GVINS｜SLAM 點雲圖鑑

GVINS tightly fuses raw GNSS pseudorange and Doppler with visual-inertial data in a sliding-window factor graph, using a coarse-to-fine online initialisation of the local-to-ENU yaw and anchor, and explicit handling of degenerate cases, to give drift-free global 6-DoF estimates across indoor-outdoor transitions.

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。

GVINS 的技術屬性
感測輸入	monocular camera (left camera of a VI-Sensor)、IMU、GNSS receiver raw code pseudorange and Doppler (GPS, GLONASS, Galileo, BeiDou)
原文測試平台	helmet-mounted (walking: sports field; indoor-outdoor route over 3 km with stairs)、urban driving (22.9 km in Hong Kong; vehicle mounting not described)、simulation (10 km trajectory in a 30 m cube)
狀態估計	tightly coupled sliding-window non-linear optimisation (window size 10) on a factor graph inherited from VINS-Mono: IMU preintegration, visual reprojection, GNSS code pseudorange and Doppler factors, receiver clock bias and drift factors; states include the yaw offset between local world and ENU frames and per-constellation clock biases; two-way marginalisation; robust norm on GNSS factors in urban driving (Secs. V, VI, VIII-B5)
資料關聯	strong corners tracked with iterative Lucas-Kanade optical flow (VINS-Mono front end); GNSS satellites filtered by elevation, health and continuous lock (Secs. V, VI-C)
時間表示	discrete keyframe states; GNSS time aligned to local time beforehand via the receiver PPS signal triggering the VI-Sensor (Secs. III-A, VIII-B)
去畸變	不適用
迴圈閉合	none (VINS-Mono and VINS-Fusion loop closure also disabled in the comparisons)
全域最佳化	none; global drift-free estimation comes from GNSS raw-measurement factors inside the sliding window rather than from loop closure
地圖表示	sparse features with inverse depth in the sliding window
先驗資訊	broadcast ephemeris; Saastamoinen tropospheric and Klobuchar ionospheric models; no base station needed for the estimator (RTK used only as ground truth) (Secs. IV-D, VIII-B)
可輸出幾何	global 6-DoF pose in ECEF and a local ENU frame, including global yaw; no dense map
計算需求	Intel i7-8700K at 3.7 GHz, 32 GB: feature detection and tracking 7.28 ms per frame and window optimisation 21.91 ms on the urban sequence, real time with a 20 Hz camera (Sec. VIII-B5)

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。

原文使用的設備
類別	型號（原文寫法）	角色	資料集	原文規格	出處
慣性量測單元（IMU）	ADIS16448	方法輸入	GVINS-Dataset	200 Hz; gyroscope noise density 7.0e-3 deg/s/sqrt(Hz); accelerometer noise density 6.6e-4 m/s^2/sqrt(Hz)	(Cao et al., 2022, Table IV)
GNSS 接收器	u-blox ZED-F9P	方法輸入	GVINS-Dataset	low-cost multi-band, multi-constellation receiver; raw measurements at 10 Hz	(Cao et al., 2022, Sec. VIII-B; Table IV)
GNSS 接收器	u-blox ZED-F9P internal RTK engine with RTCM from a nearby base station	參考或真值量測	GVINS-Dataset	RTK solution at 10 Hz; about 1 cm accuracy in open areas; fails indoors and under heavy blockage	(Cao et al., 2022, Sec. VIII-B; Table IV)
GNSS 接收器	Tallysman TW3882 antenna	方法輸入	GVINS-Dataset	GNSS antenna of the receiver	(Cao et al., 2022, Table IV)
雙目相機	VI-Sensor (Aptina MT9V034 camera sensors; left camera used)	方法輸入	GVINS-Dataset	global shutter, 752 x 480, horizontal FOV 98 deg, vertical FOV 73 deg, 20 Hz; camera and IMU synchronised by the VI-Sensor and triggered by the GNSS PPS	(Cao et al., 2022, Sec. VIII-B; Table IV)
載具平台	helmet carrying the VI-Sensor and ZED-F9P	方法輸入	GVINS-Dataset	used in the real-world experiments	(Cao et al., 2022, Fig. 10)
運算硬體	Intel i7-8700K at 3.7 GHz, 32 GB	執行運算平台	未標示	desktop PC used for all experiments	(Cao et al., 2022, Sec. VIII)

作者報告的優勢與限制

優勢

RMSE 0.806 m (sports field), 3.700 m (indoor-outdoor) and 4.508 m (22.9 km urban driving) versus 2.835 m, 6.036 m and 11.106 m for RTKLIB SPP (Table III)
Survives the whole urban sequence where VINS-Mono failed at 54% of the distance and VINS-Fusion oscillated with RMSE on the order of 10^5 m (Sec. VIII-B5)
Still benefits from fewer than four satellites, even a single satellite, compared with pure VIO (Sec. VIII-B2; Figs. 13-14)
Keeps global estimates indoors and on stairs where RTK showed errors up to 80 m during outages (Sec. VIII-B3; Figs. 1, 16)
Window optimisation 21.91 ms versus an average 404.83 ms (up to 1018.46 ms) for the VINS-Fusion pose-graph fusion (Sec. VIII-B5)

限制

Global positions carry a bias from satellite orbit error, imperfect atmospheric delay models and multipath, especially in the up direction (Secs. VIII-B1, VIII-B3, VIII-B5)
With three or fewer satellites the up-direction drift is no longer removed; with two or one satellite horizontal and yaw drift appear (Sec. VIII-B2)
Low-speed or rotation-only motion cannot constrain the yaw offset, which is fixed below 0.3 m/s (Sec. VII-B1)
Initialisation needs at least four satellites (N+3 for N constellations) and about 4 m of travel, and can wait several seconds for ephemerides (Secs. VII-A, VIII-B3)
Doppler-only operation drifts and keeps the initialisation bias (Sec. VIII-B4)

營建工程相關證據

論文未在營建工地測試；實驗包括校園運動場、含室內樓梯、橋下與樹林的 3 km 室內外路線，以及香港 22.9 km 市區行車，並以 RTK 為參考。它能在衛星時有時無的室內外轉換中持續提供全域位姿，對把相機慣性軌跡或點雲地理參考到工地座標、以及在高樓遮蔽處維持定位有參考價值；但全域位置仍有公尺級偏差，不足以直接作為測量控制（推論）。

原文驗證環境：模擬、獨立參考量測

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

Cao et al., 2022
Shaozu Cao, Xiuyuan Lu, Shaojie Shen（2022）GVINS: Tightly Coupled GNSS–Visual–Inertial Fusion for Smooth and Consistent State EstimationIEEE Transactions on Robotics, 38(4):2004-2021
DOI 10.1109/tro.2021.3133730 arXiv 2103.07899 程式碼
同儕審查已出版已讀全文近十年查證後修正

預印本：GVINS: Tightly Coupled GNSS-Visual-Inertial Fusion for Smooth and Consistent State Estimation (arXiv v1 to v3) https://arxiv.org/abs/2103.07899
程式碼釋出：HKUST-Aerial-Robotics/GVINS https://github.com/HKUST-Aerial-Robotics/GVINS
資料集：HKUST-Aerial-Robotics/GVINS-Dataset (well-synchronised camera, IMU and GNSS raw data with RTK ground truth, per the paper) https://github.com/HKUST-Aerial-Robotics/GVINS-Dataset

程式碼：https://github.com/HKUST-Aerial-Robotics/GVINS（授權：GPL-3.0 (README licence section)）。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

技術屬性

使用設備

作者報告的優勢與限制

優勢

限制

營建工程相關證據

報告的性能數據

來源

相關版本