SLICT
SLICT 以 UFOMap 八元樹維護全域多尺度面元（surfel）地圖，每個節點只存點數、座標和與散佈矩陣，因此子節點新增或刪除時可遞增更新父節點面元，不必反覆重建整張地圖的 k-d 樹。前端把一顆或多顆 LiDAR 的點合併成單一串流，每個原始點依時間戳記在滑動視窗兩個相鄰狀態間內插位姿，直接與多個尺度中平面度足夠的面元建立點到面元因子，再與 IMU 預積分因子一起以 Ceres 最佳化。後端在 K 個最近關鍵影格中找出時間差足夠大的候選作為迴圈，以 ICP 計算相對位姿，做位姿圖最佳化後重算全域地圖。
本頁內容
Multi-LiDAR continuous-time LIO and mapping that keeps an incrementally updatable global multi-scale surfel octree (UFOMap), associates each raw point to planar surfels at several scales with time-interpolated states in a sliding-window optimization with IMU preintegration, and adds ICP-based loop closure with pose-graph 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 | one or several 3D LiDARs merged into one stream (Ouster OS1-128 plus prism-based Livox Mid-70 in-house; horizontal and vertical LiDARs in NTU VIRAL; Ouster 64-channel in Newer College)、IMU (VectorNav VN100 in-house; built-in 100 Hz Ouster IMU in Newer College) |
|---|---|
| 原文測試平台 | UAV (NTU VIRAL)、handheld (Newer College)、ATV (all-terrain vehicle, in-house, up to 30 km/h) |
| 狀態估計 | sliding-window MAP optimization in Ceres over several states per LiDAR sweep (2 to 8 new states per cloud; 400 ms window with 16 intervals in NTU VIRAL), with IMU preintegration factors and continuous-time point-to-surfel factors; the deskew, association and optimization steps can be iterated (Sec. II-C, III-A, III-F) |
| 資料關聯 | multi-scale point-to-surfel: each point is matched to every surfel at octree depths 1 to Dmax whose voxel intersects a sphere around the point, with enough points and planarity above a threshold, and whose plane distance is below dmax; five scales from 2 to 32 times the 0.1 m leaf size were used (Sec. III-D, IV-A) |
| 時間表示 | continuous-time in the sense of per-point factors: each raw point is coupled to the two bounding sliding-window states by linear interpolation of position and SO(3) interpolation of rotation at its normalized timestamp (Sec. II-C, III-D) |
| 去畸變 | IMU-propagated poses interpolated with slerp per point for association; the optimization itself uses raw points with time-interpolated states (Sec. III-C, III-D) |
| 迴圈閉合 | yes; proximity-based candidate among K nearest keyframes with sufficient time difference, relative pose by ICP (Sec. III-I) |
| 全域最佳化 | pose-graph optimization over keyframes with relative-pose priors and loop constraints, followed by recomputing the global map (Eq. 8, Sec. III-I) |
| 地圖表示 | global multi-scale surfel map in an octree built on UFOMap; each node stores point count, sum and scatter so surfels can be added or removed incrementally with Welford-type updates (Sec. II-D) |
| 先驗資訊 | none (LiDAR extrinsics for merging multiple LiDARs assumed known; not described) |
| 可輸出幾何 | trajectory, keyframe poses with deskewed keyframe clouds, and a global point and surfel map (Figs. 8 and 12) |
| 計算需求 | not guaranteed real time: about 165 ms per cycle on NTU VIRAL nya_02 (about 50 ms solving) on a Core i7 PC, versus 100 ms LiDAR period; authors expect real time with more CPU threads (Sec. IV-A) |
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 | Ouster OS1-128 | 方法輸入 | SLICT in-house NTU campus dataset | spinning LiDAR, merged with the Livox stream | (Nguyen et al., 2023, Sec. IV-C; Sec. I) |
| LiDAR | Livox Mid-70歸入：Livox MID70 | 方法輸入 | SLICT in-house NTU campus dataset | Livox LiDAR (the paper describes Livox as prism-based, box-shaped, Sec. I); merged with the Ouster stream | (Nguyen et al., 2023, Sec. IV-C) |
| LiDAR | Ouster 64-channel LiDAR (model not stated) | 資料集感測器 | Newer College Dataset | 64 channels, 90-degree vertical field of view | (Nguyen et al., 2023, Sec. IV-B) |
