GNSS-Visual-Inertial Odometry (GVIO) Overview

• GVIO is a sensor fusion framework that integrates high-frequency visual-inertial odometry with drift-free GNSS measurements to achieve globally referenced, drift-suppressed state estimates.
• It employs diverse architectures—including loosely coupled, tightly coupled, and multi-layer formulations—that balance computational efficiency with enhanced estimation accuracy.
• GVIO systems demonstrate significant performance gains, with studies reporting up to 50% reduction in position error and sub-decimeter accuracy in urban, multi-agent, and dynamic environments.

Searching arXiv for recent and foundational GVIO papers to ground the article in the literature. GNSS-Visual-Inertial Odometry (GVIO) denotes a class of state-estimation systems that fuse Global Navigation Satellite System (GNSS) measurements with visual and inertial data to obtain globally referenced, drift-suppressed motion estimates. In the literature, GVIO appears in loosely coupled, tightly coupled, multi-layer, filter-based, and optimization-based forms, but the common objective is consistent: to combine the high-rate local motion tracking of visual-inertial odometry (VIO) with the drift-free but noisy or intermittent global information supplied by GNSS, including global positions, GNSS velocity, raw pseudorange and Doppler, and, in high-precision variants, carrier phase (Cao et al., 2021, Dong et al., 2023).

1. Problem formulation and scope

GVIO is motivated by a characteristic complementarity between sensing modalities. Pure VIO is high-frequency and locally smooth, but it drifts in position and yaw over long trajectories; standalone GNSS is drift-free in a global frame, but its measurements are noisy, can be intermittent, and degrade severely in urban canyons or under non-line-of-sight (NLOS) conditions. The literature therefore treats GVIO as a global-local fusion problem in which camera and IMU measurements stabilize short-term motion estimation, while GNSS constrains long-term drift and supplies a global reference frame (Cioffi et al., 2020, Song et al., 2023).

The global frame varies with formulation. Some systems estimate directly in Earth-centered Earth-fixed (ECEF), as in carrier-phase fusion, where the IMU state is expressed in the ECEF frame and the goal is robust global positioning in ECEF (Dong et al., 2023). Others use East-North-Up (ENU) or a local world frame aligned to gravity, with an additional alignment variable such as a yaw offset between the VIO world and ENU (Cao et al., 2021). This variation is not merely notational: it determines how GNSS measurements are modeled, how lever arms are compensated, and which calibration parameters are observable.

The measurement regime also varies substantially. Conventional fusion methods use low-accuracy pseudorange based GNSS measurements with $>>5m$ errors and therefore provide only a coarse registration to the global frame (Dong et al., 2023). By contrast, tightly coupled raw-GNSS methods incorporate pseudorange and Doppler directly into a joint estimator (Liu et al., 2020, Cao et al., 2021), while high-precision systems leverage double-differenced carrier phase and integer ambiguity resolution to attain centimeter-level accuracy when ambiguity fixing succeeds (Dong et al., 2023). A related line of work fuses only global positional measurements, rather than raw GNSS observables, into optimization-based VIO (Cioffi et al., 2020).

2. Estimation architectures

GVIO systems divide broadly into loosely coupled, tightly coupled, multi-layer, and decentralized formulations. In a loosely coupled design, the visual-inertial subsystem first produces a local odometry estimate, and GNSS is fused with that estimate at a higher level. The carrier-phase GVIO system of 2023 is loosely coupled at the system level: an EKF runs at the GNSS rate, consumes incremental odometry from an IMU+camera SLAM pipeline, and augments the state with double-differenced carrier-phase integer ambiguities $N$ (Dong et al., 2023). In that formulation, the EKF state is

$x = [R_I^E,\; p^E,\; v^E,\; b_g,\; b_a,\; N]^T,$

with orientation, position, velocity, IMU biases, and ambiguity states defined in ECEF.

Tightly coupled systems move the GNSS constraints into the same estimator that handles visual and inertial data, so that correlations among measurements are explicitly exploited. In optimization-based formulations, the estimator typically uses a fixed-lag sliding window of keyframes, landmarks, IMU preintegration factors, GNSS factors, and a marginalized prior. A representative cost is

$$J(X) = \sum r_{vis}^T W_{vis} r_{vis} + \sum r_{imu}^T W_{imu} r_{imu} + \sum r_{gps}^T W_{gps} r_{gps} + r_{prior}^T W_{prior} r_{prior},$$

which is the structure used for tightly coupled fusion of global positional measurements in optimization-based VIO (Cioffi et al., 2020). GVINS extends this pattern to raw GNSS code pseudorange and Doppler shift measurements in a factor graph, together with visual and inertial information, and maintains GNSS receiver clock bias for each constellation as part of the keyframe state (Cao et al., 2021).

