GS-LIVO
GS-LIVO 以三維高斯(3D Gaussians)取代傳統彩色點雲與稀疏區塊地圖:全域高斯地圖以空間雜湊索引的八元樹管理,只將視野內的高斯放入 GPU 上的滑動視窗即時最佳化,以控制顯示記憶體用量。高斯由光達點與影像聯合初始化,里程計沿用 FAST-LIVO2 的序列更新 IESKF,但視覺殘差改為渲染影像與實際影像的光度誤差。作者宣稱這是首個可在 Jetson Orin NX 嵌入式平台即時運作並線上更新地圖的高斯式 SLAM。
本頁內容
Replaces colored point clouds with a hash-octree Gaussian map optimized in a GPU sliding window and uses rendered-image photometric residuals inside a FAST-LIVO2-derived sequential IESKF.
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、IMU、camera |
|---|---|
| 原文測試平台 | aerial robotic vehicles (MARS-LVIG) | mobile chassis carrying the sensor suite and Jetson Orin NX | carrying mode of the FAST-LIVO2 sequences, the Playground sequences and (T-RO) Oxford Spires not stated in this paper |
| 狀態估計 | iterated error-state Kalman filter with sequential updates, modified from FAST-LIVO2 |
| 資料關聯 | LiDAR update with planar features of a size-adaptive voxel map (FAST-LIVO2 and VoxelMap-type LIO); visual update minimizes the photometric loss between the image rendered from Gaussians in the current FoV at the LiDAR-updated pose and the captured image, with Jacobians derived as in MonoGS and chained to the IMU pose inside the IESKF |
| 時間表示 | discrete poses; emulated PPS hardware synchronization |
| 去畸變 | Not described; the LiDAR-inertial update is taken from FAST-LIVO2 and size-adaptive voxel LIO ([57], [59] in T-RO) without re-describing motion compensation |
| 迴圈閉合 | none reported |
| 全域最佳化 | none |
| 地圖表示 | Planar 3D Gaussians initialized from LiDAR leaf voxels (normal from LiDAR, color by bilinear sampling) in a global hash-indexed octree in CPU RAM; Gaussians in the current FoV kept in a contiguous CPU buffer mirrored in GPU memory and optimized with Adam; root voxel 0.03 or 0.06 m indoors and 1.0 or 0.5 m outdoors with 2 subdivision levels; window of 100,000 Gaussians (desktop) or 20,000 (Orin NX) |
| 先驗資訊 | offline camera intrinsic and LiDAR-camera extrinsic calibration |
| 可輸出幾何 | Gaussian map with photorealistic rendering; 2D occupancy grid derived for navigation (Sec. III-D); point-cloud export 原文未報告 |
| 計算需求 | GPU required; desktop i9-13900KF, 128 GB RAM, RTX-4090: 48.5 to 94.8 ms per frame; map updates over 10 Hz indoors and about 3 Hz outdoors (Sec. I-B); Jetson Orin NX 16 GB: 15.3 ms optimization, 18.9 ms map maintenance, 48.3 ms total at 256x216 images and 20,000 Gaussians |
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 | LiDAR (model not stated) | 方法輸入 | 未標示 | 原文未報告 | (Hong et al., 2025, Sec. II; Fig. 8(e) arXiv, Fig. 12(e) T-RO) |
| 地面雷射掃描儀(TLS) | TLS (model not stated in this paper) | 參考或真值量測 | Oxford Spires (Radcliffe01; T-RO only) | LiDAR-TLS map registration giving 1 to 2 cm ground-truth trajectories | (Hong et al., 2025, T-RO Sec. III-A) |
| 慣性量測單元(IMU) | IMU (model not stated) | 方法輸入 | 未標示 | 原文未報告 | (Hong et al., 2025, Sec. II) |
| GNSS 接收器 | D-RTK (DJI Differential Real-Time Kinematic GNSS system) | 參考或真值量測 | MARS-LVIG | precise ground truth for odometry | (Hong et al., 2025, Sec. III-A (T-RO wording)) |
| 相機 | camera (model not stated; pinhole projection model) | 方法輸入 | 未標示 | intrinsics calibrated with a checkerboard | (Hong et al., 2025, Sec. II; Sec. III-A) |
| 載具平台 | mobile chassis | 方法輸入 | 未標示 | carries the sensor suite and Jetson Orin NX | (Hong et al., 2025, Fig. 8(e) arXiv; Fig. 12(e) T-RO) |
| 載具平台 | aerial robotic vehicles | 資料集感測器 | MARS-LVIG | MARS-LVIG data collection over mountains and seas | (Hong et al., 2025, Sec. III-A) |
