FAST-LIVO2 on Resource-Constrained Platforms
此研究針對邊緣運算平台精簡 FAST-LIVO2:以光達退化評估決定何時需要影像更新,在光達約束充足時減少視覺幀,降低計算量;地圖改為小範圍的統一視覺光達局部地圖加上稀疏的長期視覺地圖,以限制記憶體。作者在 Hilti(資料集描述含工地序列)16 個序列與私人序列(礦坑隧道、黑暗樹林、野外公園等)上測試,並在約 100 美元的 RK3588 ARM 板上於地下停車場與夜間街道進行即時定位測試。
本頁內容
A lightweight FAST-LIVO2 variant with degeneration-aware adaptive visual frame selection and a compact local plus long-term visual map for ARM edge devices.
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 (Hilti dataset LiDARs、Livox Mid-360 on the authors' rig、a small-FoV AVIA LiDAR, written 'Aivia' in Sec. V-D2, in the degeneration-detection test of Fig. 9, sequence not named) | IMU | camera (B/W fisheye on the authors' rig) | 15 W onboard illuminator for extremely dark scenes |
|---|---|
| 原文測試平台 | handheld | ground robot (Hilti robot-mounted sequences、type 未查證) | aerial dataset (MARS-LVIG HKIsland03 map ablation) | ARM real-time tests in an underground parking lot and a nighttime street (carrying mode not stated) |
| 狀態估計 | ESIKF with sequential updates (from FAST-LIVO2) plus a LiDAR-degeneration-aware adaptive visual frame selector |
| 資料關聯 | LiDAR point-to-plane and patch photometric errors as in FAST-LIVO2; images used only when LiDAR constraints are weak or keyframe criteria met |
| 時間表示 | discrete poses; scan recombination (Sec. III) |
| 去畸變 | Uses undistorted points of recombined scans (scan recombination); the undistortion step itself is inherited from FAST-LIVO2 and not re-described |
| 迴圈閉合 | none reported |
| 全域最佳化 | none |
| 地圖表示 | Compact unified local visual-LiDAR voxel map (hash of 0.5 m root voxels with three-level octrees; typical edge 200 m, slid every 20 m of motion) plus a sparse long-term visual map (edge 800 m, slid every 100 m) that keeps visual points leaving the local map |
| 先驗資訊 | none |
| 可輸出幾何 | sharp point cloud maps and colored point clouds (Sec. V-C1; Fig. 10) |
| 計算需求 | Hilti mean per-frame time 25.99 ms on the x86 laptop (13th Gen Intel Core i9-13900HX) versus 35.66 ms for FAST-LIVO2, and 57.82 ms on the RK3588 board (4x Cortex-A76 + 4x Cortex-A55, up to 2.4 GHz, about 100 USD, CPU only) versus 75.87 ms; about 37 ms per frame in the onboard ARM tests; mean memory 1.7 GB versus 2.5 GB on Hilti |
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 | Livox Mid-360歸入:Livox MID-360 | 方法輸入 | private dataset | 原文未報告 | (Zhou et al., 2025, Sec. V-A2; Fig. 5) |
| LiDAR | AVIA (written 'Aivia' in Sec. V-D2)歸入:Livox Avia | 資料集感測器 | unnamed LiDAR-degeneration test sequence of Fig. 9 (dataset not stated) | small-FoV LiDAR, facing a wall in the degeneration test | (Zhou et al., 2025, Sec. IV-A2; Sec. V-D2; Fig. 9) |
| GNSS 接收器 | RTK | 參考或真值量測 | MARS-LVIG HKIsland03 | RTK trajectory used as ground truth | (Zhou et al., 2025, Sec. V-E1) |
| 相機 | B/W camera with a fisheye lens | 方法輸入 | private dataset | equidistant projection model | (Zhou et al., 2025, Sec. V-A2, V-A3) |
| 運算硬體 | 13th Gen Intel Core i9-13900HX (personal laptop) | 執行運算平台 | 未標示 | x86 platform | (Zhou et al., 2025, Sec. V-A3) |
| 運算硬體 | RK3588 | 執行運算平台 | 未標示 | octa-core 4x Cortex-A76 + 4x Cortex-A55, max 2.4 GHz, about 100 USD; CPU only | (Zhou et al., 2025, Sec. V-A3; Fig. 1) |
| 其他 | STM32 microcontroller | 資料集感測器 | private dataset | hardware synchronization of LiDAR and camera | (Zhou et al., 2025, Sec. V-A2) |
| 其他 | onboard illuminator | 資料集感測器 | private dataset | 15 W, used in extremely dark environments | (Zhou et al., 2025, Sec. V-A2) |
| 其他 | Hilti handheld and robot-mounted rigs (LiDAR, cameras, IMUs; models not stated in this paper) | 資料集感測器 | Hilti'22 and Hilti'23 | sensors at different frequencies | (Zhou et al., 2025, Sec. V-A1) |
作者報告的優勢與限制
優勢
- On Hilti data, 33% lower per-frame runtime and 47% lower memory than FAST-LIVO2 with about 3 cm higher RMSE (abstract)
- Return-to-origin drift below 2 cm on private sequences including a mining tunnel (Sec. V-C1)
- 37 ms per frame on RK3588 ARM (Sec. V-C2)
限制
- Accuracy slightly lower than FAST-LIVO2 on Hilti (average 0.063 m vs 0.034 m) (abstract
