulated ground truth; composition models keep 500 discrete poses (Park et al., 2022, Table II)
Final t accuracy (absolute trajectory accuracy after optimisation)，simulation (local trajectory optimisation) · 5 s window
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 II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Composition model, Linear interpolation, Linear Composition, SO(3)+R3 update, 11 compositions, 66 states [3] | 39 mm | (Park et al., 2022, Table II (VoR; identical in arXiv v1)) |
| Composition model, Linear se(3) interpolation, Spline Composition, SE(3) update, 11 controls, 66 states (Ours)本方法原文提出 | 10.3 mm | (Park et al., 2022, Table II (VoR; identical in arXiv v1)) |
| Approximation model, Spline Direct, SE(3), 11 controls, 66 states [9] | 103 mm | (Park et al., 2022, Table II (VoR; identical in arXiv v1)) |
| Approximation model, Spline Direct, SE(3), 51 controls, 306 states [9] | 21.8 mm | (Park et al., 2022, Table II (VoR; identical in arXiv v1)) |
| Approximation model, Spline Direct, SE(3), 101 controls, 606 states [9] | 23.1 mm | (Park et al., 2022, Table II (VoR; identical in arXiv v1)) |
其他比較組
來源
Park et al., 2022
（2022）Elasticity Meets Continuous-Time: Map-Centric Dense 3D LiDAR SLAMIEEE Transactions on Robotics, 38(2): 978-997
DOI 10.1109/tro.2021.3096650arXiv 2008.02274
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：Elastic LiDAR Fusion: Dense Map-Centric Continuous-Time SLAM 10.1109/ICRA.2018.8462915
- 預印本：arXiv 2008.02274 v1 (submitted version) https://arxiv.org/abs/2008.02274