SLAM 點雲圖鑑Point clouds for construction
M-01 · 2026

PA-LVIO

PA-LVIO 提出僅含位姿的光束法平差（pose-only bundle adjustment），把光達與視覺的多幀幾何約束轉為幀間位姿約束，在滑動視窗因子圖中與 IMU 預積分緊密融合，以降低計算量。另加入不需邊緣化的幀對地圖（frame-to-map）光達位姿約束抑制漂移，並以 IMU 為中心線上估計光達、相機的時空參數，達到像素級對齊後產生 RGB 上色點雲地圖。

里程計＋局部建圖近十年已讀全文預印本LiDAR、視覺與慣性融合

閱讀版本
arXiv 2603.16228 v2 (2026-03-24), all 14 pages; no journal version found
全文取得途徑
arXiv 預印本
本頁內容

A sliding-window LVIO that converts LiDAR and visual multi-frame constraints into pose-only factors, adds a marginalization-free frame-to-map LiDAR factor, and calibrates spatio-temporal parameters online to render RGB point clouds.

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 (Livox AVIA 10 Hz on MARS-LVIG, R3LIVE and HandNav data; Hesai AT128 10 Hz on i2Nav-Robot)、IMU (MEMS; BMI088 on MARS-LVIG, R3LIVE and HandNav; ADIS16465 on i2Nav-Robot; 200 Hz)、RGB camera (models not reported; 1280x1024 to 2448x2048 at 10 to 15 Hz)
原文測試平台wheeled UGV、UAV、handheld
狀態估計INS-centric sliding-window factor graph optimization with pose-only bundle adjustment factors for LiDAR and visual measurements, IMU preintegration, and a marginalization-free frame-to-map LiDAR pose factor
資料關聯same-plane LiDAR points associated across keyframes (following BA-LINS); visual features tracked with INS prior; frame-to-map LiDAR pose optimization against a global ikd-Tree map
時間表示discrete poses; LiDAR frames projected to visual keyframe times; online camera-IMU and LiDAR time-delay and extrinsic calibration
去畸變point clouds undistorted using high-rate INS pose (Sec. II)
迴圈閉合none (authors note place recognition or loop closure could be added; Sec. I)
全域最佳化none
地圖表示global point cloud map in ikd-Tree; RGB-rendered point-cloud map
先驗資訊none
可輸出幾何RGB-rendered point-cloud map (abstract, Sec. III end)
計算需求desktop AMD Ryzen 9 9950X with CUDA-accelerated OpenCV on an NVIDIA RTX 5090 (Sec. IV-A; Sec. IV-D names RTX 5080); NVIDIA Orin NX (6-core CPU, 8 GB) on HandNav with images resized to 900x600: average LiDAR 3.27 vs 20.40 ms, visual 7.07 vs 28.91 ms and FGO 39.06 vs 85.35 ms per frame on PC vs ARM (Table VI), equivalent 33.2 Hz and 13.0 Hz (Table VII); on i2Nav-Robot the F2M pose optimization averages 3.19 ms and the loosely coupled FGO 26.96 ms vs 33.95 ms tightly coupled (Table IV)