- Table I) | Much larger error than FAST-LIVO2 on Construction Stairs (0.170 vs 0.016 m), Cupola (0.220 vs 0.121 m) and Attic to Upper Gallery (0.180 vs 0.069 m) (Table I) | Slightly worse than FAST-LIO2 in visually challenging sequences such as overexposed Outside Building and dark Large Room (Sec. V-B1) | Long-term visual map adds memory overhead (Sec. V-E2)
營建工程相關證據
作者以 Hilti 2022 與 2023 共 16 個序列評估軌跡精度,論文描述其場景含營建工地、辦公室與地下室,並使用 Hilti 官方評估工具。Hilti 2022 三個工地序列的 RMSE 為 Construction Ground 0.010 m、Construction Multilevel 0.023 m、Construction Stairs 0.170 m,其中樓梯序列明顯遜於 FAST-LIVO2 的 0.016 m。另在礦坑隧道與地下停車場做定性測試。證據屬軌跡層級,未報告點雲幾何精度。
原文驗證環境:公開基準、施工中工地、地下或隧道、獨立參考量測
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 30 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。
Zhou et al., 2025 · Table I 本方法 17 筆
表格設定(擷取紀錄原文):ATE RMSE on 16 Hilti'22 and Hilti'23 sequences computed with the official Hilti evaluation tools; parameters of all methods tuned by the authors; x = system totally failed (Zhou et al., 2025, Table I)
ATE (RMSE),Hilti'22 · Construction Ground
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 Zhou et al., 2025 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(Zhou et al., 2025, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Ours本方法原文提出 | 0.01 m | (Zhou et al., 2025, Table I) |
| FAST-LIVO2 | 0.01 m | (Zhou et al., 2025, Table I) |
| FAST-LIO2 | 0.013 m | (Zhou et al., 2025, Table I) |
| FAST-LIVO | 0.022 m | (Zhou et al., 2025, Table I) |
| R3LIVE | 0.021 m | (Zhou et al., 2025, Table I) |
| SDV-LOAM | 25.121 m | (Zhou et al., 2025, Table I) |
| LVI-SAM | 無數值失敗註記(擷取紀錄):failed (x) | (Zhou et al., 2025, Table I) |
Zhou et al., 2025 · Table II 本方法 6 筆
資料集與序列Hilti'22 and Hilti'23 · mean over 16 sequences
表格設定(擷取紀錄原文):Mean per-frame time over the 16 Hilti sequences, standard error given in metric_as_written; ARM runtime measured on CPU only (Zhou et al., 2025, Table II)
LiDAR Part time, mean (standard error 5.35 ms),Hilti'22 and Hilti'23 · mean over 16 sequences
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 Zhou et al., 2025 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 |
|---|---|---|
| Ours本方法原文提出硬體:RK3588 (4x Cortex-A76 + 4x Cortex-A55, up to 2.4 GHz), CPU only | 53.83 ms | (Zhou et al., 2025, Table II) |
Zhou et al., 2025 · Table III 本方法 4 筆
指標memory usage
表格設定(擷取紀錄原文):Memory usage on the x86 laptop; only the three Hilti'22 construction sequences and the Average row kept for the row cap (Zhou et al., 2025, Table III)
memory usage,Hilti'22 · Construction Ground
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 Zhou et al., 2025 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(Zhou et al., 2025, Table III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Ours本方法原文提出硬體:laptop, 13th Gen Intel Core i9-13900HX | 3 GB | (Zhou et al., 2025, Table III) |
| FAST-LIVO2硬體:laptop, 13th Gen Intel Core i9-13900HX | 4.1 GB | (Zhou et al., 2025, Table III) |
Zhou et al., 2025 · Text Sec.V-C1 本方法 1 筆
指標return-to-origin drift
資料集與序列private sequences · Mining Tunnel, Dark Woods, Wild Park, HIT Graffiti Wall, HW Corridor and others
表格設定(擷取紀錄原文):Return-to-origin drift on private sequences that physically return to the start; value is an upper bound (Zhou et al., 2025, Text Sec.V-C1)
return-to-origin drift,private sequences · Mining Tunnel, Dark Woods, Wild Park, HIT Graffiti Wall, HW Corridor and others
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 Zhou et al., 2025 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 |
|---|---|---|
| Ours本方法原文提出 | 0.02 m僅報告範圍註記(擷取紀錄):reported as less than 2 cm (upper bound) | (Zhou et al., 2025, Sec. V-C1; Fig. 6) |
其他比較組
來源
Zhou et al., 2025
(2025)FAST-LIVO2 on Resource-Constrained Platforms: LiDAR-Inertial-Visual Odometry With Efficient Memory and ComputationIEEE Robotics and Automation Letters, 10(8): 7931-7938
DOI 10.1109/lra.2025.3581125arXiv 2501.13876
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:arXiv v1 (2025-01-23) https://arxiv.org/abs/2501.13876