Velodyne SLAM
Velodyne SLAM 專為 Velodyne HDL-64E 的連續旋轉取樣與較高量測雜訊設計，只使用 LiDAR 資料。每轉一圈的資料排成 870 乘 64 的距離影像，先估計各點的法向量與平面信心，再以位置加法向量的 6D 最近鄰 ICP 將掃描對整張地圖配準，並在配準前後各做一次以線性內插的去畸變。地圖是每格最多保存一個曲面元素的 3D 網格；在平面信心高的區域，新量測會先沿法向量調整再加入，以降低雜訊造成的厚度。作者另提出離線精修步驟，以儲存的原始量測重建更細緻的地圖。
本頁內容
LiDAR-only scan-to-map ICP SLAM for the Velodyne HDL-64E with two-pass linear de-skewing and normal-guided measurement adaptation in a surface-grid map, plus offline map refinement; no loop closure.
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 spinning LiDAR only (Velodyne HDL-64E S2); no wheel-speed, inertial or other information |
|---|---|
| 原文測試平台 | vehicle (experimental vehicle AnnieWay) |
| 狀態估計 | scan-to-map ICP minimizing point-to-plane distances (Chen-Medioni) with nearest neighbours searched in 6D (position and normal), initialized by constant-motion prediction; 1000 surfaces sampled from the upper and 500 from the lower half of the range image (Sec. II.C) |
| 資料關聯 | 6D nearest neighbour (px, py, pz, nx, ny, nz) between scan surfaces and map surfaces (Sec. II.C) |
| 時間表示 | one pose at the end of each 360 deg turn; poses within a turn linearly interpolated assuming constant velocity (Sec. II.C) |
| 去畸變 | linear-interpolation de-skewing applied twice, before ICP with the predicted pose and after ICP with the final pose (Sec. II.D) |
| 迴圈閉合 | none (named as future work) |
| 全域最佳化 | none |
| 地圖表示 | 3D grid of resolution g (5 cm default) holding at most one surface (point, normal, normal confidence) per cell; in flat regions new measurements are moved along their normals to fit neighbouring surfaces before insertion; cells are replaced by closer or more confident measurements and never erased (Sec. II.E) |
| 先驗資訊 | none |
| 可輸出幾何 | vehicle trajectory and a detailed 3D point cloud of surfaces; an offline refinement builds a new map from the stored original measurements (Sec. II.F) |
| 計算需求 | 原文未報告 (the authors state that a low grid value gives precise maps and a higher value faster on-line processing) |
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 HDL-64E S2歸入：Velodyne HDL-64E | 方法輸入 | Velodyne SLAM dataset (AnnieWay, two scenarios) | 64 laser diodes over about 26 deg pitch, continuous 360 deg rotation at about 10 Hz; each turn arranged as an 870 x 64 range image; relatively high measurement noise | (Moosmann & Stiller, 2011, Sec. I; Sec. II-A; Sec. II-D) |
| GNSS 接收器 | integrated navigation system (INS) fusing GPS, wheel speed sensors and inertial measurements (model not reported) | 比較對象設備 | Velodyne SLAM dataset (AnnieWay, two scenarios) | local errors higher than those of the proposed method; not usable as ground truth | (Moosmann & Stiller, 2011, Sec. III; Table I) |
| 載具平台 | experimental vehicle AnnieWay | 方法輸入 | Velodyne SLAM dataset (AnnieWay, two scenarios) | scanner mounted on top of the vehicle; scenario 1 about 1.3 km and scenario 2 about 1.1 km, both with a bridge | (Moosmann & Stiller, 2011, Sec. I; Sec. III; Fig. 4) |
作者報告的優勢與限制
優勢
- With mapping, de-skewing and adaptation the end-point error was 2.29 m after the 1.3 km loop and 4.10 m after the 1.1 km loop without loop closure, versus 3.30 m and 2.64 m for the INS (Table I).
- Pairwise scan matching without mapping gave 19.60 m and 22.25 m, so scan-to-map matching is the main factor; de-skewing is the second (Table I; Sec. III).
- The grid resolution can vary over a wide range with little effect on localization error; only map detail suffers (Sec. III; Fig. 5).
