LeGO-LOAM
LeGO-LOAM 針對地面載具,先把點雲投影為距離影像(range image),分離地面點,並以影像式分割剔除少於 30 點的小群集(如樹葉);邊緣特徵只取自非地面點,以避開草地造成的不穩定特徵,再依 LOAM 的粗糙度指標擷取邊緣與平面特徵。其兩步驟 Levenberg-Marquardt 最佳化先以地面平面特徵求 [tz, roll, pitch],再以邊緣特徵求 [tx, ty, yaw],以降低計算量。地圖改為儲存每次掃描的特徵集合與對應位姿,並可選擇性地接上以 ICP 建立迴圈約束、iSAM2 最佳化的位姿圖(pose graph)。
本頁內容
LeGO-LOAM adds ground segmentation and cluster filtering to LOAM, solves the pose in two ground-then-edge LM steps for embedded real-time use, and optionally closes loops with ICP constraints in an iSAM2 pose graph.
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 VLP-16; HDL-64E via KITTI)、IMU (low-cost CH Robotics UM6, used only for initial guess) |
|---|---|
| 原文測試平台 | wheeled UGV (Clearpath Jackal)、vehicle (KITTI) |
| 狀態估計 | two-step Levenberg-Marquardt: ground planar features estimate [tz, roll, pitch], then edge features estimate [tx, ty, yaw]; optional pose graph optimized with iSAM2 (Sec. III-D, III-E, IV-D) |
| 資料關聯 | range-image ground separation and image-based segmentation (clusters < 30 points discarded); LOAM-style roughness-based edge/planar features; label-consistent point-to-edge / point-to-plane matching (Sec. III-B to III-D) |
| 時間表示 | discrete scan poses (per-scan transformation) (Sec. III) |
| 去畸變 | 原文未報告 in the paper (feature and matching details deferred to LOAM [20]); IMU provides the initial guess (Sec. IV-A) |
| 迴圈閉合 | optional: ICP between current and earlier feature sets adds pose-graph constraints, optimized by iSAM2; used only in the KITTI seq. 00 test (Sec. III-E, IV-D) |
| 全域最佳化 | optional iSAM2 pose graph (Sec. III-E, IV-D) |
| 地圖表示 | per-scan edge/planar feature sets stored with sensor poses; local map assembled from sets within 100 m or the k most recent sets (Sec. III-E) |
| 先驗資訊 | none (assumes presence of ground for ground-optimized steps) |
| 可輸出幾何 | feature-set point cloud map and 6-DoF poses (Sec. III-E); export of full-resolution map not described in paper |
| 計算需求 | CPU only on an Nvidia Jetson TX2 (ARM Cortex-A57) and a 2.5 GHz i7-4710MQ laptop; LeGO-LOAM per scan on the Jetson: segmentation 29.3 to 36.8 ms, feature extraction 6.1 to 9.9 ms, odometry 18.1 to 19.3 ms, mapping 253.3 to 278.2 ms; on the i7: 16.7 to 20.0, 2.3 to 4.4, 6.1 to 6.8 and 101.7 to 116.7 ms; LOAM extraction plus odometry exceeded 100 ms on the Jetson so scans were skipped |
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 | 未標示 | 16 channels; range up to 100 m, accuracy +/-3 cm; vertical FOV 30 deg (+/-15 deg), 2 deg vertical resolution; 360 deg horizontal FOV, 0.1 to 0.4 deg horizontal resolution; scan rate set to 10 Hz (0.2 deg); projected to a 1800 x 16 range image | (Shan & Englot, 2018, Sec. II; Sec. III-B) |
| LiDAR | Velodyne HDL-64E | 資料集感測器 | KITTI odometry (sequence 00) | 360 deg horizontal FOV, 48 more channels than the VLP-16, vertical FOV 26.9 deg; downsampled to the VLP-16 range image (75% of points omitted) for real time on the Jetson | (Shan & Englot, 2018, Sec. II; Sec. IV-D) |
| 慣性量測單元(IMU) | CH Robotics UM6 Orientation Sensor | 方法輸入 | 未標示 | low-cost IMU on the Jackal; supplies the identical initial translational and rotational guess to LOAM and LeGO-LOAM | (Shan & Englot, 2018, Sec. II; Sec. IV-A) |
| 載具平台 | Clearpath Jackal | 方法輸入 | 未標示 | UGV, 270 Wh lithium battery, maximum speed 2.0 m/s, maximum payload 20 kg | (Shan & Englot, 2018, Sec. II; Fig. 1a) |
| 運算硬體 | Nvidia Jetson TX2歸入:NVIDIA Jetson TX2 | 執行運算平台 | 未標示 | embedded device with ARM Cortex-A57 CPU; CPU only | (Shan & Englot, 2018, Sec. II) |
