D2D-NDT
此研究把固定與移動兩片掃描都轉成三維常態分布轉換（3D-NDT）模型，也就是在規則網格的每個格子以一個高斯分布描述局部表面，再直接最小化兩個模型之間的 L2 距離（分布對分布，D2D），不像點對分布（P2D）或 ICP 那樣逐點計算。目標函數只在兩模型彼此最近的高斯成分之間評估，具有解析梯度與 Hessian，以牛頓法搭配 More-Thuente 線搜尋求解，並在 4、2、1、0.5 m 的網格上逐層配準；每次迭代直接轉換整個 NDT 模型並累積齊次轉換矩陣，以處理 SE(3) 的結構。作者另以 3D-NDT 直方圖對齊主要平面法向量來估計初始旋轉，並依 Censi 的方法推導封閉形式共變異數。在 AASS 室內迴圈與 Hannover2 戶外資料上，D2D 的精度與 ICP、P2D 相當，在 AASS 上快將近一個數量級；模擬中直方圖初始化平均約 150 ms，FPFH 約 15 秒。結果多以箱形圖呈現且未報告硬體；在稀疏的戶外資料上各方法都有大量失敗，共變異數估計也偏大。
本頁內容
D2D-NDT converts both scans into 3D-NDT Gaussian grids and minimizes the L2 distance between the two mixtures over pairwise closest components with Newton's method and analytic derivatives on a 4, 2, 1, 0.5 m grid cascade; a 3D-NDT histogram supplies an initial rotation and a Censi-style closed-form covariance is derived; on AASS and Hannover2 it matches ICP and P2D accuracy while running up to about an order of magnitude faster.
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 | ["rotating SICK laser (AASS loop and Hannover2 data sets)", "simulated 3D range sensor in ROS/Gazebo with SICK LMS 200 error models, 180 x 120 deg field of view"] |
|---|---|
| 原文測試平台 | 未記錄 |
| 狀態估計 | minimizes an L2 distance between the fixed and moving 3D-NDT models, written as a sum of negative Gaussian terms over component pairs (Eq. 18); analytic gradient and Hessian; Newton's method with More-Thuente line search; parameters d1 = 1, d2 = 0.05; optimization repeated at grid sizes of 4, 2, 1 and 0.5 m; each pose increment is applied by transforming the whole moving NDT model and increments are accumulated in a homogeneous matrix so derivatives are always taken at the zero pose |
| 資料關聯 | distribution-to-distribution: each Gaussian component of the moving 3D-NDT is paired only with the closest component of the fixed 3D-NDT; no point-level correspondences; baselines ICP (PCL) and NDT-P2D run on point sets sub-sampled on a 0.1 m grid |
| 時間表示 | 不適用 |
| 去畸變 | 原文未報告 |
| 迴圈閉合 | none |
| 全域最佳化 | none; pairwise scan registration only; coupling with loop detection and a global pose-graph optimization is stated as future work |
| 地圖表示 | 3D-NDT: one Gaussian (mean and covariance) per occupied cell of a regular grid, built for both scans at cell sizes 4, 2, 1 and 0.5 m; about 1,500 components at 0.5 m for a typical AASS scan |
| 先驗資訊 | optional initial rotation from the 3D-NDT histogram: per range band 1 linear, 20 flat (orientation) and 5 spherical bins, three range bands, 78 values; the n = 3 dominant flat-patch directions of the two histograms are matched over 6 permutations and the rotation with the most similar histogram is kept; orientation only, no translation; in registration the top three candidate rotations and the zero pose each start a run and the best score is kept (up to four times the runtime) |
| 可輸出幾何 | 6-DoF rigid transformation (homogeneous matrix) plus a closed-form covariance derived after Censi (2007) with the NDT component covariances as the measurements; in the simulation test the estimate over-estimated the sample covariance, which the authors consider acceptable |
| 計算需求 | hardware not reported; NDT-histogram initialization averages on the order of 150 ms vs about 15 s for the ROS FPFH baseline (Sec. 6.1); D2D runs almost an order of magnitude faster than PCL ICP and the point-to-distribution variant on AASS (the sentence names 'D2D' for the second method, evidently a typo for P2D) and is consistently faster on Hannover as well (Figs. 5c, 6c, text only); objective evaluation costs O(q log q) for q Gaussian components vs O(n log n) for ICP, and D2D time is usually dominated by building the NDT models |