| LiDAR | NTU VIRAL horizontal and vertical LiDARs (models not stated) | 資料集感測器 | NTU VIRAL | two LiDARs merged as input | (Nguyen et al., 2023, Sec. IV-A) |
| 地面雷射掃描儀（TLS） | Leica RTC360 | 參考或真值量測 | SLICT in-house NTU campus dataset | survey scanner used to build a static map with centimetre accuracy; Ouster scans registered to it for ground-truth poses | (Nguyen et al., 2023, Sec. IV-C) |
| 慣性量測單元（IMU） | VectorNav VN100 | 方法輸入 | SLICT in-house NTU campus dataset | 原文未報告 | (Nguyen et al., 2023, Sec. IV-C) |
| 慣性量測單元（IMU） | Ouster built-in IMU | 資料集感測器 | Newer College Dataset | 100 Hz | (Nguyen et al., 2023, Sec. IV-B) |
| 全測站 | laser-tracker total station (model not stated) | 參考或真值量測 | NTU VIRAL | centimetre-level ground truth | (Nguyen et al., 2023, Sec. IV-A) |
| 載具平台 | All-Terrain-Vehicle (ATV) | 方法輸入 | SLICT in-house NTU campus dataset | about 1.5 km loop, up to 30 km/h | (Nguyen et al., 2023, Sec. IV-C) |
| 運算硬體 | Core i7 PC (model not stated) | 執行運算平台 | 未標示 | 原文未報告 | (Nguyen et al., 2023, Sec. IV-A) |
作者報告的優勢與限制
優勢
- Lowest ATE in 15 of the 18 NTU VIRAL sequences against MARS, LIO-SAM and FAST-LIO2; MARS and LIO-SAM diverged on some sequences (Table I)
- Lowest ATE on 4 of 5 Newer College sequences (Table II)
- Without loop closure, lowest ATE on all three in-house ATV sequences where MARS diverged in all and LIO-SAM in one; loop closure lowered ATE further to 0.74 to 0.98 m (Table III)
- Global map lets current scans associate with the earliest features instead of only recent local-map features (Sec. I, Fig. 1)
- Accepts any number of LiDARs, including spinning Ouster and prism-based Livox units, as one merged stream (Sec. I, III-A, IV-C)
限制
- Real time is not guaranteed: about 165 ms per cycle versus a 100 ms LiDAR period on a Core i7 PC (Sec. IV-A)
- Association over five scales costs at least several times the single-scale direct method (Sec. IV-A)
- Largest errors appear in the vertical direction on sites with elevation change (Sec. IV-B, IV-C, Figs. 7 and 11)
- Point-to-surfel association uses simple predicates and loop closure is basic (Sec. V)
- FAST-LIO2 had lower ATE on NTU VIRAL spms_02 and spms_03 and Newer College 06, and LIO-SAM on tnp_03 (Tables I and II)
營建工程相關證據
論文未在營建工地驗證；資料為 NTU VIRAL 無人機、Newer College 手持與 NTU 校園 ATV 序列。作者自建資料以 Leica RTC360 靜態掃描地圖加 ICP 配準產生參考軌跡，屬獨立幾何參考，但只評估軌跡 ATE，未評估點雲幾何精度。多 LiDAR 合併與全域面元地圖的做法，可對應工地機器人同時搭載旋轉式與固態 LiDAR 以提高覆蓋的需求（推論）。GLIM （Koide et al., 2024）在 NTU VIRAL 比較中列入 SLICT，RESPLE （Cao et al., 2025）比較其後繼版本 SLICT2。
原文驗證環境：公開基準、獨立參考量測、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 5 個比較組，合計 36 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 1 組列在最後，並連到性能比較頁。
Nguyen et al., 2023 · Table I 本方法 18 筆
指標ATE
表格設定（擷取紀錄原文）：NTU VIRAL; horizontal and vertical LiDARs merged as input for all methods; SLICT loop closure disabled; 400 ms window with 16 intervals; x = divergence (Nguyen et al., 2023, Table I)
ATE，NTU VIRAL · eee_01
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 Nguyen et al., 2023 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(Nguyen et al., 2023, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| MARS | 0.2471 m | (Nguyen et al., 2023, Table I) |
| LIO-SAM | 0.0624 m | (Nguyen et al., 2023, Table I) |
| FAST-LIO2 | 0.0585 m | (Nguyen et al., 2023, Table I) |
| SLICT本方法原文提出 | 0.0316 m | (Nguyen et al., 2023, Table I) |
Nguyen et al., 2023 · Table III 本方法 6 筆
指標ATE
表格設定（擷取紀錄原文）：In-house NTU campus ATV dataset (Ouster OS1-128 + Livox Mid-70 merged, VN100 IMU), ground truth by registering scans to a Leica RTC360 map; LC = loop closure and pose-graph optimization; x = divergence (Nguyen et al., 2023, Table III)