| 運算硬體 | laptop with Intel i7-4710MQ | 執行運算平台 | 未標示 | 2.5 GHz i7-4710MQ CPU, chosen to match the LOAM papers' hardware; CPU only | (Shan & Englot, 2018, Sec. II) |
作者報告的優勢與限制
優勢
- Feature extraction and odometry runtime reduced by about an order of magnitude and mapping runtime by at least 60% relative to LOAM on the same hardware (Sec. IV-C, Table IV)
- Two-step LM reduces odometry runtime by 34-48% with similar accuracy (Sec. IV-C, Table III)
- On a forested trail (about 35 min, 19 m elevation change) end-to-start translation error 13.93 m (Jetson) vs 69.40 m for LOAM (Sec. IV-B3, Table V)
限制
- Ground-optimized steps assume a ground plane is visible; UAV use would require segmentation without ground extraction (Sec. V)
- Accuracy evaluated only by end-to-start pose difference in campus/forest runs, not by an external trajectory or map reference (Sec. IV-B)
- In KITTI loop-closure test, HDL-64E scans were reduced to a 16-ring range image (75% of points omitted) to run in real time on the Jetson (Sec. IV-D)
- Follow-up work states its IMU use is the same loosely coupled scheme as LOAM (Shan et al., 2020, Sec. II)
營建工程相關證據
原論文僅在校園與森林步道等室外地形測試。Feng 等人(Feng et al., 2025)在施工中醫院門診大樓與模擬工地以預設參數評估 LeGO-LOAM;作者在結論中報告其因地面點特徵與迴圈閉合而在 LiDAR-only 方法中表現最佳,實際工地 APE RMSE 為 7.97 m(Table 4);但在模擬工地中 HDL-Graph-SLAM 的 RMSE(12.12 m)略低於 LeGO-LOAM(12.33 m)(Table 3)。論文未說明實際工地 APE 所用參考軌跡的來源(全文僅描述 Gazebo 模擬的真實軌跡外掛),故實際工地 APE 只能視為作者報告值,不能當作已驗證的幾何精度。
原文驗證環境:公開基準、跨場域
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 43 個比較組,合計 300 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 39 組列在最後,並連到性能比較頁。
Shan & Englot, 2018 · Table IV 本方法 24 筆
表格設定(擷取紀錄原文):Runtime of each module for processing one scan, averaged over 10 real-time trials; LOAM has no segmentation module (Shan & Englot, 2018, Table IV)
runtime of segmentation module per scan,Own Jackal UGV datasets · Experiment 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。
- 不適用
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 Shan & Englot, 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(Shan & Englot, 2018, Table IV)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| LOAM(Nvidia Jetson TX2 (ARM Cortex-A57), CPU only) | 無數值不適用註記(擷取紀錄):不適用 (N/A in table) | (Shan & Englot, 2018, Table IV) |
| LOAM(laptop, 2.5 GHz Intel i7-4710MQ, CPU only) | 無數值不適用註記(擷取紀錄):不適用 (N/A in table) | (Shan & Englot, 2018, Table IV) |
| LeGO-LOAM(Nvidia Jetson TX2 (ARM Cortex-A57), CPU only)本方法原文提出 | 29.3 ms | (Shan & Englot, 2018, Table IV) |
| LeGO-LOAM(laptop, 2.5 GHz Intel i7-4710MQ, CPU only)本方法原文提出 | 16.7 ms | (Shan & Englot, 2018, Table IV) |
Yarovoi & Cho, 2024 · Table 2 本方法 24 筆
表格設定(擷取紀錄原文):Hilti 2022 handheld construction-site sequences; errors vs motion-capture GT; translation = Euclidean distance, rotation = smallest angle; RMSE and STD; mostly default parameters; * = loss of tracking (Yarovoi & Cho, 2024, Table 2)
Translation RMSE (m),Hilti SLAM Challenge Dataset 2022 · Exp04
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。
- 失敗
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 Yarovoi & Cho, 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(Yarovoi & Cho, 2024, Table 2)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Lego-LOAM (IMU)本方法 | 3.713 m失敗註記(擷取紀錄):failed (loss of tracking; value as reported) | (Yarovoi & Cho, 2024, Table 2) |
| Lego-LOAM (no IMU)本方法 | 2.55 m | (Yarovoi & Cho, 2024, Table 2) |
| LIO-SAM | 0.167 m | (Yarovoi & Cho, 2024, Table 2) |
| ART-SLAM odom | 2.755 m | (Yarovoi & Cho, 2024, Table 2) |