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 | rotating SICK laser | 資料集感測器 | AASS loop and Hannover2 (3D scans online repository) | real-world data sets acquired with SICK laser scanners on a rotating mount; AASS: 60 point clouds of about 90,000 points; Hannover: 923 point clouds of about 15,000 points (Table 1) | (Stoyanov et al., 2012, Sec. 6.1, 6.2, Table 1) |
| LiDAR | SICK LMS 200 | 方法輸入 | simulated Willow garage office world and asphalt-mill environment (ROS/Gazebo) | simulated only: the noise and mixed-measurement error models reported for this scanner (Ye and Borenstein 2002; Tuley et al. 2005) were implemented in ROS/Gazebo for a simulated 3D sensor with a 180 x 120 deg field of view | (Stoyanov et al., 2012, Sec. 6.1, 6.3) |
作者報告的優勢與限制
優勢
- D2D final alignments comparable to ICP and NDT-P2D on the AASS loop, 59 scan pairs (Sec. 6.2, Fig. 5)
- with histogram initialization D2D has the largest share of successfully registered AASS pairs, success being at most 0.5 m and 0.2 rad error (Sec. 6.2, Fig. 7)
- D2D consistently faster than ICP and P2D (Sec. 6.2)
- in the 343-offset robustness test on five AASS pairs (success within 0.2 m and 0.05 rad) D2D slightly outperforms P2D with stable results, while ICP fails on corridor pairs 1 and 5 (Sec. 6.2, Fig. 9)
- histogram initialization more accurate and much faster than FPFH (Sec. 6.1, Fig. 4)
- inexpensive closed-form covariance (Sec. 5, 7)
限制
- as an iterative method it remains susceptible to local minima (Sec. 7)
- without initialization all methods show large translational error variance on AASS, with failures in corridors and locally similar geometry (Sec. 6.2)
- on the sparser outdoor Hannover2 set all six combinations are worse, with many outliers and an apparent systematic bias, and histogram initialization does not help and slightly lowers P2D success (Sec. 6.2, Figs. 6, 8)
- the histogram initial guess corrects orientation only (Sec. 6.1)
- the covariance estimate over-estimates the observed covariance (Sec. 6.3)
- no comparison with Generalized ICP, 3DTK ICP or global registration methods (Sec. 6.2)
- evaluation on scan pairs, not full trajectories, with results only as box plots (Sec. 6.2)
- not tested on noisier or narrower field-of-view sensors (Sec. 7)
- NDT variants depend on voxel-size choice (Koide et al., 2021b Sec. II, secondary)
營建工程相關證據
原文未報告（測試資料為 AASS 室內迴圈、Hannover2 戶外掃描，以及模擬的辦公室與瀝青廠場景，論文未討論營建應用）
原文驗證環境：公開基準、模擬
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 4 個比較組，合計 27 筆紀錄。
Stückler & Behnke, 2014 · Table 1 本方法 22 筆
指標median relative pose error (RPE) in mm
表格設定（擷取紀錄原文）：Incremental (frame-to-frame) registration on TUM Freiburg sequences; median translational relative pose error in mm (maximum values in brackets in the table not extracted); warp is the OpenCV reimplementation (Stückler & Behnke, 2014, Table 1)
median relative pose error (RPE) in mm，TUM RGB-D (Freiburg) · fr1 360
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 Stückler & Behnke, 2014 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(Stückler & Behnke, 2014, Table 1)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Ours (MRSMap)原文提出 | 5.1 mm | (Stückler & Behnke, 2014, Table 1) |
| Warp [17] (OpenCV) | 5.9 mm | (Stückler & Behnke, 2014, Table 1) |
| GICP [5] | 18.8 mm | (Stückler & Behnke, 2014, Table 1) |
| 3D-NDT [7]本方法 | 7.8 mm | (Stückler & Behnke, 2014, Table 1) |
| Fovis [12] | 7.1 mm | (Stückler & Behnke, 2014, Table 1) |