ATE，SLICT in-house NTU campus dataset · seq 01
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。
- 發散
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 Nguyen et al., 2023 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(Nguyen et al., 2023, Table III)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| MARS | 無數值發散註記（擷取紀錄）：diverged | (Nguyen et al., 2023, Table III) |
| LIO-SAM | 4.0784 m | (Nguyen et al., 2023, Table III) |
| FAST-LIO2 | 2.008 m | (Nguyen et al., 2023, Table III) |
| SLICT本方法原文提出 | 1.6929 m | (Nguyen et al., 2023, Table III) |
| LIO-SAM (LC) | 1.2931 m | (Nguyen et al., 2023, Table III) |
| SLICT (LC)本方法原文提出 | 0.9815 m | (Nguyen et al., 2023, Table III) |
Nguyen et al., 2023 · Table II 本方法 5 筆
指標ATE
表格設定（擷取紀錄原文）：Newer College (Ouster 64-channel, built-in 100 Hz IMU); settings as NTU VIRAL but 2 states per interval; LIO-SAM not run because it needs IMU orientation; x = divergence (Nguyen et al., 2023, Table II)
ATE，Newer College Dataset · 01 short experiment
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。
- 未執行
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 Nguyen et al., 2023 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(Nguyen et al., 2023, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| MARS | 2.1521 m | (Nguyen et al., 2023, Table II) |
| LIO-SAM | 無數值未執行註記（擷取紀錄）：未執行 | (Nguyen et al., 2023, Table II) |
| FAST-LIO2 | 0.3883 m | (Nguyen et al., 2023, Table II) |
| SLICT本方法原文提出 | 0.3843 m | (Nguyen et al., 2023, Table II) |
Burnett et al., 2025 · Table II 本方法 5 筆
指標root mean squared ATE
表格設定（擷取紀錄原文）：Newer College Dataset (handheld, 6 km); RMS ATE after Umeyama alignment; star = explicit loop closures, dagger = results from DLIOM [71], double dagger = uses camera; other baselines as originally published (Burnett et al., 2025, Table II)
root mean squared ATE，Newer College Dataset · 01-Short
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 Burnett et al., 2025 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(Burnett et al., 2025, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| CT-ICP* [18] (explicit loop closures) | 0.36 m | (Burnett et al., 2025, Table II) |
| KISS-ICP [5] (result from [71]) | 0.6675 m | (Burnett et al., 2025, Table II) |
| FAST-LIO2 [10] (result from [71]) | 0.3775 m | (Burnett et al., 2025, Table II) |
| DLIO [11] | 0.3606 m | (Burnett et al., 2025, Table II) |
| SLICT* [52] (explicit loop closures)本方法 | 0.3843 m | (Burnett et al., 2025, Table II) |
| CLIO* [60] (loop closures, uses camera) | 0.408 m | (Burnett et al., 2025, Table II) |
| Constant Velocity (ablation baseline) | 0.8558 m | (Burnett et al., 2025, Table II) |
| STEAM-LO (Ours)原文提出 | 0.3398 m | (Burnett et al., 2025, Table II) |
| STEAM-LO + Gyro (Ours)原文提出 | 0.3055 m | (Burnett et al., 2025, Table II) |
| STEAM-LIO (Ours)原文提出 | 0.3042 m | (Burnett et al., 2025, Table II) |
其他比較組
列出其餘 1 個比較組
來源
Nguyen et al., 2023
（2023）SLICT: Multi-Input Multi-Scale Surfel-Based Lidar-Inertial Continuous-Time Odometry and MappingIEEE Robotics and Automation Letters, 8(4):2102-2109
DOI 10.1109/lra.2023.3246390arXiv 2211.03900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
- 預印本：arXiv 2211.03900 (v1 2022-11-07, v2 2022-11-09) https://arxiv.org/abs/2211.03900
- 預印本：DR-NTU repository copy (submitted version, not opened) https://hdl.handle.net/10356/228819
- 程式碼釋出：brytsknguyen/slict (code and in-house datasets) https://github.com/brytsknguyen/slict
程式碼：https://github.com/brytsknguyen/slict（授權：GPL-2.0 (LICENSE file checked)）。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。