VGICP
VGICP 延伸 GICP，以體素化取代耗時的最近鄰搜尋：每個體素彙整其內各點的分布（而非像 NDT 直接由點位置計算分布），形成分布對多分布的對應，即使體素內點數少也能得到有效分布。體素化使最佳化容易平行化，作者報告 CPU 約 30 Hz、GPU 約 120 Hz，精度與 GICP 相當且對體素解析度較不敏感。
本頁內容
VGICP voxelizes GICP by aggregating per-point covariances per voxel, avoiding nearest-neighbour search and enabling parallel CPU/GPU registration with GICP-level accuracy.
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 (Velodyne HDL-32E real and simulated) |
|---|---|
| 原文測試平台 | simulation (authors' ray-casting LiDAR simulator using Velodyne HDL-32e parameters)、real HDL-32e recordings, eight sequences of about 120 m; carrying platform not stated |
| 狀態估計 | GICP-style least squares with voxel-based distribution-to-multi-distribution residuals; implementation uses Gauss-Newton-type optimizer (Sec. III, IV-B) |
| 資料關聯 | per-point covariances from k nearest neighbours (for example k = 20) found with a KD-tree and regularised to eigenvalues (1, 1, epsilon); the target is voxelized by averaging point means and covariances per voxel; each source point is matched to the voxel it falls in (distribution-to-multi-distribution), so no nearest-neighbour search is needed during optimization |
| 時間表示 | discrete poses (scan-to-scan odometry in experiments) |
| 去畸變 | 原文未報告 |
| 迴圈閉合 | none |
| 全域最佳化 | none |
| 地圖表示 | voxel map of aggregated point distributions (Gaussian voxelmap) |
| 先驗資訊 | initial guess required (Algorithm 1) |
| 可輸出幾何 | 6-DoF rigid transformation |
| 計算需求 | single-threaded GICP 189 ms and VGICP 156 ms versus PCL GICP 201 ms; multi-threaded GICP 68 ms and VGICP 50 ms; GPU VGICP 6 ms to optimize (simulated data, Sec. IV-A); on real data about 30 fps on CPU and 120 fps on GPU for about 15,000-point frames (Table II); Intel Core i9-9900K and NVIDIA GeForce RTX2080Ti (Sec. IV-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 HDL-32e歸入：Velodyne HDL-32E | 方法輸入 | authors' real HDL-32e sequences | about 15,000 points per frame; eight sequences of about 120 m | (Koide et al., 2021b, Sec. IV-B) |
| 運算硬體 | Intel Core i9-9900K | 執行運算平台 | 未標示 | CPU used for all methods | (Koide et al., 2021b, Sec. IV-A) |
| 運算硬體 | NVIDIA Geforce RTX2080Ti | 執行運算平台 | 未標示 | GPU for VGICP GPU version | (Koide et al., 2021b, Sec. IV-A) |
作者報告的優勢與限制
優勢
- accuracy comparable to GICP but substantially faster (abstract)
- robust to voxel resolution changes, unlike NDT (Sec. IV-B, Sec. V)
限制
- voxelization may affect convergence when the initial guess is far from the true pose (Sec. V, future work)
- real-sequence reference built by aligning last frame to first with GICP, so reference is not independent (Sec. IV-B; independence concern is reviewer inference)
- GPU brute-force nearest-neighbour search was slower than the CPU parallel KD-tree in this evaluation (Sec. IV-A)
- (reviewer observation) evaluation is consecutive-frame odometry only; no loop closure, map-quality or engineering-accuracy evaluation (Sec. IV)
營建工程相關證據
原文未報告（實驗為模擬與約 120 m 的 HDL-32E 序列，場域類型未明示）
原文驗證環境：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 6 個比較組，合計 24 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 2 組列在最後，並連到性能比較頁。
Koide et al., 2021b · Table II 本方法 8 筆
