MRS continuous-time surfel SLAM (Droeschel and Behnke)
這個方法延續作者的局部多解析度網格地圖：每個 3D 掃描以面元（surfel）配準到以機器人為中心的局部地圖，多個局部地圖再以面元配準連成全域位姿圖。新意在於把每個局部地圖內的掃描位姿建成子圖，形成階層式圖：當地圖累積更多資訊後，可挑選配準不確定性最大的掃描重新對齊並最佳化子圖，再更新上層位姿圖；子圖內以 SE(3) 三次 B 樣條表示連續時間軌跡，內插每條掃描線的位姿以修正掃描期間的運動畸變。作者以平均地圖熵量化點雲清晰度。
本頁內容
Hierarchical LiDAR SLAM built on local multiresolution surfel maps: scans in each local map form a sub-graph that can be re-aligned and re-optimized online when more data arrive, local maps form a g2o pose graph with loop closures, and a cubic B-spline in SE(3) per sub-graph interpolates scan-line poses; map crispness is quantified by mean map entropy.
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 rotating multi-beam LiDAR (Velodyne VLP-16; two VLP-16 on the Deutsches Museum backpack)、optional IMU or wheel odometry used as registration prior and for motion during acquisition (Sec. III); the MAV carried an IMU measuring attitude (Sec. V-A) |
|---|---|
| 原文測試平台 | UAV (DJI Matrice 600)、backpack (Deutsches Museum dataset) |
| 狀態估計 | hierarchical graph optimization in g2o: an allocentric pose graph of local multiresolution maps, a sub-graph of 3D scan poses per local map, and a continuous-time cubic B-spline over scan poses; sub-graphs are refined in parallel and global optimization is triggered by loop closures or when a sub-graph reference pose changes by more than 0.01 m or 1 degree (Sec. IV) |
| 資料關聯 | surfel-based registration of each 3D scan to a robot-centric local multiresolution grid map (surfel = sample mean and covariance of the points in a cell), and surfel-based map-to-map registration between local maps; scans with the largest entropy of registration covariance are selected for realignment (Sec. III; Sec. IV-A) |
| 時間表示 | continuous-time: cumulative cubic B-spline in SE(3) with scan nodes as control points, used to interpolate the pose of each scan line (one VLP-16 data packet of 24 firing sequences) (Sec. IV-C; Sec. V-A) |
| 去畸變 | motion during acquisition compensated with IMU or wheel odometry when available, then scan-line poses refined by B-spline interpolation within each sub-graph (Sec. III; Sec. IV-C; Fig. 4) |
| 迴圈閉合 | after each new local map one candidate map node is drawn with a probability that decays with distance and registered by surfel map-to-map alignment; on revisits, scans from neighbouring map nodes enlarge the local window (Sec. IV-B; Sec. IV-D) |
| 全域最佳化 | pose graph over local maps optimized with g2o; changes are propagated to the sub-graphs and vice versa (Sec. IV; Sec. IV-D) |
| 地圖表示 | robot-centric local multiresolution grid maps that store measurements, occupancy (ray casting with an approximated 3D Bresenham) and surfels per cell; allocentric graph of local maps with a new map node every 5 m (Sec. III; Sec. V) |
| 先驗資訊 | none |
| 可輸出幾何 | refined 3D point cloud map and trajectory (Figs. 5 and 7) |
| 計算需求 | Intel Core i7-6700HQ at 2.6 GHz with 32 GB RAM; one refinement iteration takes 54 ms for a single map node and 380 ms for all 16 map nodes in parallel on the courtyard data, averaged over 10 runs (Sec. V; Sec. V-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 | Velodyne VLP-16 | 方法輸入 | courtyard MAV flight (own data) | about 300,000 range measurements per second, 16 rings, 30 deg vertical field of view, 100 m maximum range, up to 1200 rpm; a scan line is one data packet of 24 firing sequences of 1.33 ms | (Droeschel & Behnke, 2018, Sec. IV; Sec. V-A) |
| LiDAR | Velodyne VLP-16 (two units)歸入：Velodyne VLP-16 | 資料集感測器 | Deutsches Museum dataset (Google Cartographer team) | one mounted horizontally and one vertically on a backpack; calibration provided with the data and refined in the graph | (Droeschel & Behnke, 2018, Sec. V-B) |