Filter-based tightly coupled variants occupy a different point in the design space. Open-VINS-based GPS-VIO fusion uses an MSCKF-style EKF with cloned camera states, plus explicit yaw and translation variables relating the VIO frame to ENU (Song et al., 2023). InGVIO uses an invariant filter with raw GNSS pseudorange and Doppler, cloned camera poses, and GNSS clock states, and emphasizes intrinsic consistency and strictly linear error propagation in the chosen invariant error coordinates (Liu et al., 2022). Sky-GVIO retains an MSCKF core but augments GNSS/INS/vision fusion with an FCN-based sky segmentation module for NLOS handling (Wang et al., 2024). The 2025 multi-view pose-only formulation keeps the state dimension fixed by removing landmarks from the filter state and replacing them with a multi-view pose-only visual measurement model (Hua et al., 7 Aug 2025).

Multi-layer systems separate local motion estimation from long-term global correction. In the multi-layer VI-GNSS framework, an inner-layer tightly couples camera, IMU preintegration, and GNSS velocity to estimate local motion, while an outer-layer loosely couples local motion with GNSS position and course in a 4-DoF pose graph over a longer time scale (Han et al., 2022). Decentralized GVIO generalizes the problem to multiple agents: D-GVIO assigns each robot a local buffer-driven filter, performs private GNSS and visual updates locally, and uses covariance intersection for consistent peer-to-peer collaborative updates without explicit cross-covariances (Luo et al., 2 Mar 2026). This suggests that “GVIO” is best understood as a family of fusion strategies rather than a single canonical estimator.

3. Measurement models and factor construction

The inertial component is structurally consistent across the literature. IMU measurements drive state propagation by integrating angular velocity and specific force, with gravity, gyroscope bias, and accelerometer bias included in the process model. Optimization-based approaches commonly use IMU preintegration between keyframes (Cioffi et al., 2020, Han et al., 2022), whereas several filter-based methods propagate the nominal state directly at IMU rate and update cloned camera states as images arrive (Song et al., 2023, Wang et al., 2024).

The visual component is usually a reprojection model. In optimization-based GVIO, a landmark observed in keyframe $k$ contributes a residual of the form

$$r_{vis}^{j,k} = z^{j,k} - \pi\!\left(R_C^W(q_k)(\ell_j^W - p_k^W)\right),$$

weighted by the inverse image-noise covariance (Cioffi et al., 2020). Carrier-phase EKF fusion also uses the standard reprojection linearization

$$r = z - \pi(\cdot) \approx H_x \delta x + H_p \delta p_i + \eta_v,$$

stacked across visual observations (Dong et al., 2023). MSCKF-derived systems instead exploit cloned camera poses and nullspace projection, so that landmarks need not remain in the state (Liu et al., 2022, Wang et al., 2024). The multi-view pose-only formulation goes further by expressing a feature’s 3D position directly as a function of multiple camera poses and observations, then using a pose-only residual with a perfect null space independent of estimated poses (Hua et al., 7 Aug 2025).

The GNSS component is the main axis of variation. Some systems use global-position factors: at time $t_j$, a GNSS-type measurement of a point rigidly fixed to the body becomes a residual after lever-arm removal and IMU-based propagation to the measurement time, allowing asynchronous global updates without full re-propagation (Cioffi et al., 2020). Some systems use GNSS velocity rather than position in the inner layer, specifically to constrain scale and yaw and to penalize scale drift (Han et al., 2022). Raw-GNSS systems formulate pseudorange and Doppler residuals directly against the platform position, velocity, clock bias, and clock drift, often accounting for image-GNSS asynchronism via IMU interpolation at GNSS timestamps (Liu et al., 2020, Cao et al., 2021). InGVIO uses raw pseudorange and Doppler, then single-differences between satellites to eliminate receiver clock bias and frequency bias from the raw measurement model (Liu et al., 2022).

High-precision GVIO replaces code-only GNSS with carrier phase. In the carrier-phase EKF, double-differenced carrier phase and pseudorange are formed over a reference satellite and reference station,

$$\phi_j = h_{\phi,j}(x) + \epsilon_{\phi,j}, \qquad \rho_j = h_{\rho,j}(x) + \epsilon_{\rho,j},$$

with the ambiguity vector $N \in \mathbb{Z}^M$ treated temporarily as a real “float” state before integer fixing (Dong et al., 2023). Sky-GVIO spans both single point positioning (SPP) and real-time kinematic (RTK) modes: in SPP mode the GNSS update uses pseudorange directly, whereas in RTK mode it forms double-differenced pseudorange and carrier phase and includes ambiguity states in the filter (Wang et al., 2024).