Magnusson et al., 2015 · Fig. 3 (execution-time table) 本方法 2 筆
資料集與序列ETH Challenging Laser Registration (six data sets) · all scan pairs and pose offsets
表格設定（擷取紀錄原文）：execution time per registration over all six ETH data sets, including pre-processing and excluding file loading; single-threaded; different CPUs per method, so the authors call the comparison coarse (Magnusson et al., 2015, Fig. 3 (execution-time table))
execution time Q50，ETH Challenging Laser Registration (six data sets) · all scan pairs and pose offsets
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 Magnusson et al., 2015 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(Magnusson et al., 2015, Fig. 3 (execution-time table))
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Plane ICP (libpointmatcher point-to-plane baseline)硬體：i7 2.2 GHz | 2.58 s | (Magnusson et al., 2015, Fig. 3 (table)) |
| P2D-NDT (perception_oru)硬體：i7 3.5 GHz | 1.48 s | (Magnusson et al., 2015, Fig. 3 (table)) |
| D2D-NDT (perception_oru)本方法硬體：i7 3.5 GHz | 0.37 s | (Magnusson et al., 2015, Fig. 3 (table)) |
| MUMC DC-OFF硬體：i7 3.4 GHz | 2.95 s | (Magnusson et al., 2015, Fig. 3 (table)) |
| MUMC DC-ON硬體：i7 3.4 GHz | 2.41 s | (Magnusson et al., 2015, Fig. 3 (table)) |
Stückler & Behnke, 2014 · Table 2 本方法 2 筆
指標average runtime in milliseconds
表格設定（擷取紀錄原文）：Average runtime per incremental registration in ms (standard deviations in the table not extracted) (Stückler & Behnke, 2014, Table 2)
average runtime in milliseconds，TUM RGB-D (Freiburg) · fr1 desk
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 Stückler & Behnke, 2014 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(Stückler & Behnke, 2014, Table 2)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Ours (MRSMap)原文提出硬體：Intel Core i7 3610QM 2.3 GHz notebook CPU, VGA images | 75.15 ms | (Stückler & Behnke, 2014, Table 2) |
| Warp [17] (OpenCV)硬體：Intel Core i7 3610QM 2.3 GHz notebook CPU, VGA images | 108.64 ms | (Stückler & Behnke, 2014, Table 2) |
| GICP [5]硬體：Intel Core i7 3610QM 2.3 GHz notebook CPU, VGA images | 4015.4 ms | (Stückler & Behnke, 2014, Table 2) |
| 3D-NDT [7]本方法硬體：Intel Core i7 3610QM 2.3 GHz notebook CPU, VGA images | 414.87 ms | (Stückler & Behnke, 2014, Table 2) |
| Fovis [12]硬體：Intel Core i7 3610QM 2.3 GHz notebook CPU, VGA images | 15.98 ms | (Stückler & Behnke, 2014, Table 2) |
Stoyanov et al., 2012 · Text Sec. 6.1 本方法 1 筆
指標average runtimes on the order of 150 ms
資料集與序列simulated Willow and Terrain data sets (ROS/Gazebo) · 160 scan pairs per data set
表格設定（擷取紀錄原文）：Initial orientation estimation on simulated scan pairs (20 positions per environment, 10 scans 15 deg apart, 8 pairs 30 deg apart per position, 160 pairs per data set, about 80% overlap); average runtime stated in text for both data sets; FPFH baseline is the ROS implementation (Stoyanov et al., 2012, Text Sec. 6.1)
average runtimes on the order of 150 ms，simulated Willow and Terrain data sets (ROS/Gazebo) · 160 scan pairs per data set
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 Stoyanov et al., 2012 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 |
|---|---|---|
| 3D-NDT histogram initialization本方法原文提出 | 150 ms有附註註記（擷取紀錄）：approximate, stated as on the order of 150 ms | (Stoyanov et al., 2012, Sec. 6.1, Fig. 4(b)) |
來源
Stoyanov et al., 2012
（2012）Fast and accurate scan registration through minimization of the distance between compact 3D NDT representationsThe International Journal of Robotics Research, 31(12):1377-1393
同儕審查已出版已讀全文經典