NDT-LOAM
NDT-LOAM 把 LOAM 的特徵式前端改成加權的常態分布轉換(NDT)直接配準:每個 NDT 格依量測距離與格內形狀(平面、線狀或立體)給不同權重,並以目前幀對最近關鍵影格配準(Scan2Key)降低逐幀累積誤差。得到的初始位姿再交給沿用 LOAM 建圖模組的局部特徵調整(LFA),以角點與平面點對局部地圖精修。系統只處理前端,沒有迴圈閉合;在 KITTI 上平均平移漂移為 0.899%。
本頁內容
LOAM variant whose odometry front end is a range- and planarity-weighted NDT registered scan-to-keyframe, followed by LOAM-style local feature adjustment against a local map; front end only, 0.899% average KITTI drift at about 10 Hz.
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 only (Velodyne HDL-64E in KITTI; horizontal Velodyne VLP-16 on the Kylin backpack) |
|---|---|
| 原文測試平台 | vehicle (KITTI)、backpack (Kylin) |
| 狀態估計 | two stages: direct odometry by weighted NDT solved with Newton's method and line search against a keyframe (Scan2Key), then local feature adjustment (LFA) that reuses the LOAM mapping module to refine the pose with corner and surface feature residuals against a local map, two iterations (Sec. III) |
| 資料關聯 | NDT cells (1 m grid in KITTI tests) weighted by point range and by cell dimensionality from covariance eigenvalues (planar 1.25, volumetric 1.0, linear 0.75); LFA corner and surface correspondences searched in a KD-tree built from map cubes intersecting the scan (Sec. III-B; Sec. III-D; Sec. IV-A) |
| 時間表示 | discrete poses |
| 去畸變 | not described |
| 迴圈閉合 | none (loop closure and graph optimization are named as future work) (Sec. III-C; Sec. V) |
| 全域最佳化 | none |
| 地圖表示 | LOAM-style corner and surface feature map stored in cubic areas for LFA; keyframe scans (selected by 10 m, 10 deg or 1 s on KITTI; 2 m or 10 deg on the backpack) as NDT targets (Sec. III-C; Sec. III-D; Sec. IV-A; Sec. IV-D) |
| 先驗資訊 | none |
| 可輸出幾何 | trajectory and 3D point cloud map (Figs. 5 and 7) |
| 計算需求 | laptop Intel i7-7700HQ at 2.8 GHz with 8 GB RAM, ROS on Ubuntu 16.04; reported as real time at 10 Hz; per-scan module times on KITTI of 38.0-48.2 ms (NDT odometry), 37.1-42.9 ms (LFA extraction) and 86.9-107.8 ms (LFA mapping) (abstract; Sec. IV; Table V) |
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 | 資料集感測器 | KITTI odometry | KITTI recording platform; only the Velodyne data are used | (Chen et al., 2022b, Sec. IV; Fig. 2) |
| LiDAR | Velodyne VLP-16 | 方法輸入 | Kylin backpack sequences K1 and K2 | two lidars on the backpack; only the horizontal one used; 10 Hz | (Chen et al., 2022b, Sec. IV; Sec. IV-D) |
| 慣性量測單元(IMU) | Xsens-300 | 資料集感測器 | Kylin backpack sequences K1 and K2 | on the backpack; use by the method not stated | (Chen et al., 2022b, Sec. IV) |
| GNSS 接收器 | OXTS RT 3003 | 參考或真值量測 | KITTI odometry | high-accuracy GPS/INS localization system providing ground truth | (Chen et al., 2022b, Sec. IV; Fig. 2) |
| 雙目相機 | four stereo camera systems (grayscale and color) | 資料集感測器 | KITTI odometry | KITTI platform cameras; not used by the method | (Chen et al., 2022b, Sec. IV; Fig. 2) |
| 雙目相機 | Mynak D1000-IR-120 color binocular camera | 資料集感測器 | Kylin backpack sequences K1 and K2 | left camera images at 640 x 480 used in the experiment (their role is not stated) | (Chen et al., 2022b, Sec. IV) |
| 載具平台 | Kylin backpack | 方法輸入 | Kylin backpack sequences K1 and K2 | walked at about 1 m/s; no GPS device | (Chen et al., 2022b, Sec. IV; Fig. 2) |
| 運算硬體 | Intel i7-7700HQ | 執行運算平台 | 未標示 | laptop CPU at 2.8 GHz, 8 GB RAM, ROS on Ubuntu 16.04 | (Chen et al., 2022b, Sec. IV) |
作者報告的優勢與限制
優勢
- Average KITTI 00-10 translational error 0.899% after LFA, versus 1.809% for A-LOAM and 1.27% for F-LOAM, and at the level of the LOAM paper values (Table II)
- Weighted NDT lowers odometry error by about 12% relative to classic NDT (1.041% to 0.910% with Scan2Key) and Scan2Key roughly halves the error of Scan2Scan (Sec. IV-A)
- Odometry-only initial pose error 0.910% versus 4.913% for the A-LOAM odometry (Table I)