Some GVIO systems extend the measurement set beyond ego-motion. DynaVIG fuses monocular vision, INS, GNSS, and object-dynamics priors in a single sliding-window factor graph, where the state includes a global monocular scale, dynamic-object poses, object velocities, and object points. The GNSS factor is loosely coupled and provides ENU-frame positions, while object reprojection and random-constant-velocity priors are optimized jointly with navigation states (Jin et al., 2022).

4. Initialization, frame alignment, and calibration

Initialization is a central problem because GNSS and VIO generally begin in different reference frames and often operate at different rates. A recurring design is coarse-to-fine alignment. GVINS first runs vision-only structure-from-motion and inertial initialization to recover scale, gravity, and IMU biases, then performs coarse anchor localization with single-epoch SPP, yaw calibration using Doppler over a short window, and anchor refinement by minimizing pseudorange and clock factors while holding VIO poses fixed (Cao et al., 2021). The optimization-based GPS-VIO system of 2022 similarly performs fast initialization by collecting raw GPS and stereo VIO positions during VIO startup and solving for ENU-to-VIO alignment via Umeyama’s closed form, which the paper states saves up to $N$0 s versus sequential initialization (Han et al., 2022).

Calibration encompasses extrinsics, time offsets, and frame-rotation parameters. The 2022 optimization-based GPS-VIO system includes GPS–IMU extrinsic transform and time offset in the state vector and lets the GPS residual drive them to their optimal values, stabilized by weak priors on the calibration parameters (Han et al., 2022). The carrier-phase EKF paper addresses GNSS-antenna-to-IMU extrinsics through an explicit batch problem over a short initialization segment: $N$1 solved in two passes and repeated when the current segment has sufficient excitation and an RMSE $N$2 m (Dong et al., 2023). The multi-layer VI-GNSS framework also treats the GNSS-to-IMU lever arm and a global yaw correction quaternion as outer-layer variables, while its initialization includes a MAP refinement over initial frames (Han et al., 2022).

A common misconception in earlier discussions of GNSS-VIO alignment is that rotational extrinsics between the GNSS ENU frame and the VIO frame are unobservable. Song et al. show that the yaw between $N$3 and $N$4 is observable non-linearly and can be estimated online within a filter, whereas the translation extrinsic is unobservable in their formulation (Song et al., 2023). Their observability analysis uses Lie derivatives, leading to the conclusion that horizontal motion excites yaw calibration, while stationary intervals stall it. This result sharpens the distinction between what can be initialized once and what can be refined continuously.

5. Robustness, degeneracy, and adverse environments

Robustness is the dominant practical concern in GVIO because the environments that most require global localization are often those that most damage GNSS. The literature repeatedly emphasizes urban canyons, intermittent GNSS, NLOS propagation, tunnels, stop lines, low-speed segments, dynamic scenes, and long GNSS outages (Liu et al., 2020, Cao et al., 2021, Wang et al., 2024). The corresponding mitigation strategies span residual gating, outlier pruning, explicit NLOS detection, degeneracy analysis, and estimator design for consistency.

Raw-GNSS sliding-window fusion handles measurement corruption partly through preprocessing and robust costs. One optimization-based VI-SLAM system with raw pseudorange and Doppler removes GNSS measurements whose pseudorange residual exceeds $N$5 m or Doppler residual exceeds $N$6 m/s after a one-epoch GNSS-only check, de-activates GNSS factors when vehicle speed is $N$7 m/s to avoid ill-conditioning at stop lines, and wraps visual and GNSS residuals in Huber or Cauchy losses (Liu et al., 2020). The carrier-phase EKF prunes obviously non-LOS satellites by comparing GNSS innovations against a threshold derived from VIO uncertainty and may prune satellites with the largest residuals if ambiguity validation fails (Dong et al., 2023).

Sky-GVIO introduces an explicit vision-assisted NLOS pipeline. A sky-pointing fisheye camera feeds an FCN-8s sky-segmentation network; the resulting sky mask is combined with satellite elevation and azimuth to classify each satellite as line-of-sight or NLOS, after which per-satellite pseudorange and carrier-phase covariances are inflated for NLOS measurements by the S-NDM algorithm (Wang et al., 2024). This architecture exemplifies a broader tendency in GVIO to use vision not only for motion estimation but also for GNSS quality control.