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
LiDARLivox AVIA歸入：Livox Avia方法輸入HandNav (private)10 Hz (Table I)(Tang et al., 2026, Table I; Fig. 4)
LiDARHesai AT128資料集感測器i2Nav-Robot10 Hz (Table I)(Tang et al., 2026, Table I)
LiDARLivox AVIA歸入：Livox Avia資料集感測器MARS-LVIG10 Hz (Table I)(Tang et al., 2026, Table I)
LiDARLivox AVIA歸入：Livox Avia資料集感測器R3LIVE dataset10 Hz (Table I)(Tang et al., 2026, Table I)
慣性量測單元（IMU）BMI088方法輸入HandNav (private)200 Hz (Table I)(Tang et al., 2026, Table I)
慣性量測單元（IMU）ADIS16465資料集感測器i2Nav-Robot200 Hz (Table I)(Tang et al., 2026, Table I)
慣性量測單元（IMU）BMI088資料集感測器MARS-LVIG200 Hz (Table I)(Tang et al., 2026, Table I)
慣性量測單元（IMU）BMI088資料集感測器R3LIVE dataset200 Hz (Table I)(Tang et al., 2026, Table I)
GNSS 接收器RTK (low-rate, provided with MARS-LVIG) post-processed with RTK/INS software into high-rate ground truth參考或真值量測MARS-LVIG原文未報告(Tang et al., 2026, Sec. IV-B2)
相機RGB camera (model not reported)方法輸入HandNav (private)1800x1200, 10 Hz; resized to 900x600 on the ARM computer (Table I; Sec. IV-D)(Tang et al., 2026, Table I; Sec. IV-D)
相機camera of i2Nav-Robot (model not reported)資料集感測器i2Nav-Robot1600x1200, 10 Hz; pixel size 5.86 um, focal length 6 mm (Table I; Sec. IV-C1)(Tang et al., 2026, Table I; Sec. IV-C1)
相機camera of MARS-LVIG (model not reported)資料集感測器MARS-LVIG2448x2048, 10 Hz (Table I)(Tang et al., 2026, Table I)
相機camera of the R3LIVE dataset (model not reported)資料集感測器R3LIVE dataset1280x1024, 15 Hz (Table I)(Tang et al., 2026, Table I)
載具平台HandNav handheld device方法輸入HandNav (private)3 sequences, 1.5 km (Table I)(Tang et al., 2026, Table I; Fig. 4)
載具平台low-speed wheeled robot (model not reported)資料集感測器i2Nav-Robot10 sequences, 17.1 km(Tang et al., 2026, Sec. IV-A; Table I)
載具平台UAV (model not reported)資料集感測器MARS-LVIG3 to 12 m/s, altitude 80 to 130 m; 9 sequences, 32.6 km(Tang et al., 2026, Sec. IV-B2; Table I)
運算硬體AMD Ryzen 9 9950X CPU執行運算平台未標示desktop PC(Tang et al., 2026, Sec. IV-A; Sec. IV-D)
運算硬體NVIDIA RTX 5090 GPU (Sec. IV-A); RTX 5080 GPU (Sec. IV-D)執行運算平台未標示CUDA-accelerated OpenCV for visual processing(Tang et al., 2026, Sec. IV-A; Sec. IV-D)
運算硬體NVIDIA Orin NX執行運算平台HandNav (private)6-core CPU, 8 GB RAM(Tang et al., 2026, Sec. IV-A; Sec. IV-D)

作者報告的優勢與限制

優勢

  • Evaluated on 28 sequences exceeding 50 km across wheeled robot, UAV and handheld data (abstract)
  • Online LiDAR-camera rotation estimates consistent across ten sequences (standard deviation below 0.04 deg per axis) (Table V)
  • Real-time on onboard ARM computer (abstract; Sec. IV)

限制

  • Marginalization-free F2M measurement may reduce odometry accuracy in certain circumstances (Sec. V; Sec. IV-B4)
  • Factor graph optimization is computationally heavy; MSCKF variant planned (Sec. V)
  • Mapping quality evaluated only qualitatively on selected scenes (Sec. IV-C2; Figs. 9-10)
  • On the short campus02 and hku-park0 sequences (under 500 s) the end-to-end error is decimeter-level (0.18 m and 0.13 m) while FAST-LIO2, R3LIVE and FAST-LIVO2 reach centimeter level (Table III; Sec. IV-B3)
  • No loop-closure or place-recognition module is described in the pipeline; drift is addressed by the marginalization-free F2M pose factor (Sec. II; Sec. III-D)

營建工程相關證據

原文未報告（i2Nav-Robot 為建物、停車場、操場與街道序列；MARS-LVIG 為 80 至 130 m 高度的無人機序列；R3LIVE 資料集與自採 HandNav 為手持序列，HandNav 拍攝武漢大學傳統建築；未見營建工地）

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 5 個比較組，合計 39 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 1 組列在最後，並連到性能比較頁。

Tang et al., 2026 · Table II 本方法 21 筆

指標absolute translation error (RMSE, meters)

表格設定（擷取紀錄原文）：Absolute translation error (RMSE, m), all systems in real-time mode on the desktop PC; 'x' = system totally failed. Ablation columns (Ours VIO, LIO, w/o F2M, Marg. F2M, w/o calib.) omitted here. MARS-LVIG ground truth re-derived by the authors with post-processed RTK/INS; i2Nav-Robot ground-truth source not described in this paper. Average rows printed by the authors. (Tang et al., 2026, Table II)

absolute translation error (RMSE, meters)，i2Nav-Robot · building00

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

這些是 Tang 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。

統計量：均方根誤差（RMSE）；對齊方式：原文未報告；單位：m；場景：low-speed wheeled robot; building, parking, playground and street sequences (indoor-outdoor)

資料來源作者報告值(Tang et al., 2026, Table II)

