interactive_slam
interactive_slam 讓使用者透過圖形介面修正自動 3D LiDAR SLAM 產生的地圖。系統把自動 SLAM 的位姿約束與使用者建立的修正約束放在同一個位姿圖中，以 g2o 最佳化並立即顯示結果。修正工具有三種：使用者指定迴圈兩端後以 FPFH 初配準、再以 GICP 精配準的半自動迴圈閉合，以及之後的自動迴圈搜尋；依卡方距離抽樣、重新匹配不一致邊的位姿約束精修；以及點選平面後以區域成長與 RANSAC 擷取平面，並在遠處平面之間加入相同、平行或垂直約束，以修正整體彎曲。
本頁內容
Human-in-the-loop correction of 3D LiDAR pose graphs: user-guided and automatic loop closing, chi-square-sampled edge refinement and user-selected plane constraints (identity, parallel, perpendicular) applied on top of any ROS SLAM output.
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 (16-line in the indoor-outdoor sequence; model not reported)、input is any ROS SLAM pose graph or odometry sequence |
|---|---|
| 原文測試平台 | mobile mapping platform (PASCO Mobile Measurement System)、原文未報告 (carrier of the Kaarta Stencil not stated) |
| 狀態估計 | pose-graph optimization (g2o, Levenberg-Marquardt) over automatic SLAM constraints plus user-created correction constraints, re-optimized after every edit (Sec. III.A) |
| 資料關聯 | for user-selected loop pairs, FPFH-based initial alignment the user can fine-tune, then GICP (ICP or NDT selectable); automatic loop search by distance, graph-path and matching-score thresholds with a Huber kernel (Sec. III.B) |
| 時間表示 | discrete poses (keyframes every 5 m outdoors, every 3 m in the indoor-outdoor sequence) |
| 去畸變 | 不適用 (operates on keyframes from an upstream SLAM) |
| 迴圈閉合 | semi-automatic: the user picks the two ends of a loop, then automatic scan-matching loop search runs on the roughly corrected graph (Sec. III.B) |
| 全域最佳化 | pose graph with plane vertices; identity, parallel and perpendicular constraints between user-selected planes with a large information matrix (e.g., 10^3 I) as semi-global corrections; pose-edge refinement re-matches inconsistent edges sampled by chi-square distance (Sec. III.C-D) |
| 地圖表示 | keyframe point clouds on a pose graph with plane vertices |
| 先驗資訊 | user knowledge of planar structure (identity, parallel, perpendicular) added through the GUI; initial pose graph or odometry from an automatic SLAM |
| 可輸出幾何 | corrected, globally consistent 3D point cloud map and pose graph |
| 計算需求 | 原文未報告; correction took about 15 minutes for the 20-minute outdoor dataset and about 20 minutes for the indoor-outdoor sequence, including user interaction (Sec. IV) |
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 |
|---|---|---|---|---|---|
| 行動掃描設備 | Mobile Measurement System | 方法輸入 | 未標示 | mobile platform with a 3D LIDAR, rangefinders, GNSS and cameras | (Koide et al., 2021a, Sec. IV.A; Fig. 4) |
| 行動掃描設備 | Kaarta Stencil | 方法輸入 | 未標示 | commercial visual-LIDAR-IMU mapping system; its own trajectory is shown as a baseline (Fig. 7a) | (Koide et al., 2021a, Sec. IV.B) |
| 全測站 | static total stations (PASCO Mobile Measurement System) | 參考或真值量測 | 未標示 | several static total stations placed in the environment; 3D LiDAR position measured to a few millimetres (translation only) | (Koide et al., 2021a, Sec. IV.A) |
作者報告的優勢與限制
優勢
- On the PASCO outdoor dataset ATE fell from 1.731 m (LOAM base) to 1.683 m after loop closing, 1.340 m after edge refinement and 0.517 m with plane constraints, below all automatic baselines (1.731 to 5.248 m) (Table I; Sec. IV.A).
- The outdoor correction took about 15 minutes for a 20-minute dataset with one manual loop and 15 planes (Sec. IV.A).
