SuMa
SuMa 以面元(surfel,帶法向量與半徑的小圓盤)地圖表示環境,將掃描投影成球面頂點圖與法向量圖,並從面元地圖繪製(render)同視角的模型圖,以投影式資料關聯(projective data association)執行密集的點到面 frame-to-model ICP,避免最近鄰搜尋。面元以穩定度對數勝算比過濾動態物體與雜訊;迴圈則在非活動地圖中搜尋候選,並以「合成虛擬視圖」檢驗一致性、連續多幀驗證後才加入位姿圖。由於面元綁定建立時的位姿,位姿圖最佳化後可直接更新地圖。
本頁內容
SuMa performs dense frame-to-model ICP by projective association against views rendered from a GPU surfel map, verifies loop closures with composed virtual views, and updates the pose-anchored surfel map after pose-graph optimization.
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-64E S2 via KITTI) |
|---|---|
| 原文測試平台 | vehicle (KITTI) |
| 狀態估計 | frame-to-model point-to-plane ICP with projective data association, Gauss-Newton with Huber weights; pose graph optimized with gtsam (Levenberg-Marquardt) in a separate thread (Sec. III-C, III-E, III-F) |
| 資料關聯 | dense projective data association between the current vertex/normal maps (spherical projection) and vertex/normal maps rendered from the surfel map (Sec. III-A, III-C) |
| 時間表示 | discrete scan poses |
| 去畸變 | 原文未報告 |
| 迴圈閉合 | single candidate within a radius searched in the inactive map; ICP with multiple initializations; accepted only if a composed virtual map view is consistent with the scan, then verified over subsequent scans (Sec. III-E) |
| 全域最佳化 | pose graph of odometry and verified loop closures (gtsam); surfels are anchored to creation poses so the map is updated without re-integration (Sec. III-B, III-E) |
| 地圖表示 | surfel map (position, normal, radius, creation/update timestamps, stability log-odds); active/inactive partition; GPU rolling-grid submaps (Sec. III-B, III-F) |
| 先驗資訊 | none |
| 可輸出幾何 | globally consistent surfel map / registered point cloud (Fig. 1, Fig. 4) |
| 計算需求 | OpenGL 4.0 GPU implementation with a 2D rolling-grid of submaps offloaded between GPU and main memory; test machine Intel i7-6700 @3.4 GHz with 16 GB RAM and Nvidia GeForce GTX 960 with 4 GB RAM; odometry and map update 31 ms average (max 71 ms), up to 189 ms with loop closure detection and verification, 48 ms average overall (about 20 Hz) on KITTI 00; pose graph optimized with gtsam 4.0 Levenberg-Marquardt in a separate thread (Sec. III-F, IV, Fig. 7) |
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 | 資料集感測器 | KITTI odometry | provided KITTI point clouds recorded at 10 Hz; parameters assume vertical FoV fup 3.0 deg and fdown 25 deg, vertex map 900 x 64 (Table I) | (Behley & Stachniss, 2018, Sec. IV, Table I) |
| GNSS 接收器 | KITTI GPS-based inertial navigation system ground truth (model not stated) | 參考或真值量測 | KITTI odometry | authors observed height inconsistencies in the training ground truth | (Behley & Stachniss, 2018, Sec. IV, Fig. 6) |
| 運算硬體 | Intel i7-6700 | 執行運算平台 | 未標示 | 3.4 GHz, 16 GB RAM | (Behley & Stachniss, 2018, Sec. IV) |
| 運算硬體 | Nvidia GeForce GTX 960 | 執行運算平台 | 未標示 | 4 GB RAM; OpenGL 4.0 implementation | (Behley & Stachniss, 2018, Sec. III-F, IV) |
作者報告的優勢與限制
優勢
- KITTI training set average 0.3 deg/100 m and 0.8% relative translational error with loop closure (Table II)
- Full pipeline including loop closure at about 20 Hz on average on KITTI 00 (Sec. IV, Fig. 7)
- Loop detection works with small overlap between scans via the virtual-view criterion (Sec. III-E, Fig. 5)
限制
- Test set: 1.4% translational error vs 0.7% reported for LOAM (Sec. IV)
