LAMM
LAMM 是離線的多時段光達點雲地圖合併框架，輸入各代理人由前端 SLAM（如 FAST-LIO2）得到的掃描與初始位姿。先以 M-Detector 為基礎，在正向與反向時間序列各做一次遮擋測試的雙向濾波移除動態點；再以 BTC 描述子在各序列的資料庫中搜尋序列內與序列間迴圈；序列間迴圈以「用每個迴圈把另一序列起點投影到本序列座標」後做 RANSAC 聚類來剔除離群；最後依連通性把序列分組，對每組子位姿圖以 GTSAM 做位姿圖最佳化，輸出一或多張全域一致的點雲地圖。
本頁內容
LAMM merges multi-agent, multi-session LiDAR sub-maps from different LiDAR types, adding temporal bidirectional dynamic filtering and robust loop outlier removal.
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 | 3D LiDAR of different scanning patterns: Velodyne (KITTI, WildPlaces), Ouster OS2-128 and Livox Avia (HeLiPR), Hesai 128-line (Shenzhen)、four Hikvision cameras on the Shenzhen backpack, used with R3LIVE to produce colored point clouds |
|---|---|
| 原文測試平台 | backpack device (self-collected Shenzhen dataset)、public datasets KITTI, HeLiPR and WildPlaces (carrier platforms not described in the paper) |
| 狀態估計 | Offline back end: bidirectional M-Detector dynamic removal per sequence, BTC place recognition, false-loop filtering, connectivity check that splits sequences into sub-pose graphs, and standard pose graph optimization of each sub-pose graph with the first node anchored to a reference pose, solved in GTSAM (Sec. III-B to III-E, Eq. 2). |
| 資料關聯 | BTC (Binary Triangle Combined) descriptors stored in one database per loaded sequence; each scan is searched against all databases in descending order for inner- and inter-sequence loops; rough detection by hash-table matching, fine detection by clustering transforms between triangle pairs, geometric verification by point-to-plane distance using fewer than 50 key points (Sec. III-D). |
| 時間表示 | discrete scan poses from the front-end odometry (Sec. III-A) |
| 去畸變 | not performed by LAMM; inputs are registered scans and poses from a front-end SLAM such as FAST-LIO2 (Sec. III-A, III-B) |
| 迴圈閉合 | Inner-sequence loops are rejected when the initial poses of the matched frames are geographically far apart; for inter-sequence loops, each loop projects the start position of one sequence into the other sequence's frame (Eq. 1) and RANSAC clustering of these projected points removes outliers (Sec. III-E1, Fig. 3). |
| 全域最佳化 | Pose graph optimization per connected sub-pose graph (odometry and intra- and inter-sequence loop edges, anchor prior on the first node), solved with GTSAM; no bundle adjustment or map-level refinement (Sec. III-E2, III-E3). |
| 地圖表示 | registered LiDAR scans with poses per sequence; output is one merged global point-cloud map per connected group of sequences (Sec. III-B) |
| 先驗資訊 | LiDAR scans with initial poses from a front-end SLAM (FAST-LIO2 for KITTI and HeLiPR; R3LIVE for Shenzhen); no prior relative pose between sequences, which may start anywhere (Sec. III-A, IV-A1, IV-C) |
| 可輸出幾何 | merged multi-session point-cloud map (abstract) |
| 計算需求 | Offline, C++; all experiments on a desktop with an Intel i9-13900K @ 3.0 GHz and 128 GB memory (Sec. IV); runtime is not reported. |
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 | Hesai 128-line LiDAR | 方法輸入 | Shenzhen (self-collected) | 128 lines | (Wei et al., 2025b, Sec. IV-C2; Table I) |
| LiDAR | Livox Avia | 資料集感測器 | HeLiPR (Town, Roundabout) | 原文未報告 | (Wei et al., 2025b, Sec. IV-C1; Table I) |
| LiDAR | OS2-128歸入：Ouster OS2-128 | 資料集感測器 | HeLiPR (Town, Roundabout) | 原文未報告 | (Wei et al., 2025b, Sec. IV-C1; Table I) |
| LiDAR | Velodyne (model not named) | 資料集感測器 | KITTI; WildPlaces | 原文未報告 | (Wei et al., 2025b, Table I) |
| 相機 | Hikvision cameras (four) | 方法輸入 | Shenzhen (self-collected) | four cameras; colored point cloud via R3LIVE | (Wei et al., 2025b, Sec. IV-C2) |
| 載具平台 | backpack device | 方法輸入 | Shenzhen (self-collected) | carries a Hesai 128-line LiDAR and four Hikvision cameras | (Wei et al., 2025b, Sec. IV-C2) |
| 運算硬體 | Intel i9-13900K | 執行運算平台 | 未標示 | 3.0 GHz; 128 GB memory; desktop computer | (Wei et al., 2025b, Sec. IV) |
作者報告的優勢與限制
優勢
- Merged all six KITTI test sequences; BTC-only merging failed on 00 and 02 and Disco-SLAM and DCL-SLAM failed on 02 (Sec. IV-A1, Table II)
- ATE RMSE 2.040 m on KITTI 00 and 4.746 m on 02 versus 3.330 m and 8.749 m for a single FAST-LIO2 run (Table II)
- Bi-M-Detector gave the most balanced static and dynamic classification on 141 HeLiPR Town01 frames (AA 80.65%, HA 79.86%) (Sec. IV-A3, Table IV)
- Merged Ouster OS2-128 and Livox Avia sequences into one map on HeLiPR Town and Roundabout (merged multi-LiDAR ATE 4.008 m and 2.101 m) (Table V)
- Open-source code (Sec. I)
限制
- Merging uses pose graph optimization only, with no map-level refinement of overlaps (Sec. III-E3); follow-up work reports local divergence in overlaps (Wang et al., 2026 Sec. II-B, Sec. VII-C, Fig. 9)