資料集與序列authors' HDL-32e sequences · 8 sequences (aggregate)
表格設定（擷取紀錄原文）：eight real Velodyne HDL-32e sequences of about 120 m, about 15,000 points per frame; consecutive-frame registration; error of the last frame against a reference obtained by aligning the last frame to the first with GICP; the ± term is not defined in the paper (presumably spread over the eight sequences) (Koide et al., 2021b, Table II)
FPS (CPU)，authors' HDL-32e sequences · 8 sequences (aggregate)
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., 2021b 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., 2021b, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| VGICP (0.5m)本方法原文提出硬體：not restated in Sec. IV-B; Sec. IV-A states all methods were run on an Intel Core i9-9900K and NVIDIA GeForce RTX2080Ti | 28.7 fps | (Koide et al., 2021b, Table II) |
| VGICP (1.0m)本方法原文提出硬體：not restated in Sec. IV-B; Sec. IV-A states all methods were run on an Intel Core i9-9900K and NVIDIA GeForce RTX2080Ti | 30.4 fps | (Koide et al., 2021b, Table II) |
| GICP (ours)硬體：not restated in Sec. IV-B; Sec. IV-A states all methods were run on an Intel Core i9-9900K and NVIDIA GeForce RTX2080Ti | 20.5 fps | (Koide et al., 2021b, Table II) |
| GICP (PCL)硬體：not restated in Sec. IV-B; Sec. IV-A states all methods were run on an Intel Core i9-9900K and NVIDIA GeForce RTX2080Ti | 5.2 fps | (Koide et al., 2021b, Table II) |
| NDT (0.5m)硬體：not restated in Sec. IV-B; Sec. IV-A states all methods were run on an Intel Core i9-9900K and NVIDIA GeForce RTX2080Ti | 8.3 fps | (Koide et al., 2021b, Table II) |
| NDT (1.0m)硬體：not restated in Sec. IV-B; Sec. IV-A states all methods were run on an Intel Core i9-9900K and NVIDIA GeForce RTX2080Ti | 10.3 fps | (Koide et al., 2021b, Table II) |
| NDT (2.0m)硬體：not restated in Sec. IV-B; Sec. IV-A states all methods were run on an Intel Core i9-9900K and NVIDIA GeForce RTX2080Ti | 10.4 fps | (Koide et al., 2021b, Table II) |
| NDT (4.0m)硬體：not restated in Sec. IV-B; Sec. IV-A states all methods were run on an Intel Core i9-9900K and NVIDIA GeForce RTX2080Ti | 9.3 fps | (Koide et al., 2021b, Table II) |
| ICP硬體：not restated in Sec. IV-B; Sec. IV-A states all methods were run on an Intel Core i9-9900K and NVIDIA GeForce RTX2080Ti | 11 fps | (Koide et al., 2021b, Table II) |
Lim et al., 2024 · Table 6 本方法 6 筆
表格設定（擷取紀錄原文）：KITTI Seq. 00 odometry test with frame interval Delta (source i+Delta, target i); trel [%] and rrel [deg/100m] by RPG evaluation tools; c2f = global registration then local registration (G-ICP); deep-learning rows copied by the authors from the original papers; † = Seq. 00 used for training (Lim et al., 2024, Table 6)
trel，KITTI · Seq. 00, Delta = 1
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 Lim et al., 2024 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(Lim et al., 2024, Table 6)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| ICP | 6.88% | (Lim et al., 2024, Table 6) |
| G-ICP | 1.26% | (Lim et al., 2024, Table 6) |
| VGICP本方法 | 1.03% | (Lim et al., 2024, Table 6) |
| FGR | 2.73% | (Lim et al., 2024, Table 6) |
| TEASER++ | 2.11% | (Lim et al., 2024, Table 6) |