Degeneracy analysis is especially prominent in invariant-filter formulations. InGVIO defines infinitesimal symmetries and identifies unobservable subspaces tied to satellite geometry and motion. If the number of satellites is $N$8, translation degeneracy remains along the nullspace induced by the GNSS geometry, and under aligned motion and geometry an $N$9 yaw degeneracy can also appear (Liu et al., 2022). The choice of invariant error is then used to preserve consistency in these regimes. Song et al. arrive at a related but narrower conclusion: the rotational extrinsic yaw is observable under non-zero horizontal velocity, but translation remains unobservable (Song et al., 2023).

Dynamic scenes create an additional failure mode for vision. DynaVIG addresses this by separating static and object features, introducing a prior height model to initialize object pose and the monocular global scale, and adding an adaptive object-dynamics prior whose covariance is inflated via $x = [R_I^E,\; p^E,\; v^E,\; b_g,\; b_a,\; N]^T,$0 when speed or direction changes (Jin et al., 2022). This suggests that, in some applications, robustness to moving objects is inseparable from robustness of the GVIO backbone itself.

6. Accuracy, efficiency, applications, and research trajectories

Reported performance covers a wide range of sensor qualities and operating regimes. For tightly coupled global-position fusion in optimization-based VIO, the mean position error is reduced up to $x = [R_I^E,\; p^E,\; v^E,\; b_g,\; b_a,\; N]^T,$1 with respect to the loosely coupled approach in outdoor UAV flights, with negligible increase of the optimization computational cost (Cioffi et al., 2020). The 2022 optimization-based GPS-VIO system reports mean RMS ATE $x = [R_I^E,\; p^E,\; v^E,\; b_g,\; b_a,\; N]^T,$2 m on EuRoC with simulated $x = [R_I^E,\; p^E,\; v^E,\; b_g,\; b_a,\; N]^T,$3 m GPS noise, $x = [R_I^E,\; p^E,\; v^E,\; b_g,\; b_a,\; N]^T,$4 m RMSE on a $x = [R_I^E,\; p^E,\; v^E,\; b_g,\; b_a,\; N]^T,$5 km KAIST urban trajectory, $x = [R_I^E,\; p^E,\; v^E,\; b_g,\; b_a,\; N]^T,$6 m on outdoor MAV loops, and convergence of random-start extrinsic errors $x = [R_I^E,\; p^E,\; v^E,\; b_g,\; b_a,\; N]^T,$7 m) and $x = [R_I^E,\; p^E,\; v^E,\; b_g,\; b_a,\; N]^T,$8 s to hand-measured values in $x = [R_I^E,\; p^E,\; v^E,\; b_g,\; b_a,\; N]^T,$9 s of real time (Han et al., 2022).

Carrier-phase GVIO targets a higher-precision regime. In simulation, under 5-satellite exclusion, it reports RMSE $$J(X) = \sum r_{vis}^T W_{vis} r_{vis} + \sum r_{imu}^T W_{imu} r_{imu} + \sum r_{gps}^T W_{gps} r_{gps} + r_{prior}^T W_{prior} r_{prior},$$0 m versus $$J(X) = \sum r_{vis}^T W_{vis} r_{vis} + \sum r_{imu}^T W_{imu} r_{imu} + \sum r_{gps}^T W_{gps} r_{gps} + r_{prior}^T W_{prior} r_{prior},$$1 m for RTKLIB and a fixed-solution rate $$J(X) = \sum r_{vis}^T W_{vis} r_{vis} + \sum r_{imu}^T W_{imu} r_{imu} + \sum r_{gps}^T W_{gps} r_{gps} + r_{prior}^T W_{prior} r_{prior},$$2 versus $$J(X) = \sum r_{vis}^T W_{vis} r_{vis} + \sum r_{imu}^T W_{imu} r_{imu} + \sum r_{gps}^T W_{gps} r_{gps} + r_{prior}^T W_{prior} r_{prior},$$3; in real urban-canyon tests with baseline $$J(X) = \sum r_{vis}^T W_{vis} r_{vis} + \sum r_{imu}^T W_{imu} r_{imu} + \sum r_{gps}^T W_{gps} r_{gps} + r_{prior}^T W_{prior} r_{prior},$$4 km, it reports RMSE $$J(X) = \sum r_{vis}^T W_{vis} r_{vis} + \sum r_{imu}^T W_{imu} r_{imu} + \sum r_{gps}^T W_{gps} r_{gps} + r_{prior}^T W_{prior} r_{prior},$$5 m versus $$J(X) = \sum r_{vis}^T W_{vis} r_{vis} + \sum r_{imu}^T W_{imu} r_{imu} + \sum r_{gps}^T W_{gps} r_{gps} + r_{prior}^T W_{prior} r_{prior},$$6 m on UC1 and $$J(X) = \sum r_{vis}^T W_{vis} r_{vis} + \sum r_{imu}^T W_{imu} r_{imu} + \sum r_{gps}^T W_{gps} r_{gps} + r_{prior}^T W_{prior} r_{prior},$$7 m versus $$J(X) = \sum r_{vis}^T W_{vis} r_{vis} + \sum r_{imu}^T W_{imu} r_{imu} + \sum r_{gps}^T W_{gps} r_{gps} + r_{prior}^T W_{prior} r_{prior},$$8 m on UC2, again outperforming RTKLIB in both fix rate and RMSE (Dong et al., 2023). Sky-GVIO reports RMS East/North/Up errors of $$J(X) = \sum r_{vis}^T W_{vis} r_{vis} + \sum r_{imu}^T W_{imu} r_{imu} + \sum r_{gps}^T W_{gps} r_{gps} + r_{prior}^T W_{prior} r_{prior},$$9 m in SPP mode and $k$0 m in RTK mode over an $k$1 km urban-canyon drive, both better than corresponding tightly coupled GNSS/INS/vision systems without the sky-mask-based S-NDM mechanism (Wang et al., 2024).