- Produced a consistent multi-floor indoor-outdoor map where Kaarta Stencil, LOAM variants, SuMa, ethzasl_icp_mapping and hdl_graph_slam gave corrupted trajectories (Sec. IV.B).
- Pose-constraint refinement reduced the example graph's chi-square distance from 0.447 to 0.061 (Sec. III.C).
限制
- Requires manual effort: 19 loops and 17 planes, about 20 minutes, for the indoor-outdoor sequence (Sec. IV.B.3).
- Automatic loop detectors (distance histogram, ScanContext, LiDAR-Iris, OverlapNet) degraded strongly indoors, and repeated structures across floors gave near-identical descriptor distances (Sec. IV.B.2).
- Loop closing alone can worsen global accuracy: LeGO-LOAM's ATE rose from 1.846 m to 2.418 m with loops because of altitude bending (Table I; Fig. 6).
- Ground truth gives translation only, so RTE uses a modified sub-trajectory alignment (Sec. IV.A).
- (inference) Quantitative accuracy is reported only for the outdoor dataset; the indoor-outdoor result is qualitative.
營建工程相關證據
未在營建工地驗證。戶外評估以 PASCO 行動量測系統記錄，並以環境中多台靜置全測站追蹤 3D LiDAR 位置作為毫米級參考軌跡（Sec. IV.A）；室內外資料包含多樓層與樓梯（Sec. IV.B）。以人工指定平面的相同、平行或垂直關係修正整體彎曲，可直接用於竣工點雲品質管制中修正樓板與牆面的彎曲，但這也把人工先驗注入地圖，應記錄所加約束，以免掩蓋結構真實的變形或施工偏差（推論）。
原文驗證環境：獨立參考量測、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 2 個比較組，合計 14 筆紀錄。
Koide et al., 2021a · Table I 本方法 12 筆
資料集與序列PASCO Mobile Measurement System outdoor dataset (released via SMRT-AIST) · outdoor sequence (about 20 min)
表格設定（擷取紀錄原文）：PASCO Mobile Measurement System outdoor dataset (about 200 m x 400 m, 20 min); ground truth is the 3D LiDAR position tracked by static total stations (translation only); errors given as mean +/- std; RTE uses a modified routine that aligns each sub-trajectory within the evaluation window; the proposed rows start from LOAM odometry with 5 m keyframes (Koide et al., 2021a, Table I)
RTE(5m) [m], mean +/- std = 0.044 +/- 0.017，PASCO Mobile Measurement System outdoor dataset (released via SMRT-AIST) · outdoor sequence (about 20 min)
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 Koide et al., 2021a 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(Koide et al., 2021a, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Proposed: Edge refinement (with loop closure)本方法原文提出 | 0.044 m | (Koide et al., 2021a, Table I) |
| Proposed: Plane constraints (with loop closure)本方法原文提出 | 0.044 m | (Koide et al., 2021a, Table I) |
Koide et al., 2021a · Text Sec. III.C 本方法 2 筆
資料集與序列example dataset of Fig. 3
表格設定（擷取紀錄原文）：Example correction sequence (Fig. 3): chi-square distance of the pose graph before pose-constraint refinement (Koide et al., 2021a, Text Sec. III.C)
chi-square distance (sum of weighted constraint errors) before refinement，example dataset of Fig. 3
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 Koide et al., 2021a 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 |
|---|---|---|
| before pose constraint refinement本方法 | 0.447 | (Koide et al., 2021a, Sec. III.C) |
來源
Koide et al., 2021a
（2021）Interactive 3D Graph SLAM for Map CorrectionIEEE Robotics and Automation Letters, 6(1):40-47
DOI 10.1109/lra.2020.3028828程式碼
同儕審查已出版已讀全文近十年
相關版本
- 程式碼釋出：koide3/interactive_slam (GPL-3.0) https://github.com/koide3/interactive_slam
- 資料集：Outdoor evaluation dataset announced as available at github.com/SMRT-AIST (footnote 2; not checked) https://github.com/SMRT-AIST
- companion：hdl_graph_slam (Koide et al., 2019), whose pose graphs interactive_slam can edit https://github.com/koide3/hdl_graph_slam
程式碼：https://github.com/koide3/interactive_slam（授權：GPL-3.0 (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。