- Kylin backpack indoor and indoor-outdoor runs: start-to-end error 0.20 m and 0.25 m versus 7.99 m and 2.60 m for LeGO-LOAM (Table IV)
- APE RMSE on KITTI 00, 05 and 09 of 3.98, 2.39 and 1.58 m versus 5.98, 2.97 and 2.30 m for LeGO-LOAM with loop closure; the authors describe this as almost 50% better (Table III; Sec. IV-C)
限制
- No loop closure or back-end optimization (Sec. V)
- Accuracy is poorer on KITTI 01, 02, 08 and 10 (above 1.0%), attributed to highway speed and few features, calibration of the sensors, ground-truth error at the start of 08 and vibration on rough roads (Sec. IV-C)
- NDT cell weights (1.25, 1.0, 0.75) were set empirically (Sec. III-B)
- Backpack evaluation uses only the start-to-end offset because no GPS ground truth was available (Sec. IV-D)
- LOAM results are copied from its original paper rather than rerun (Sec. IV-C)
營建工程相關證據
論文未在施工現場測試;以 Kylin 背包在室內及室內外混合路線(含上下樓梯)測試,屬既有建築環境,但只以起訖點偏移估計誤差。依距離與平面性為 NDT 格加權的做法,可能有利於樓板與牆面等平面為主的建築場景,這屬推論。
原文驗證環境:公開基準、已完工建築
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 54 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。
Chen et al., 2022b · Table V 本方法 15 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)
DLO 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 |
|---|---|---|
| NDT-LOAM DLO本方法原文提出硬體:laptop Intel i7-7700HQ 2.8 GHz, 8 GB RAM | 48.2 ms | (Chen et al., 2022b, Table V) |
Chen et al., 2022b · Table I 本方法 12 筆
指標position error (%)
表格設定(擷取紀錄原文):KITTI 00-10, odometry part only (initial pose): A-LOAM feature odometry versus NDT-LOAM weighted NDT with Scan2Key; position error (%) from the KITTI development kit (Chen et al., 2022b, Table I)
position error (%),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 Chen et al., 2022b 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., 2022b, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| ALOAM (odometry only) | 4.12% | (Chen et al., 2022b, Table I) |
| NDT-LOAM (odometry only, wNDT + Scan2Key)本方法原文提出 | 0.76% | (Chen et al., 2022b, Table I) |
Chen et al., 2022b · Table II 本方法 12 筆
指標position error (%)
表格設定(擷取紀錄原文):KITTI 00-10 after refinement (LFA); position error (%); A-LOAM and NDT-LOAM run at 10 Hz; LOAM values from the original LOAM paper (1 Hz); F-LOAM values from [36] (Chen et al., 2022b, Table II)
position error (%),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 Chen et al., 2022b 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., 2022b, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| ALOAM | 0.88% | (Chen et al., 2022b, Table II) |
| FLOAM [36] | 0.92% | (Chen et al., 2022b, Table II) |
| LOAM (from original paper) | 0.78% | (Chen et al., 2022b, Table II) |
| NDT-LOAM本方法原文提出 | 0.79% | (Chen et al., 2022b, Table II) |
Chen et al., 2022b · Table III 本方法 9 筆
表格設定(擷取紀錄原文):KITTI 00, 05, 09; absolute pose error (m) computed with evo; LeGO-LOAM with loop closure; alignment not stated (Chen et al., 2022b, Table III)
Absolute Pose Error, mean (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 Chen et al., 2022b 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., 2022b, Table III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| LeGO-LOAM | 5.16 m | (Chen et al., 2022b, Table III) |
| NDT-LOAM本方法原文提出 | 3.38 m | (Chen et al., 2022b, Table III) |
其他比較組
來源
Chen et al., 2022b
(2022)NDT-LOAM: A Real-Time Lidar Odometry and Mapping With Weighted NDT and LFAIEEE Sensors Journal, 22(4):3660-3671
DOI 10.1109/jsen.2021.3135055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
- 程式碼釋出:BurryChen/lv_slam (named in the paper abstract) https://github.com/BurryChen/lv_slam
程式碼:https://github.com/BurryChen/lv_slam(授權:not stated (no LICENSE file at the repository root))。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。