- Fails to estimate motion in sequences with very few structures such as highways (Sec. IV)
- Consistently moving objects (e.g., cars in a jam) can be integrated as static surfels and corrupt the map (Sec. IV)
- Loop-closure gains are hard to assess with KITTI's GPS/INS ground truth, which has height inconsistencies (Sec. IV)
- Requires a GPU (OpenGL) implementation (Sec. III-F)
營建工程相關證據
原文未報告(僅在 KITTI 城市、鄉間與高速公路資料測試;(推論)工地常見的長廊與開闊樓板等結構稀少場景可能類似其報告的退化情境,需實測)。
原文驗證環境:公開基準
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 46 個比較組,合計 387 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 42 組列在最後,並連到性能比較頁。
Behley & Stachniss, 2018 · Table II 本方法 72 筆
表格設定(擷取紀錄原文):KITTI odometry training set; relative errors averaged over 100-800 m trajectories; values written rot [deg/100m] / trans [%]; * = sequence contains loop closures; LOAM, S-LSD and SOFT-SLAM values as reported by their authors (Behley & Stachniss, 2018, Table II)
relative translational error (%),KITTI odometry (training) · 00*
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 Behley & Stachniss, 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(Behley & Stachniss, 2018, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Frame-to-Frame本方法原文提出 | 2.1% | (Behley & Stachniss, 2018, Table II) |
| Frame-to-Model本方法原文提出 | 0.7% | (Behley & Stachniss, 2018, Table II) |
| Frame-to-Model with loop closure本方法原文提出 | 0.7% | (Behley & Stachniss, 2018, Table II) |
| LOAM [35] | 0.8% | (Behley & Stachniss, 2018, Table II) |
| S-LSD [6] (Stereo LSD-SLAM) | 0.6% | (Behley & Stachniss, 2018, Table II) |
| SOFT-SLAM [2] | 0.7% | (Behley & Stachniss, 2018, Table II) |
Chen et al., 2019 · Table I 本方法 26 筆
表格設定(擷取紀錄原文):KITTI raw road-category drives 2011_09_26_drive_0015_sync to 2011_10_03_drive_0047_sync renamed 30-41; relative errors averaged over segments of 5 to 400 m (rot deg/100 m, trans %); these drives have no semantic labels, so they are unseen by RangeNet++ (Chen et al., 2019, Table I)
relative rotational error,KITTI raw (road category) · 30
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 Chen et al., 2019 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(Chen et al., 2019, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| SuMa本方法 | 0.38 deg/100 m | (Chen et al., 2019, Table I) |
| SuMa_nomovable | 0.39 deg/100 m | (Chen et al., 2019, Table I) |
| SuMa++原文提出 | 0.38 deg/100 m | (Chen et al., 2019, Table I) |
Chebrolu et al., 2021 · Table I 本方法 24 筆
表格設定(擷取紀錄原文):Frame-to-frame point-to-plane projective ICP inside SuMa on KITTI odometry; only the robust kernel or outlier scheme differs between rows (Chebrolu et al., 2021, Table I)
relative rot. error in degrees per 100 m,KITTI odometry · 00
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 Chebrolu et al., 2021 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(Chebrolu et al., 2021, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Our Approach原文提出 | 1.5 deg/100 m | (Chebrolu et al., 2021, Table I) |
| Adaptive Kernel (Barron [6]) | 1.6 deg/100 m | (Chebrolu et al., 2021, Table I) |
| Fixed Kernel (Huber) | 0.93 deg/100 m | (Chebrolu et al., 2021, Table I) |
| Fixed Kernel (Geman-McClure) | 1.8 deg/100 m | (Chebrolu et al., 2021, Table I) |