| 運算硬體 | NVIDIA Jetson Orin NX | 執行運算平台 | 未標示 | 8-core CPU, 1024 CUDA cores, 16 GB LPDDR5 | (Hong et al., 2025, Abstract footnote; Sec. III-D) |
| 運算硬體 | desktop with Intel i9-13900KF CPU and NVIDIA RTX-4090 GPU | 執行運算平台 | 未標示 | 128 GB RAM | (Hong et al., 2025, Sec. III) |
| 其他 | emulated pulse-per-second (PPS) synchronization | 方法輸入 | 未標示 | temporal alignment of LiDAR, IMU and camera | (Hong et al., 2025, Sec. II) |
| 其他 | motion capture system (MoCap) | 參考或真值量測 | proprietary Playground sequences | ground truth for small indoor sequences; tracker-odometry alignment calibrated | (Hong et al., 2025, Sec. III-A) |
作者報告的優勢與限制
優勢
- Outdoor trajectory RMSE of 0.58 m (HKisland03), 0.63 m (HKairport01) and, in the T-RO version, 0.75 m (Radcliffe01), versus 1.71, 1.22 and 1.85 m for R3LIVE and 4.12, 5.21 and 6.15 m for LVI-SAM
- the FAST-LIVO baseline was better outdoors (0.51, 0.56, 0.63 m). The 0.042 m outdoor RMSE in the Sec. III-B2 text contradicts the table, whereas the quoted R3LIVE 1.465 m and LVI-SAM 4.665 m equal the means of their MARS-LVIG rows, so 0.042 m should not be cited | Sliding window keeps map update time below 100 ms and PSNR around 25 dB while limiting GPU memory (Sec. III-C) | Runs on Jetson Orin NX (Sec. III-D)
限制
- Localization slightly less precise than the FAST-LIVO baseline, which the tables label 'FAST-LIVO [7]' (IROS 2022) while the text cites [8] (FAST-LIVO2), so the baseline identity is ambiguous
- per-frame time 48.5 to 94.8 ms versus 8.75 to 49.8 ms for that baseline (Sec. III-B2
- T-RO Table II) | Level-of-detail adaptation and merging of homogeneous Gaussians left for future work (arXiv Sec. III-E) | T-RO conclusion: the system still struggles with indoor-outdoor transitions
- size-adaptive voxels proposed (T-RO Sec. IV) | Mapping quality evaluated by rendering metrics (PSNR), not metric geometry accuracy (inference from Sec. III) | GS-LIVO memory for Playground01 and Playground02 is swapped between the rendering table (1.5, 1.2 GB) and the GS-SLAM table (1.2, 1.5 GB), and the rendering-table 'Dur./s' values equal the per-frame ms values (T-RO Tables I and III)
營建工程相關證據
原文未報告(資料為 FAST-LIVO2 校園序列、MARS-LVIG 空拍與小型室內動作捕捉場地)
原文驗證環境:公開基準、受控實驗、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 5 個比較組,合計 36 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。
Hong et al., 2025 · Table III 本方法 15 筆
表格設定(擷取紀錄原文):Gaussian-based SLAM comparison (T-RO Table III; arXiv v1 Table IV without Radcliffe01); MonoGS* uses LiDAR-projected depth, MonoGS is monocular; x = failed on all outdoor sequences (one row per failed method and sequence); Dur./ms column header carries an upward arrow in the table (Hong et al., 2025, Table III)
RMSE/m,proprietary (MoCap) · Playground01
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 Hong 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(Hong et al., 2025, Table III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| SplaTAM | 0.28 m | (Hong et al., 2025, T-RO Table III; arXiv v1 Table IV) |
| MonoGS* | 0.09 m | (Hong et al., 2025, T-RO Table III; arXiv v1 Table IV) |
| MonoGS | 0.18 m | (Hong et al., 2025, T-RO Table III; arXiv v1 Table IV) |
| GS-LIVO (Ours)本方法原文提出 | 0.006 m | (Hong et al., 2025, T-RO Table III; arXiv v1 Table IV) |
Hong et al., 2025 · Table II 本方法 10 筆
