OKVIS2-X
OKVIS2-X 以關鍵影格式視覺慣性 SLAM（OKVIS2）為核心，可選擇加入深度網路估計的稠密深度、LiDAR 或 GNSS。系統把 supereight2 體素占據子地圖綁定在關鍵影格上，並以影格對地圖、地圖對地圖的占據對齊殘差把子地圖與狀態估計緊耦合，使迴圈閉合或 GNSS 修正後地圖可隨位姿一起更新，形成全域一致、可直接用於導航的稠密地圖。GNSS 以可觀測性準則處理座標對齊與訊號中斷，並支援相機外參線上校正。
本頁內容
Extends OKVIS2 visual-inertial SLAM with keyframe-anchored volumetric occupancy submaps built from learned depth or LiDAR and tightly coupled by occupancy alignment factors, plus optional GNSS with observability-aware alignment and online extrinsic calibration, giving globally consistent dense maps at large scale.
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 | one or more cameras、IMU、optional learned depth (stereo and multi-view stereo networks)、optional LiDAR、optional GNSS |
|---|---|
| 原文測試平台 | GVINS dataset complex_environment sequence (carrying platform not described in this paper) |
| 狀態估計 | OKVIS2 keyframe-based visual-inertial SLAM (BRISK features, realtime sliding-window estimator plus asynchronous full-graph optimization with posegraph edges from marginalized landmarks, DBoW2 loop closure) extended with volumetric occupancy submaps anchored to keyframes and tightly coupled through frame-to-map and map-to-map occupancy alignment factors (Tukey robustifier), optional GNSS position factors and online camera-IMU extrinsic calibration (Secs. IV-V) |
| 資料關聯 | BRISK keypoint matching for landmarks; dense submap alignment by occupancy residuals of depth or LiDAR points against submaps; sky segmentation (Fast-SCNN) to remove sky pixels from depth (Sec. V) |
| 時間表示 | discrete multi-camera frames with IMU pre-integration between states; LiDAR points and GNSS measurements are related to states through IMU propagation (Sec. V) |
| 去畸變 | not described as a separate step |
| 迴圈閉合 | yes, DBoW2 place recognition with full-graph optimization (Sec. V) |
| 全域最佳化 | asynchronous full-graph optimization and optional final full bundle adjustment ('-ba' variants); GNSS alignment with a yaw-observability criterion and global re-alignment after GNSS dropouts (Sec. V) |
| 地圖表示 | supereight2 volumetric occupancy submaps anchored to keyframes, meshable into a globally consistent map (Figs. 1, 9, 10) |
| 先驗資訊 | camera intrinsics and initial extrinsics; networks fine-tuned on synthetic data |
| 可輸出幾何 | trajectory (causal, non-causal and full-BA variants) and dense occupancy submaps or meshes |
| 計算需求 | desktop with Intel i7-13700 and RTX-3080 10 GB; Ours-vi up to 47 Hz on EuRoC MH05; MVS network at 13 Hz on 8-view images; GPU memory 3.51 GB for the networks; also runs on a drone with NVIDIA Orin NX (Sec. VI) |
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 | Hilti-Oxford 32-channel LiDAR (model not reported) | 資料集感測器 | Hilti-Oxford | 32-channel point clouds on the handheld device; far plane 30 m in mapping | (Boche et al., 2025, Sec. VI-C) |
| LiDAR | VBR LiDAR (model not reported) | 資料集感測器 | VBR (Vision Benchmark in Rome) | far plane limited to 30 m | (Boche et al., 2025, Sec. VI-D) |
| 慣性量測單元（IMU） | EuRoC MAV IMU (model not reported in this paper) | 資料集感測器 | EuRoC MAV | IMU measurements recorded with the stereo images by a drone | (Boche et al., 2025, Sec. VI-B) |
| 慣性量測單元（IMU） | Hilti-Oxford handheld device IMU (model not reported in this paper) | 資料集感測器 | Hilti-Oxford | IMU measurements; known camera-IMU time offset accounted for | (Boche et al., 2025, Secs. VI, VI-C) |
| 慣性量測單元（IMU） | VBR IMU (model not reported in this paper) | 資料集感測器 | VBR (Vision Benchmark in Rome) | IMU measurements; the driving sequences Campus* and Ciampino* contain episodes of missing IMU measurements | (Boche et al., 2025, Sec. VI-D) |
| 慣性量測單元（IMU） | GVINS-Dataset IMU (model not reported in this paper) | 資料集感測器 | GVINS-Dataset | IMU recorded with the stereo camera and the ZED-F9P GNSS sensor | (Boche et al., 2025, Sec. VI-E) |
| GNSS 接收器 | ZED-F9P GNSS sensor | 資料集感測器 | GVINS-Dataset | raw measurements and RTK solutions provided by the GVINS dataset; SPP computed with RTKLIB | (Boche et al., 2025, Sec. VI-E) |