| Quatro (Ours)原文提出 | 1.45% | (Lim et al., 2024, Table 6) |
| Quatro++ (Ours)原文提出 | 1.9% | (Lim et al., 2024, Table 6) |
| LO-Net | 1.47% | (Lim et al., 2024, Table 6) |
| LO-Net+M | 0.78% | (Lim et al., 2024, Table 6) |
| DMLO† | 0.83% | (Lim et al., 2024, Table 6) |
| DMLO+M† | 0.73% | (Lim et al., 2024, Table 6) |
| A-LOAM + StickyPillars† | 0.65% | (Lim et al., 2024, Table 6) |
| SuMa | 0.68% | (Lim et al., 2024, Table 6) |
| A-LOAM | 0.7% | (Lim et al., 2024, Table 6) |
| Quatro-c2f (Ours)原文提出 | 0.65% | (Lim et al., 2024, Table 6) |
| Quatro++-c2f (Ours)原文提出 | 0.68% | (Lim et al., 2024, Table 6) |
Koide, 2024 · Text BENCHMARK.md Accuracy 本方法 4 筆
資料集與序列KITTI odometry sequence 00 · 00
表格設定（擷取紀錄原文）：repository documentation linked from the paper (BENCHMARK.md, master branch, fetched 2026-09-25), not peer-reviewed text; odometry benchmark on KITTI 00; units and the meaning of '+-' and of the RPE window (100, 400, 800) are not stated; value is the number before '+-' (Koide, 2024, Text BENCHMARK.md Accuracy)
APE = 6.791 +- 3.215，KITTI odometry sequence 00 · 00
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, 2024 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 |
|---|---|---|
| fast_vgicp本方法 | 6.791 | (Koide, 2024, BENCHMARK.md, Accuracy) |
Koide et al., 2021b · Text Sec. IV-A 本方法 3 筆
資料集與序列authors' simulated LiDAR sequence · simulated sequence
表格設定（擷取紀錄原文）：average processing time per scan on the simulated sequence, as stated in the Sec. IV-A text (Fig. 4 is a plot and was not read off); the text does not say explicitly whether each stated time includes the covariance preprocessing; the GPU figure is described as the time to optimize (Koide et al., 2021b, Text Sec. IV-A)
processing time，authors' simulated LiDAR sequence · simulated sequence
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., 2021b 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., 2021b, Text Sec. IV-A)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| GICP (ours, single-thread)硬體：Intel Core i9-9900K CPU; NVIDIA GeForce RTX2080Ti GPU | 189 ms | (Koide et al., 2021b, Sec. IV-A) |
| VGICP (single-thread)本方法原文提出硬體：Intel Core i9-9900K CPU; NVIDIA GeForce RTX2080Ti GPU | 156 ms | (Koide et al., 2021b, Sec. IV-A) |
| GICP (PCL)硬體：Intel Core i9-9900K CPU; NVIDIA GeForce RTX2080Ti GPU | 201 ms | (Koide et al., 2021b, Sec. IV-A) |
| GICP (ours, multi-thread)硬體：Intel Core i9-9900K CPU; NVIDIA GeForce RTX2080Ti GPU | 68 ms | (Koide et al., 2021b, Sec. IV-A) |
| VGICP (multi-thread)本方法原文提出硬體：Intel Core i9-9900K CPU; NVIDIA GeForce RTX2080Ti GPU | 50 ms | (Koide et al., 2021b, Sec. IV-A) |
其他比較組
來源
Koide et al., 2021b
（2021）Voxelized GICP for Fast and Accurate 3D Point Cloud Registration2021 IEEE International Conference on Robotics and Automation (ICRA), pp. 11054-11059
DOI 10.1109/icra48506.2021.9560835程式碼
同儕審查已出版已讀全文近十年
相關版本
- 預印本：EasyChair Preprint 2703 (dated 2020-02-18) https://easychair.org/publications/preprint/ftvV
- 程式碼釋出：fast_gicp https://github.com/koide3/fast_gicp
程式碼：https://github.com/koide3/fast_gicp（授權：BSD-3-Clause (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。