- Bi-M-Detector is not best in static accuracy (92.86%) or dynamic accuracy (70.05%) alone (Table IV)
- Without the loop filter, merging fails on KITTI 00 and 02 (Sec. IV-B, Table II)
- Dynamic-removal evaluation uses 141 manually selected frames of one HeLiPR sequence (Sec. IV-A3)
- No open-source map-merging baseline was available; comparison is against multi-robot SLAM systems and BTC-only merging (Sec. IV-A1)
- WildPlaces and Shenzhen results and HeLiPR merged maps appear only in supplementary material; efficiency improvement left to future work (Sec. IV-C, V)
- Runtime not reported
營建工程相關證據
未於工地測試。資料為 KITTI 市區、HeLiPR 住宅區（Town、Roundabout）、WildPlaces 自然環境，以及深圳北站附近以背包式設備（Hesai 128 線光達與四部 Hikvision 相機）蒐集的都市彩色點雲（Sec. IV、Table I）。能把機械旋轉式與 Livox Avia 等不同掃描型態、不同時間的序列合併成一張地圖，符合工地多次、多設備掃描整合的需求（推論）；但只以位姿圖合併，重疊區的局部幾何一致性沒有以參考點雲量化。
原文驗證環境：公開基準
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 40 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。
Wang et al., 2026 · Table V 本方法 12 筆
表格設定（擷取紀錄原文）：Mapping quality against the MARS-LVIG ground-truth map (DJI L1 LiDAR processed with DJI Terra); AWD average Wasserstein distance, CD Chamfer distance, SCS spatial consistency score, MME mean map entropy; all lower is better; metrics as defined in MapEval (Wang et al., 2026, Table V)
AWD(m)，MARS-LVIG · Island
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 Wang et al., 2026 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(Wang et al., 2026, Table V)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LAMM本方法 | 1.3 m | (Wang et al., 2026, Table V) |
| Ours原文提出 | 0.3 m | (Wang et al., 2026, Table V) |
Wang et al., 2026 · Table IV 本方法 10 筆
指標RMSE of the ATE (m)
表格設定（擷取紀錄原文）：Multi-robot localization; RMSE of ATE (m); failure (x) = any sequence with RMSE above 30 m; GEODE, MARS-LVIG and S3E sequences split into sessions with unknown relative transforms (Wang et al., 2026, Table IV)
RMSE of the ATE (m)，S3E · Campus 3
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 Wang et al., 2026 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(Wang et al., 2026, Table IV)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| DCL-SLAM | 17.89 m | (Wang et al., 2026, Table IV) |
| LAMM本方法 | 12.51 m | (Wang et al., 2026, Table IV) |
| Ours原文提出 | 3.51 m | (Wang et al., 2026, Table IV) |
Wei et al., 2025b · Table II 本方法 6 筆
指標RMSE of the ATE (m)
表格設定（擷取紀錄原文）：KITTI sequences split into overlapping sessions; FAST-LIO2 initial odometry; FAST-LIO2 column is the single-run result; 'Fail' = merging failed; ablations without M-detector or loop filter, and variants replacing BTC by SOLiD or the RANSAC filter by PCM; the 'Origin' per-session column is not extracted (Wei et al., 2025b, Table II)
RMSE of the ATE (m)，KITTI · 00
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。
- 失敗
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 Wei et al., 2025b 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(Wei et al., 2025b, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| FAST-LIO2 (single run) | 3.33 m | (Wei et al., 2025b, Table II) |
| Disco-SLAM | 2.097 m | (Wei et al., 2025b, Table II) |
| DCL-SLAM | 4.024 m | (Wei et al., 2025b, Table II) |
| BTC (loop-detection-only merging) | 無數值失敗註記（擷取紀錄）：failed | (Wei et al., 2025b, Table II) |
| LAMM full model本方法原文提出 | 2.04 m | (Wei et al., 2025b, Table II) |
| LAMM wo M-detector | 2.392 m | (Wei et al., 2025b, Table II) |
| LAMM wo loop filter | 無數值失敗註記（擷取紀錄）：failed | (Wei et al., 2025b, Table II) |
| LAMM-SOLiD | 2.061 m | (Wei et al., 2025b, Table II) |
| LAMM-PCM | 2.707 m | (Wei et al., 2025b, Table II) |
Wei et al., 2025b · Table V 本方法 6 筆
指標RMSE of the ATE (m)
表格設定（擷取紀錄原文）：HeLiPR Town and Roundabout: FAST-LIO2 ATE of each input sequence, ATE after merging the three sequences of the same LiDAR type, and after merging all six Ouster and Avia sequences (single value per environment) (Wei et al., 2025b, Table V)
RMSE of the ATE (m)，HeLiPR · Town Ouster 1-3 merged
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 Wei et al., 2025b 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 |
|---|---|---|
| LAMM merged (Ouster sequences)本方法原文提出 | 2.818 m | (Wei et al., 2025b, Table V) |
其他比較組
來源
Wei et al., 2025b
（2025）Large-Scale Multi-Session Point-Cloud Map MergingIEEE Robotics and Automation Letters, 10(1):88-95
DOI 10.1109/lra.2024.3504317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
- 程式碼釋出：hku-mars/LAMM https://github.com/hku-mars/LAMM
程式碼：https://github.com/hku-mars/LAMM。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。