| 相機 | Hilti-Oxford handheld rig cameras (5 cameras; model not reported) | 資料集感測器 | Hilti-Oxford | all 5 used by the estimator, 2 front cameras for the stereo network, front-left for MVS | (Boche et al., 2025, Sec. VI-C) |
| 雙目相機 | EuRoC MAV stereo camera (model not reported) | 資料集感測器 | EuRoC MAV | stereo images with IMU on a drone | (Boche et al., 2025, Sec. VI-B) |
| 雙目相機 | VBR stereo camera (model not reported; high resolution) | 資料集感測器 | VBR (Vision Benchmark in Rome) | with IMU, LiDAR and ground-truth poses | (Boche et al., 2025, Sec. VI-D) |
| 雙目相機 | GVINS-Dataset stereo camera (model not reported in this paper) | 資料集感測器 | GVINS-Dataset | stereo camera recorded together with an IMU and a ZED-F9P GNSS sensor | (Boche et al., 2025, Sec. VI-E) |
| 運算硬體 | i7-13700 | 執行運算平台 | 未標示 | desktop CPU; all timings in Sec. VI-H | (Boche et al., 2025, Sec. VI-H) |
| 運算硬體 | NVIDIA Orin NX (on a drone) | 執行運算平台 | 未標示 | system including networks and depth or LiDAR integration can run with parameter trade-offs (supplementary video; not quantified) | (Boche et al., 2025, Sec. VI-H) |
| 運算硬體 | RTX-3080 10 GB | 執行運算平台 | 未標示 | desktop GPU running the stereo and MVS networks and depth fusion; Ours-vid uses 3.51 GB GPU memory for the networks | (Boche et al., 2025, Sec. VI-H; Fig. 12 note) |
| 其他 | EuRoC Vicon-room reference point clouds | 參考或真值量測 | EuRoC MAV | mm-level accurate point clouds used for mesh accuracy and completeness | (Boche et al., 2025, Sec. VI-B) |
| 其他 | Hilti-Oxford ground truth (sparse control positions and mm-accurate dense point clouds) | 參考或真值量測 | Hilti-Oxford | sparse ground-truth positions for scoring; dense point clouds for exp04 to exp06 | (Boche et al., 2025, Sec. VI-C) |
論文圖片
只收錄原文以開放授權（open license）釋出的圖片，並依授權條件標示出處、圖號、授權與修改方式。
Fig. 1VBR Spagna 序列的三維重建：上為 LiDAR 版、下為深度網路版，黑線為估計軌跡，不同顏色代表不同子地圖
出處：Boche et al., 2025，Fig. 1。授權：CC BY 4.0 (arXiv v1)。原始圖檔。修改：縮小至寬度不超過 1400 px，並轉存為 WebP 格式。
Fig. 9由體積子地圖產生的三維重建，左為學習式深度、右為 LiDAR，上排 Hilti exp06、下排 VBR Ciampino0
出處：Boche et al., 2025，Fig. 9。授權：CC BY 4.0 (arXiv v1)。原始圖檔。修改：縮小至寬度不超過 1400 px，並轉存為 WebP 格式。
Fig. 7EuRoC V1_02 重建網格以網格到點誤差著色，比較 OKVIS2-X 深度版與 SimpleMapping
出處：Boche et al., 2025，Fig. 7。授權：CC BY 4.0 (arXiv v1)。原始圖檔。修改：轉存為 WebP 格式。
Fig. 10VBR Campus1 所有子地圖疊合的重建與加入 GNSS 的即時軌跡，標示 GNSS 可用與中斷區段
出處：Boche et al., 2025，Fig. 10。授權：CC BY 4.0 (arXiv v1)。原始圖檔。修改：縮小至寬度不超過 1400 px，並轉存為 WebP 格式。
作者報告的優勢與限制
優勢
- On VBR (1.0 to 9.0 km sequences) Ours-vil-nc average ATE RMSE 0.829 m versus 2.475 m for FAST-LIVO; visual-inertial Ours-vid-nc 2.210 m versus 5.572 m for ORB-SLAM3 and 28.770 m for OpenVINS (Table VI)
- Hilti-Oxford mesh accuracy 0.029 m and completeness 78.33 % with LiDAR, 0.042 m and 64.27 % with learned depth, at a 0.2 m threshold (Table V)
- Hilti 2022 challenge score 438.10 for Ours-vil-ba, above FAST-LIVO2 (402.93) and VILENS (325.77) but below LiDAR-inertial Wildcat (563.79) (Table IV)
- Faster than ORB-SLAM3 on EuRoC MH05 (38.1 versus 64.7 ms per frame) and on par on Hilti exp06 (Sec. VI)
- Open-source with configurable sensor setups
限制
- Dense mapping raises memory: up to 25.5 GB for Ours-vil on VBR Campus1 (Table IX)
- Higher runtime than ORB-SLAM3 on VBR Campus1 (107.3 versus 69.1 ms per frame) (Sec. VI)
- Long corridors (Hilti exp07) are degenerate for LiDAR alignment (Sec. VI)
- Learned depth networks require a GPU
營建工程相關證據
OKVIS2-X 直接輸出全域一致的稠密占據地圖與網格，並在 Hilti-Oxford 資料上以地面真值點雲評估網格精度（LiDAR 版本平均 2.9 cm，完整度 78 %）；依 Zhang et al., 2023c 的資料集說明，Hilti-Oxford 包含施工現場序列，但本文未說明 exp04 至 exp06 的場景類型。這是少數同時提供軌跡與建圖精度的多感測器系統，對施工現場掃描具參考價值（推論）。
原文驗證環境：公開基準、獨立參考量測、跨場域
報告的性能數據
性能數據仍在分批查證，目前尚未收錄此方法的報告值。
來源
Boche et al., 2025
（2025）OKVIS2-X: Open Keyframe-Based Visual-Inertial SLAM Configurable With Dense Depth or LiDAR, and GNSSIEEE Transactions on Robotics, 41, pp. 6064-6083
DOI 10.1109/tro.2025.3619051arXiv 2510.04612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：OKVIS2-X arXiv v1 (accepted version, CC BY 4.0) https://arxiv.org/abs/2510.04612
- 程式碼釋出：ethz-mrl/OKVIS2-X https://github.com/ethz-mrl/OKVIS2-X
程式碼：https://github.com/ethz-mrl/OKVIS2-X（授權：BSD-3-Clause (LICENSE file; GitHub API reports NOASSERTION)）。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。