- Adaptation and offline refinement produce visibly more detailed surfaces such as road surfaces (Figs. 6-7).
- Dataset released publicly (Sec. III).
限制
- No loop closure (Sec. IV).
- Limited to (nearly) static scenes (Sec. IV).
- Map quality is assessed only by visual inspection (Sec. III).
- Ground truth is only the end position obtained by matching the last scan to the first, because the INS had larger local errors than the method (Sec. III).
- (inference) Runtime is not reported, so real-time capability is not established in the paper.
營建工程相關證據
論文未在施工或建築環境測試，只在兩段含橋樑的戶外多層車載路線上評估。它把量測點沿法向量移動以壓低平面厚度，並提供離線地圖精修，目標是可作城市模型的細緻點雲；這種降低平面雜訊的做法與施工點雲的平面厚度問題相關，但論文只以目視檢查地圖品質，屬推論。
原文驗證環境：受控實驗
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 1 個比較組，合計 16 筆紀錄。
Moosmann & Stiller, 2011 · Table I 本方法 16 筆
指標End-point error
表格設定（擷取紀錄原文）：End-point error (Euclidean distance between estimated end position and the end position obtained by ICP of the last scan to the first scan) on two loop scenarios; no loop closure; map parts beyond 50 m discarded; g = 5 cm (Moosmann & Stiller, 2011, Table I)
End-point error，Velodyne SLAM dataset (AnnieWay) · Scenario 1 (1.3 km)
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 Moosmann & Stiller, 2011 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(Moosmann & Stiller, 2011, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| IMU (integrated navigation system: GPS, wheel speed, inertial) | 3.3 m | (Moosmann & Stiller, 2011, Table I) |
| Setting 1 (mapping no, de-skewing no, adaptation no)本方法原文提出 | 19.6 m | (Moosmann & Stiller, 2011, Table I) |
| Setting 2 (mapping no, de-skewing no, adaptation yes)本方法原文提出 | 19.33 m | (Moosmann & Stiller, 2011, Table I) |
| Setting 3 (mapping no, de-skewing yes, adaptation no)本方法原文提出 | 27.41 m | (Moosmann & Stiller, 2011, Table I) |
| Setting 4 (mapping no, de-skewing yes, adaptation yes)本方法原文提出 | 27.21 m | (Moosmann & Stiller, 2011, Table I) |
| Setting 5 (mapping yes, de-skewing no, adaptation no)本方法原文提出 | 4.47 m | (Moosmann & Stiller, 2011, Table I) |
| Setting 6 (mapping yes, de-skewing no, adaptation yes)本方法原文提出 | 4.13 m | (Moosmann & Stiller, 2011, Table I) |
| Setting 7 (mapping yes, de-skewing yes, adaptation no)本方法原文提出 | 2.9 m | (Moosmann & Stiller, 2011, Table I) |
| Setting 8 (mapping yes, de-skewing yes, adaptation yes)本方法原文提出 | 2.29 m | (Moosmann & Stiller, 2011, Table I) |
來源
Moosmann & Stiller, 2011
（2011）Velodyne SLAM2011 IEEE Intelligent Vehicles Symposium (IV), Baden-Baden, Germany, pp. 393-398
DOI 10.1109/ivs.2011.5940396程式碼
同儕審查已出版已讀全文經典
相關版本
- author copy：Moosmann_IV11.pdf on the KIT MRT server (not read; VoR read instead) http://www.mrt.kit.edu/z/publ/download/Moosmann_IV11.pdf
- 資料集：Velodyne SLAM dataset (scenario 1 and 2, results of the paper without loop closure) https://www.mrt.kit.edu/z/publ/download/velodyneslam/dataset.html
- 程式碼釋出：Source code for joint self-localization and tracking of generic objects in 3D range data (SLAM + DATMO), linked from the Velodyne SLAM page as the complete development source code https://www.mrt.kit.edu/z/publ/download/velodynetracking/code.html
程式碼：https://www.mrt.kit.edu/z/publ/download/velodynetracking/code.html（授權：GNU General Public License (version not stated on the code page)）。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。