| Hand-Crafted Outlier Rejection [7] (SuMa original: Huber + rejection of correspondences >2 m or normal angle >30 deg)本方法 | 0.9 deg/100 m | (Chebrolu et al., 2021, Table I) |
Chen et al., 2019 · Table II 本方法 24 筆
表格設定(擷取紀錄原文):KITTI odometry training sequences 00-10; relative errors averaged over 100 to 800 m segments (rot deg/100 m, trans %); asterisked sequences 00, 02, 05-09 contain loops; RangeNet++ was trained on labels of 00-10 except 08, so this table is not held out; IMLS-SLAM and LOAM rows give translation only and their source is not stated; SuMa_nomovable Average printed as 23.3/9.24, which looks swapped relative to its per-sequence values (reviewer arithmetic), recorded as printed (Chen et al., 2019, Table II)
relative translational error,KITTI odometry (training) · 00
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 Chen et al., 2019 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(Chen et al., 2019, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| SuMa本方法 | 0.68% | (Chen et al., 2019, Table II) |
| SuMa_nomovable | 58%有附註註記(擷取紀錄):(inference) diverged: authors state SuMa_nomovable diverges, particularly in urban scenes; per-sequence divergence not labelled in the paper | (Chen et al., 2019, Table II) |
| SuMa++原文提出 | 0.64% | (Chen et al., 2019, Table II) |
| IMLS-SLAM [7] | 0.5% | (Chen et al., 2019, Table II) |
| LOAM [40] | 0.78% | (Chen et al., 2019, Table II) |
其他比較組
列出其餘 42 個比較組
- Yokozuka et al., 2021 · Table III
- Ruan et al., 2023 · Table II
- Li et al., 2021a · Table II
- Vizzo et al., 2021 · Table II
- Yuan et al., 2022 · Table II
- Liu et al., 2023b · Table II
- Liu et al., 2023b · Table V
- Vizzo et al., 2023 · Table III
- Guadagnino et al., 2025a · Table III
- Koide et al., 2021a · Table I
- Pan et al., 2024 · Table VII
- Li et al., 2021a · Table III
- Guadagnino et al., 2025a · Table IV
- Guadagnino et al., 2025a · Table V
- Lim et al., 2024 · Table 6
- Nubert et al., 2021 · Table II
- Zhu et al., 2025 · Table III
- Lee et al., 2025a · Table II
- Guadagnino et al., 2025a · Table II
- Deng et al., 2023 · Table 3
- Deng et al., 2023 · Table 5
- Chen et al., 2020 · Table II
- Pan et al., 2025 · Table III
- Behley & Stachniss, 2018 · Text Sec. IV Runtime
- Vizzo et al., 2021 · Table I
- Yokozuka et al., 2021 · Table III (overall)
- Yokozuka et al., 2021 · Table IV
- Li et al., 2021a · Table IV
- Zhang et al., 2024a · Table 8
- Zhang et al., 2024a · Table 9
- Vizzo et al., 2023 · Table II
- Pan et al., 2024 · Table IV
- Behley & Stachniss, 2018 · Text Sec. IV
- Chen et al., 2019 · Text Sec.IV-B
- Yuan et al., 2022 · Table III
- Zhu et al., 2025 · Table I
- Zhu et al., 2025 · Table IV
- Lee et al., 2025a · Table III
- Ghadimzadeh Alamdari et al., 2025 · Table 3
- Guadagnino et al., 2025a · Table I
- Yokozuka et al., 2021 · Table V
- Zhu et al., 2025 · Table II
來源
Behley & Stachniss, 2018
(2018)Efficient Surfel-Based SLAM using 3D Laser Range Data in Urban EnvironmentsRobotics: Science and Systems XIV (RSS 2018), RSS XIV, paper p16 (proceedings URL rss14/p16.pdf)
DOI 10.15607/rss.2018.xiv.016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:jbehley/SuMa https://github.com/jbehley/SuMa
程式碼:https://github.com/jbehley/SuMa(授權:MIT-style permission notice (LICENSE file, Copyright 2016-2018 Jens Behley))。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。