GVIO also shows a strong efficiency-versus-accuracy design tension. InGVIO reports monocular accuracy within $k$2 of GVINS on public datasets while running at $k$3 of its per-frame time, and overall fusion remains under $k$4 ms/frame in monocular or $k$5 ms/frame in stereo (Liu et al., 2022). The multi-view pose-only IEKF-based method reports approximately $k$6 ms/frame, compared with approximately $k$7 ms/frame for traditional IEKF MSCKF and approximately $k$8 ms/frame for GVINS on the same dataset class, while keeping similar accuracy (Hua et al., 7 Aug 2025). D-GVIO extends efficiency to multi-agent operation: in a four-UAV simulation, it sent approximately $k$9 messages per UAV, used $$r_{vis}^{j,k} = z^{j,k} - \pi\!\left(R_C^W(q_k)(\ell_j^W - p_k^W)\right),$$0–$$r_{vis}^{j,k} = z^{j,k} - \pi\!\left(R_C^W(q_k)(\ell_j^W - p_k^W)\right),$$1 CPU and $$r_{vis}^{j,k} = z^{j,k} - \pi\!\left(R_C^W(q_k)(\ell_j^W - p_k^W)\right),$$2–$$r_{vis}^{j,k} = z^{j,k} - \pi\!\left(R_C^W(q_k)(\ell_j^W - p_k^W)\right),$$3 MiB RAM, compared with $$r_{vis}^{j,k} = z^{j,k} - \pi\!\left(R_C^W(q_k)(\ell_j^W - p_k^W)\right),$$4 messages and $$r_{vis}^{j,k} = z^{j,k} - \pi\!\left(R_C^W(q_k)(\ell_j^W - p_k^W)\right),$$5–$$r_{vis}^{j,k} = z^{j,k} - \pi\!\left(R_C^W(q_k)(\ell_j^W - p_k^W)\right),$$6 MiB RAM for COVINS, while reporting sub-decimeter ATE on representative cooperative trajectories (Luo et al., 2 Mar 2026).

Applications span outdoor robotics, UAV navigation, AGV navigation in dynamic scenes, autonomous driving in urban canyons, fixed-wing aircraft, augmented reality navigation, crowd sourcing high-precision map update, collaborative exploration, and search-and-rescue missions (Jin et al., 2022, Han et al., 2022, Liu et al., 2022, Luo et al., 2 Mar 2026). Several papers also highlight graceful degradation as a defining property. GVINS states that it can gain from even a single satellite while conventional GNSS algorithms need four at least (Cao et al., 2021), and InGVIO reports stable behavior under reduced satellite counts and reversion to pure invariant VIO when no satellites are available (Liu et al., 2022).

The research directions named in the literature are correspondingly diverse. These include carrier-phase integration with cycle-slip handling to tighten global accuracy further, additional aiding sensors such as magnetometer, LiDAR, and UWB, lightweight loop closure or map sharing for prolonged GNSS outages, dynamic-topology cooperation in large swarms, and online tuning to reduce the conservatism of covariance intersection (Liu et al., 2022, Song et al., 2023, Luo et al., 2 Mar 2026). Taken together, these trajectories indicate that GVIO has evolved from “GNSS-aided VIO” into a broader framework for globally consistent, multi-rate, multi-sensor state estimation under degraded sensing conditions.