| ART-SLAM final | 1.032 m | (Yarovoi & Cho, 2024, Table 2) |
Chen et al., 2022b · Table V 本方法 20 筆
表格設定(擷取紀錄原文):Runtime of modules for processing one scan (ms) on KITTI 04, 06, 07, 09 and backpack K1 (Chen et al., 2022b, Table V)
Segmentation time per scan (ms),KITTI odometry · #04
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 Chen et al., 2022b 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 |
|---|---|---|
| LeGO-LOAM Segmentation本方法硬體:laptop Intel i7-7700HQ 2.8 GHz, 8 GB RAM | 21.5 ms | (Chen et al., 2022b, Table V) |
Liu et al., 2026 · Table 2 (full SLAM with LC) 本方法 13 筆
指標absolute trajectory error (RMSE, centimeters)
表格設定(擷取紀錄原文):Hilti handheld sequences; ATE exported from the Hilti evaluation website; full SLAM with loop closure (Our (Full) adds global mapping) (Liu et al., 2026, Table 2 (full SLAM with LC))
absolute trajectory error (RMSE, centimeters),Hilti handheld sequence exp01-construction (name per Table C1) · hilti01
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 Liu et al., 2026 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(Liu et al., 2026, Table 2 (full SLAM with LC))
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| LeGO-LOAM本方法 | 8.8 cm | (Liu et al., 2026, Table 2) |
| LiLi-OM | 6.2 cm | (Liu et al., 2026, Table 2) |
| LIO-SAM | 6.1 cm | (Liu et al., 2026, Table 2) |
| LTA-OM | 1.27 cm | (Liu et al., 2026, Table 2) |
| Our (Odom+LM+LC)原文提出 | 0.78 cm | (Liu et al., 2026, Table 2) |
| Our (Full)原文提出 | 0.62 cm | (Liu et al., 2026, Table 2) |
其他比較組
列出其餘 39 個比較組
- Liu et al., 2026 · Table 2 (odometry without LC)
- Yuan et al., 2022 · Table II
- Shan & Englot, 2018 · Table V
- Lim et al., 2024 · Table 7
- Yokozuka et al., 2021 · Table III
- Feng et al., 2026 · Table 2
- Feng et al., 2026 · Table 6
- Qin et al., 2020 · Table I
- Chen et al., 2022b · Table III
- Koide et al., 2021a · Table I
- Kim et al., 2022b · Table IV
- Bavle et al., 2023 · Table II
- Wu et al., 2024b · Table I
- Yan et al., 2026a · Table 1
- Li et al., 2021b · Table 1
- Yarovoi & Cho, 2024 · Table 3
- Feng et al., 2025 · Table 3
- Feng et al., 2025 · Table 4
- Feng et al., 2026 · Table 1
- Feng et al., 2026 · Table 4
- Feng et al., 2026 · Table 5
- Bavle et al., 2023 · Table I
- Feng et al., 2026 · Table 8
- Isaacson et al., 2023 · Table II
- Zhou et al., 2021 · Table I
- Frosi & Matteucci, 2022 · Table II
- Frosi & Matteucci, 2022 · Table III
- Frosi & Matteucci, 2022 · Table IV
- Lee et al., 2024b · Table 4
- Li et al., 2021b · Table 2
- Liu & Zhang, 2021 · Table II
- Chen et al., 2022b · Table IV
- Yuan et al., 2022 · Table III
- Ghadimzadeh Alamdari et al., 2025 · Table 3
- Yokozuka et al., 2021 · Table III (overall)
- Yokozuka et al., 2021 · Table IV
- Yokozuka et al., 2021 · Table V
- Zhang et al., 2024a · Table 8
- Zhang et al., 2024a · Table 9
來源
Shan & Englot, 2018
(2018)LeGO-LOAM: Lightweight and Ground-Optimized Lidar Odometry and Mapping on Variable Terrain2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 4758-4765
DOI 10.1109/iros.2018.8594299程式碼
同儕審查已出版已讀全文近十年查證後修正
相關版本
- 預印本:Author preprint PDF hosted in the official repository (IEEE copyright notice on page 1) https://raw.githubusercontent.com/RobustFieldAutonomyLab/LeGO-LOAM/master/Shan_Englot_IROS_2018_Preprint.pdf
- 程式碼釋出:RobustFieldAutonomyLab/LeGO-LOAM https://github.com/RobustFieldAutonomyLab/LeGO-LOAM
程式碼:https://github.com/RobustFieldAutonomyLab/LeGO-LOAM(授權:BSD 3-Clause (LICENSE file))。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。