Plane-adjustment LiDAR SLAM (indoor)
這個方法以平面作為室內 LiDAR SLAM 的地標，類比視覺 SLAM 的光束法平差，聯合最佳化關鍵影格位姿與平面參數，作者稱為平面平差。定位執行緒以前向 ICP 流把上一幀的平面點追蹤到目前幀，直接得到局部對全域平面的對應，並在配準中逐點內插掃描內運動；局部建圖在 8 個關鍵影格的滑動視窗內做平面平差，全域建圖則在舊平面被再次觀測時觸發全域平差，不必回到原地即可修正漂移。平面法向量一律指向感測器，用以區分牆與門的兩個表面，避免最近鄰搜尋把兩面混在一起。
本頁內容
Indoor LiDAR SLAM with planes as landmarks: forward ICP flow tracks plane points scan to scan for real-time registration to global planes with in-scan interpolation, local plane adjustment over an 8-keyframe window, and global plane adjustment triggered when planes (not places) are revisited; sensor-facing normals separate the two sides of walls and doors.
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 only (Velodyne VLP-16 data recorded by a NavVis M6 device) |
|---|---|
| 原文測試平台 | NavVis M6 mobile mapping device (carrying mode not described) |
| 狀態估計 | three threads in an ORB-SLAM-like structure: real-time scan-to-global-plane registration linearized to first order and iterated up to 5 times (threshold 0.5 deg) with bisquare weights; local plane adjustment (LM) over a sliding window of 8 keyframes and their planes with older poses fixed; global plane adjustment (LM) over all keyframe poses and planes; all with point-to-plane costs (Sec. V-C; Sec. VI; Sec. VII) |
| 資料關聯 | forward ICP flow tracks each plane's points from scan k-1 into scan k (2 nearest neighbours, RANSAC plane fit, expansion within 5 cm, at least 30 points, normal change below 15 deg); new planes matched to global planes by normal angle below 10 deg and mean point distance below 5 cm, plus a geometric consistency check; plane normals oriented toward the sensor to separate the two sides of walls and doors (Sec. IV-2; Sec. V; Sec. VI-A) |
| 時間表示 | discrete keyframe poses; within a scan the relative motion is linearly interpolated per point (scaled angle-axis rotation and translation) (Sec. IV-1; Sec. V-C) |
| 去畸變 | per-point linear interpolation of the relative pose inside the registration cost; new keyframe scans are undistorted before plane detection; the first frame is assumed static (Sec. V-A; Sec. V-C; Sec. VI-A) |
| 迴圈閉合 | plane-revisit criterion: global plane adjustment is triggered when previously mapped planes are re-associated, which can happen far from where they were first seen, instead of when a place is revisited (Sec. I; Sec. VI; Sec. VII; Fig. 2) |
| 全域最佳化 | global plane adjustment jointly refining all keyframe poses and plane parameters (closest-point parameterization) with point-to-plane residuals, solved with the authors' efficient planar bundle adjustment method; per-plane integrated cost matrices updated afterwards (Sec. VII) |
| 地圖表示 | global plane landmarks (closest-point parameterization) with their supporting points and keyframe poses; a 4 x 4 integrated cost matrix per plane summarizes observations outside the sliding window (Sec. IV-2; Sec. VI-E; Sec. VI-F) |
| 先驗資訊 | none (assumes the LiDAR starts stationary) |
| 可輸出幾何 | keyframe trajectory and plane-segmented point map in which both sides of planar objects such as walls and doors are reconstructed separately (Fig. 8) |
| 計算需求 | 3.1 GHz Intel i7 CPU with 16 GB memory; localization about 42.1 ms and 45.5 ms per scan on datasets B and C (VLP-16 scan period 100 ms), local mapping about 25 ms, global plane adjustment 395.6 +/- 301.9 ms and 420.2 +/- 330.2 ms (Sec. VIII-B; Table II) |
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 | Velodyne VLP-16 | 方法輸入 | own indoor datasets A-D (NavVis M6) | on the NavVis M6; the only data used by the evaluated algorithms; scan period 100 ms | (Zhou et al., 2021, Sec. VIII-A; Sec. VIII-B) |
| LiDAR | Hokuyo single-layer LiDARs (three units; model not reported) | 參考或真值量測 | own indoor datasets A-D (NavVis M6) | part of the NavVis M6; used by NavVis for its reference trajectory | (Zhou et al., 2021, Sec. VIII-A) |
| 行動掃描設備 | NavVis M6 | 參考或真值量測 | own indoor datasets A-D (NavVis M6) | trajectory estimated by the NavVis system through hours of offline fusion is used as ground truth | (Zhou et al., 2021, Sec. VIII-A) |
| 慣性量測單元（IMU） | NavVis M6 IMU (model not reported) | 參考或真值量測 | own indoor datasets A-D (NavVis M6) | used by NavVis for its reference trajectory | (Zhou et al., 2021, Sec. VIII-A) |
| 運算硬體 | Intel i7 CPU at 3.1 GHz (model not reported) | 執行運算平台 | 未標示 | 16 GB memory | (Zhou et al., 2021, Sec. VIII-B) |