| 慣性量測單元（IMU） | IMU measuring attitude (model not reported) | 方法輸入 | courtyard MAV flight (own data) | 原文未報告 | (Droeschel & Behnke, 2018, Sec. V-A) |
| 載具平台 | DJI Matrice 600 | 方法輸入 | courtyard MAV flight (own data) | MAV flown by a human operator along a building front at different heights; 2000 scans in 200 s | (Droeschel & Behnke, 2018, Sec. V-A) |
| 載具平台 | backpack (carried through the museum) | 資料集感測器 | Deutsches Museum dataset (Google Cartographer team) | parts of the data contain moving persons | (Droeschel & Behnke, 2018, Sec. V-B) |
| 運算硬體 | Intel Core i7-6700HQ | 執行運算平台 | 未標示 | quad-core at 2.6 GHz, 32 GB RAM | (Droeschel & Behnke, 2018, Sec. V) |
作者報告的優勢與限制
優勢
- Lowest mean map entropy on the Deutsches Museum subset among Cartographer, the authors' previous method and Nuechter et al. continuous-time refinement (Table I)
- Refinement corrects misaligned scans that the previous method left in the courtyard map (Sec. V-A; Fig. 7)
- Covariance-based scan selection converges faster than refinement without it (Sec. V-B; Fig. 8)
- Local, allocentric and refinement stages run independently, allowing online mapping while earlier data are refined (Sec. III)
限制
- Evaluation is qualitative plus mean map entropy; no trajectory or map error against an independent reference (abstract; Sec. V)
- Deutsches Museum comparison uses only a selected part of the dataset, following Nuechter et al. (Sec. V-B)
- Only refinement iteration times are reported, not end-to-end online runtime (Sec. V-A)
- Mean map entropy measures crispness and does not show absolute geometric accuracy (inference)
營建工程相關證據
論文未在施工現場測試；多旋翼機在建築中庭沿建物立面飛行，另一組資料是背包在德意志博物館室內行走。它以平均地圖熵衡量點雲清晰度，而不是以獨立參考量測幾何精度；這類不需真值的點雲品質指標可作為施工點雲品質評估的參考，但這屬推論。
原文驗證環境：受控實驗、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 3 筆紀錄。
Droeschel & Behnke, 2018 · Text Sec. V-A 本方法 2 筆
資料集與序列courtyard MAV flight (own data) · 2000 scans, 200 s
表格設定（擷取紀錄原文）：Courtyard MAV data (16 map nodes); refinement run as post-processing, average over 10 runs (Droeschel & Behnke, 2018, Text Sec. V-A)
runtime per iteration for refining a single map node，courtyard MAV flight (own data) · 2000 scans, 200 s
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 Droeschel & Behnke, 2018 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 (single map node)本方法原文提出硬體：Intel Core i7-6700HQ quad-core 2.6 GHz, 32 GB RAM | 54 ms | (Droeschel & Behnke, 2018, Sec. V-A) |
Droeschel & Behnke, 2018 · Table I 本方法 1 筆
指標mean map entropy (MME)
資料集與序列Deutsches Museum (Cartographer dataset) · selected part (as in [27])
表格設定（擷取紀錄原文）：Best mean map entropy (MME, radius 0.5 m, lower is better) on a selected part of the Deutsches Museum backpack dataset, following Nuechter et al. [27] (Droeschel & Behnke, 2018, Table I)
mean map entropy (MME)，Deutsches Museum (Cartographer dataset) · selected part (as in [27])
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 Droeschel & Behnke, 2018 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(Droeschel & Behnke, 2018, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Cartographer [26] | -2.04 | (Droeschel & Behnke, 2018, Table I) |
| Droeschel et al. [8] (previous method) | -2.12 | (Droeschel & Behnke, 2018, Table I) |
| Nuechter et al. [27] | -2.34 | (Droeschel & Behnke, 2018, Table I) |
| Ours本方法原文提出 | -2.42 | (Droeschel & Behnke, 2018, Table I) |
來源
Droeschel & Behnke, 2018
（2018）Efficient Continuous-Time SLAM for 3D Lidar-Based Online Mapping2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 5000-5007
DOI 10.1109/icra.2018.8461000arXiv 1810.06802
同儕審查已出版已讀全文近十年
相關版本
- 預印本：Efficient Continuous-time SLAM for 3D Lidar-based Online Mapping (arXiv v1, posted after the conference) https://arxiv.org/abs/1810.06802