數值與出處
方法（原文寫法）報告值出處
FF-LINS2.32 m(Tang et al., 2026, Table II)
FAST-LIO20.68 m(Tang et al., 2026, Table II)
LE-VINS1.17 m(Tang et al., 2026, Table II)
R3LIVE無數值失敗註記（擷取紀錄）：failed(Tang et al., 2026, Table II)
FAST-LIVO21.1 m(Tang et al., 2026, Table II)
Ours (PA-LVIO)本方法原文提出0.34 m(Tang et al., 2026, Table II)

Tang et al., 2026 · Table III 本方法 7 筆

指標end-to-end error (meters)

表格設定（擷取紀錄原文）：End-to-end errors (m) on the public R3LIVE handheld dataset, computed by subtracting start positions from end positions; ablation columns omitted; Average row printed by the authors (Tang et al., 2026, Table III)

end-to-end error (meters)，R3LIVE dataset · main-building

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。

這些是 Tang 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。

統計量：原文未報告；對齊方式：未對齊；單位：m；場景：handheld campus, park and building sequences (Hong Kong)

資料來源作者報告值(Tang et al., 2026, Table III)

數值與出處
方法（原文寫法）報告值出處
FF-LINS1.2 m(Tang et al., 2026, Table III)
FAST-LIO21.38 m(Tang et al., 2026, Table III)
LE-VINS0.96 m(Tang et al., 2026, Table III)
R3LIVE0.11 m(Tang et al., 2026, Table III)
FAST-LIVO21.27 m(Tang et al., 2026, Table III)
Ours (PA-LVIO)本方法原文提出0.08 m(Tang et al., 2026, Table III)

Tang et al., 2026 · Table VI 本方法 6 筆

資料集與序列HandNav (private) · Average

表格設定（擷取紀錄原文）：Average processing time on the three private HandNav sequences (whu-building, whu-gateway, whu-library); LiDAR and visual include preprocessing and data association (Tang et al., 2026, Table VI)

LiDAR (ms) on PC，HandNav (private) · Average

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 Tang et al., 2026 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。

統計量：平均值（mean）；對齊方式：未對齊；單位：ms；場景：handheld, Wuhan University buildings

數值與出處
方法（原文寫法）報告值出處
PA-LVIO本方法原文提出硬體：desktop PC, AMD Ryzen 9 9950X CPU + NVIDIA RTX 5090 GPU (Sec. IV-A; Sec. IV-D names RTX 5080)3.27 ms(Tang et al., 2026, Table VI)

Tang et al., 2026 · Table IV 本方法 3 筆

資料集與序列i2Nav-Robot · Average

表格設定（擷取紀錄原文）：Average over the 10 i2Nav-Robot sequences of per-keyframe FGO time for tightly vs loosely coupled F2M factors, and of the F2M pose optimization (Tang et al., 2026, Table IV)

FGO (ms), tightly coupled F2M variant，i2Nav-Robot · Average

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 Tang et al., 2026 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。

統計量：平均值（mean）；對齊方式：未對齊；單位：ms；場景：wheeled robot

數值與出處
方法（原文寫法）報告值出處
PA-LVIO with tightly coupled F2M factor本方法原文提出硬體：desktop PC, AMD Ryzen 9 9950X CPU + NVIDIA RTX 5090 GPU (Sec. IV-A; Sec. IV-D names RTX 5080)33.95 ms(Tang et al., 2026, Table IV)

其他比較組

列出其餘 1 個比較組

來源

  • Tang et al., 2026

    Hailiang Tang, Tisheng Zhang, Liqiang Wang, Xin Ding, Man Yuan, Xiaoji Niu（2026）PA-LVIO: Real-Time LiDAR-Visual-Inertial Odometry and Mapping with Pose-Only Bundle AdjustmentarXiv

    arXiv 2603.16228程式碼

    預印本已讀全文近十年查證後修正

程式碼：https://github.com/i2Nav-WHU/PA-LVIO（授權：GPL-3.0 LICENSE file present and README restricts use to academic purposes; however, as of 2026-09-25 the repository contains only README.md, LICENSE and paper/ (README news 2026-03-19: 'The codes will be released soon'), although the arXiv abstract states the code is open-sourced）。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。

回到方法圖鑑

SLAM 點雲圖鑑

本站為綜述論文《SLAM-Based Point Clouds for Construction: A Review》研究過程的互動說明文件。互動圖為教學模擬，不代表任何特定系統的實測表現；方法描述、數值與結論以所引原始文獻為準。

內容基準日 2026-09-25 · 研究與網站製作使用 AI 輔助工具，見 證據標記與方法

選擇開啟Esc關閉