作者報告的優勢與限制
優勢
- Keyframe ATE of 0.033-0.048 m on four indoor datasets, versus 0.082-0.16 m for pi-LSAM and 0.21-0.34 m for BALM (Table I)
- LeGO-LOAM failed on three of four sequences and BALM on one, with large rotations; the proposed method completed all (Table I; Fig. 10)
- Local and global plane adjustment each reduce ATE substantially relative to the variants without them (Table I; Fig. 9)
- Loop correction can happen when a wall is revisited from a remote place, before returning to the start (Sec. I; Fig. 2; Fig. 9)
- Separates the two sides of walls and doors, whereas BALM mixed them on dataset B (Fig. 8; Fig. 11)
- Localization runs in about 42-46 ms per 100 ms scan and local plane adjustment is 32 to 37 times faster than direct minimization (Sec. VIII-B)
限制
- Designed for indoor environments where planes are abundant (title; Sec. I) (inference that plane-poor scenes are not covered)
- Evaluated on four own indoor datasets from one device; ground truth is the NavVis trajectory fused offline (Sec. VIII-A)
- Global plane adjustment time has a large standard deviation because it is triggered at different places (Sec. VIII-B)
- Assumes the LiDAR starts from a stationary state (Sec. V-A)
- No public code (not stated in the paper; none found)
營建工程相關證據
論文以 NavVis M6 在四個室內場景蒐集資料，並以 NavVis 結合其他 LiDAR、IMU、WiFi 訊號與人工地面控制點經數小時離線融合的軌跡作為參考，屬既有建築室內環境。它以平面為地標聯合最佳化平面與位姿，並能分開重建牆與門的兩個表面，直接對應室內點雲中牆面重影與厚度的問題；但施工中未完成的牆體、臨時構件與堆料會破壞平面假設，論文沒有驗證，屬推論。
原文驗證環境：已完工建築、獨立參考量測
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 3 個比較組，合計 38 筆紀錄。
Zhou et al., 2021 · Table II 本方法 18 筆
資料集與序列own indoor datasets A-D (NavVis M6) · B (294.0 m)
表格設定（擷取紀錄原文）：Runtime (ms, mean +/- std) of components on datasets B and C (Zhou et al., 2021, Table II)
Localization: Forward ICP Flow runtime (mean, std 5.1)，own indoor datasets A-D (NavVis M6) · B (294.0 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 Zhou et al., 2021 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 |
|---|---|---|
| Ours, Localization: Forward ICP Flow本方法原文提出硬體：3.1 GHz Intel i7 CPU, 16 GB memory | 25.6 ms | (Zhou et al., 2021, Table II) |
Zhou et al., 2021 · Table I 本方法 12 筆
指標keyframe ATE (m), median of 5 runs
表格設定（擷取紀錄原文）：Keyframe ATE (m), median of 5 runs, on four indoor NavVis M6 datasets (VLP-16 data only) with large rotations; ground truth = offline fused NavVis trajectory; '-' = failed to complete (Zhou et al., 2021, Table I)
keyframe ATE (m), median of 5 runs，own indoor datasets A-D (NavVis M6) · A (261.9 m)
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 Zhou et al., 2021 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(Zhou et al., 2021, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LeGO-LOAM [8] | 無數值失敗註記（擷取紀錄）：failed (did not complete the sequence) | (Zhou et al., 2021, Table I) |
| BALM [29] | 0.34 m | (Zhou et al., 2021, Table I) |
| pi-LSAM [22] | 0.082 m | (Zhou et al., 2021, Table I) |
| Ours - LPA - GPA本方法原文提出 | 0.29 m | (Zhou et al., 2021, Table I) |
| Ours - GPA本方法原文提出 | 0.18 m | (Zhou et al., 2021, Table I) |
| Ours本方法原文提出 | 0.039 m | (Zhou et al., 2021, Table I) |
Zhou et al., 2021 · Text Sec. VIII-B 本方法 8 筆
表格設定（擷取紀錄原文）：Runtime statements in text for datasets B and C (Zhou et al., 2021, Text Sec. VIII-B)
average runtime of the localization component，own indoor datasets A-D (NavVis M6) · B (294.0 m)
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 Zhou et al., 2021 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(Zhou et al., 2021, Text Sec. VIII-B)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Ours, localization本方法原文提出硬體：3.1 GHz Intel i7 CPU, 16 GB memory | 42.1 ms | (Zhou et al., 2021, Sec. VIII-B) |
| pi-LSAM [22], localization硬體：3.1 GHz Intel i7 CPU, 16 GB memory | 62.3 ms | (Zhou et al., 2021, Sec. VIII-B) |
來源
Zhou et al., 2021
（2021）LiDAR SLAM With Plane Adjustment for Indoor EnvironmentIEEE Robotics and Automation Letters, 6(4):7073-7080
同儕審查已出版已讀全文近十年查證後修正
相關版本
- accepted manuscript：Author copy on the Kaess lab publication page (RA-L preprint version, accepted June 2021) https://www.cs.cmu.edu/~kaess/pub/Zhou21ral2.pdf
- 前身方法：pi-LSAM: LiDAR Smoothing and Mapping With Planes (L. Zhou, S. Wang, M. Kaess; ICRA 2021); compared in Table I; not read https://doi.org/10.1109/ICRA48506.2021.9561933