表格設定(擷取紀錄原文):LIV-based SLAM comparison (T-RO Table II; arXiv v1 Table III without Radcliffe01); image 640x480; octree 0.06 m (indoor) or 0.5 m (outdoor), 2 layers; window 100,000 Gaussians; ground truth D-RTK (MARS-LVIG), MoCap (Playground), TLS-registered trajectories (Oxford Spires Radcliffe01); baseline labelled 'FAST-LIVO [7]' but the text attributes it to [8] (FAST-LIVO2), so method_id left null (Hong et al., 2025, Table II)
RMSE/m,MARS-LVIG · HKisland03
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 Hong 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(Hong et al., 2025, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| FAST-LIVO [7] | 0.51 m | (Hong et al., 2025, T-RO Table II; arXiv v1 Table III) |
| R3LIVE | 1.71 m | (Hong et al., 2025, T-RO Table II; arXiv v1 Table III) |
| LVI-SAM | 4.12 m | (Hong et al., 2025, T-RO Table II; arXiv v1 Table III) |
| GS-LIVO (Ours)本方法原文提出 | 0.58 m | (Hong et al., 2025, T-RO Table II; arXiv v1 Table III) |
Hong et al., 2025 · Table I 本方法 6 筆
指標PSNR (dB), higher is better
表格設定(擷取紀錄原文):Rendering comparison (T-RO Table I; arXiv v1 Table II without M2Mapping); 15,000 iterations per method; indoor root voxel 0.03 m, outdoor 1.0 m, 2 levels; SplaTAM and MonoGS fed with LiDAR-projected depth; x = failed; Dur./s and Mem./GB columns omitted for the row cap (Hong et al., 2025, Table I)
PSNR (dB), higher is better,FAST-LIVO2 dataset · HKU01
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 Hong 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(Hong et al., 2025, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| 3D-GS | 26.22 dB | (Hong et al., 2025, T-RO Table I; arXiv v1 Table II) |
| M2Mapping | 25.06 dB | (Hong et al., 2025, T-RO Table I; arXiv v1 Table II) |
| SplaTAM | 24.06 dB | (Hong et al., 2025, T-RO Table I; arXiv v1 Table II) |
| MonoGS | 23.51 dB | (Hong et al., 2025, T-RO Table I; arXiv v1 Table II) |
| S3GS | 無數值失敗註記(擷取紀錄):failed (x) | (Hong et al., 2025, T-RO Table I; arXiv v1 Table II) |
| LetsGo | 24.51 dB | (Hong et al., 2025, T-RO Table I; arXiv v1 Table II) |
| GS-LIVO (Ours)本方法原文提出 | 25.34 dB | (Hong et al., 2025, T-RO Table I; arXiv v1 Table II) |
Hong et al., 2025 · Text Sec.III-D 本方法 4 筆
資料集與序列原文未報告 · embedded platform run
表格設定(擷取紀錄原文):Embedded test on Jetson Orin NX 16 GB: root voxel 0.5 m, 2 layers, 256x216 images, window of 20,000 Gaussians (Hong et al., 2025, Text Sec.III-D)
optimization time,原文未報告 · embedded platform run
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 Hong 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 |
|---|---|---|
| GS-LIVO (Ours)本方法原文提出硬體:NVIDIA Jetson Orin NX 16 GB | 15.3 ms | (Hong et al., 2025, Sec. III-D; Fig. 8 (arXiv) or Fig. 12 (T-RO)) |
其他比較組
列出其餘 1 個比較組
來源
Hong et al., 2025
(2025)GS-LIVO: Real-Time LiDAR, Inertial, and Visual Multisensor Fused Odometry With Gaussian MappingIEEE Transactions on Robotics, 41: 4253-4268
DOI 10.1109/tro.2025.3582809arXiv 2501.08672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:GS-LIVO arXiv v1 (2025-01-15) https://arxiv.org/abs/2501.08672
- 程式碼釋出:HKUST-Aerial-Robotics/GS-LIVO https://github.com/HKUST-Aerial-Robotics/GS-LIVO
程式碼:https://github.com/HKUST-Aerial-Robotics/GS-LIVO(授權:GPL-2.0 (LICENSE file in src/gs-livo; root README is a demo page without a license statement; package.xml carries template metadata 'BSD' with Ji Zhang as author, an inconsistency inherited from a template))。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。