PaperlayerView exact revision on arXiv

Source-linked AI summary

VINS-Mono: A Robust and Versatile Monocular Visual-Inertial State Estimator

Tong Qin, Peiliang Li, Shaojie Shen

arXiv:1708.03852v1cs.RO

TL;DR

Monocular visual-inertial estimation must recover metric six-DOF state despite difficult initialization and nonlinear sensor fusion. VINS-Mono combines robust initialization and recovery with tightly coupled optimization, loop-based relocalization, and global optimization, achieving 0.29% position drift in an aerial-robot experiment and eliminating drift over a 264 m mobile-device trajectory.

  • Problem

    Monocular VINS lacks directly observable scale, making initialization under motion and large IMU bias particularly fragile.

  • Method

    VINS-Mono uses tightly coupled sliding-window nonlinear optimization with online calibration, bias correction, failure recovery, feature-level relocalization, and 4-DOF pose-graph optimization.

  • Results

    0.29% position drift was achieved over a 61.97 m aerial-robot trajectory, while 4-DOF pose-graph optimization eliminated drift over a 264 m mobile-device trajectory.

  • Takeaways & Limitations

    The system provides a robust, versatile estimator demonstrated in autonomous aerial flight and mobile-device applications, with PC and iOS implementations released openly.

  • Takeaways & Limitations

    The estimator can still fail under violent illumination changes or severely aggressive motions, motivating active failure detection and recovery.

Abstract

from arXiv · show

A monocular visual-inertial system (VINS), consisting of a camera and a low-cost inertial measurement unit (IMU), forms the minimum sensor suite for metric six degrees-of-freedom (DOF) state estimation. However, the lack of direct distance measurement poses significant challenges in terms of IMU processing, estimator initialization, extrinsic calibration, and nonlinear optimization. In this work, we present VINS-Mono: a robust and versatile monocular visual-inertial state estimator.Our approach starts with a robust procedure for estimator initialization and failure recovery. A tightly-coupled, nonlinear optimization-based method is used to obtain high accuracy visual-inertial odometry by fusing pre-integrated IMU measurements and feature observations. A loop detection module, in combination with our tightly-coupled formulation, enables relocalization with minimum computation overhead.We additionally perform four degrees-of-freedom pose graph optimization to enforce global consistency. We validate the performance of our system on public datasets and real-world experiments and compare against other state-of-the-art algorithms. We also perform onboard closed-loop autonomous flight on the MAV platform and port the algorithm to an iOS-based demonstration. We highlight that the proposed work is a reliable, complete, and versatile system that is applicable for different applications that require high accuracy localization. We open source our implementations for both PCs and iOS mobile devices.

I. INTRODUCTION · II. RELATED WORK

The paper presents VINS-Mono as a robust monocular visual-inertial estimator that addresses metric-scale, initialization, calibration, and long-term drift challenges. It combines tightly coupled optimization, relocalization, four-DOF pose graph optimization, and real-time deployments across drones, large-scale localization, and mobile AR.

  • I. INTRODUCTION: Monocular vision-only systems cannot recover metric scale, motivating monocular visual-inertial sensing for real-world robotic applications.Monocular VINS uses a camera and IMU while remaining applicable to mobile robots, drones, and mobile devices.
  • I. INTRODUCTION: Monocular VINS requires acceleration excitation for metric-scale observability, complicating startup from unknown motion and nonlinear estimator initialization.Camera-IMU extrinsic calibration adds further difficulty because the system combines two sensors.
  • I. INTRODUCTION: VINS-Mono uses on-the-fly initialization for startup and failure recovery, plus tightly coupled sliding-window nonlinear optimization for monocular visual-inertial odometry.The odometry estimates local pose, velocity, and orientation.
  • I. INTRODUCTION: Its tightly coupled optimization estimates camera-IMU extrinsics and IMU biases while online loop detection supports tightly coupled relocalization.These components extend the estimator beyond local odometry.
  • I. INTRODUCTION: Four DOF global pose graph optimization enforces global consistency for monocular VINS.The system is reported as complete and applicable to small-scale AR, mediumscale drone navigation, and large-scale state estimation.
  • I. INTRODUCTION: The system demonstrates real-time performance for drone navigation, large-scale localization, and mobile AR, with open-source PC and iOS implementations.The iOS version runs on iPhone6s or above, and the PC version is integrated with ROS.
  • II. RELATED WORK: Related monocular visual-inertial methods span loosely coupled EKF fusion, tightly coupled EKF or graph optimization, and direct or indirect visual residual formulations.Direct methods minimize photometric error, whereas indirect methods minimize geometric displacement and require feature extraction and matching.
  • II. RELATED WORK: IMU pre-integration avoids repeated re-integration, while prior initialization methods can be unreliable and monocular VINS retains four-DOF drift requiring loop closure.The four drifting degrees of freedom are 3D translation and rotation around gravity, or yaw.

III. OVERVIEW … B. IMU Pre-integration

VINS-Mono processes tracked visual features and pre-integrated IMU measurements, initializes the state for nonlinear visual-inertial optimization, and uses bias-aware IMU pre-integration to reduce repeated propagation. Its front-end maintains distributed features and selects keyframes, while the estimator defines consistent world, body, and camera-frame conventions.

  • III. OVERVIEW: The system preprocesses visual and inertial measurements, initializes pose, velocity, gravity, gyroscope bias, and 3D feature locations, then bootstraps nonlinear visual-inertial optimization.The overview describes feature tracking, IMU pre-integration, and initialization as the stages preceding nonlinear VIO.
  • III. OVERVIEW: The estimator uses world, body/IMU, and camera frames, with gravity aligned to the world-frame z axis and rotations represented primarily by Hamilton quaternions.Rotation matrices are also used for convenient 3D-vector rotation.
  • IV. MEASUREMENT PREPROCESSING: Visual preprocessing tracks features between consecutive frames, detects new features, undistorts and projects observations onto a unit sphere, and rejects outliers with RANSAC.The front-end uses KLT sparse optical flow for existing features and corner detection for new ones.
  • A. Vision Processing Front-end: New corner detection maintains 100-300 features per image while enforcing uniform spatial distribution through a minimum separation between neighboring features.This feature-management step complements KLT tracking and outlier rejection.
  • A. Vision Processing Front-end: Keyframes are selected when tracked-feature parallax exceeds a threshold, while short-term observations help avoid triangulation failure during rotation-only motion.Both translation and rotation can produce parallax, but rotation alone does not permit feature triangulation.
  • B. IMU Pre-integration: This work extends prior on-manifold IMU pre-integration by incorporating IMU bias correction, while recursively propagating estimated means, covariance, and first-order Jacobians.The implementation uses midpoint integration, although Euler integration is used to explain the procedure, and treats unknown additive noise as zero.
  • B. IMU Pre-integration: The IMU model accounts for acceleration bias, gyroscope bias, and additive Gaussian noise, with both biases modeled as random walks.The measurements combine gravity-countering force with platform dynamics and are measured in the body frame.
  • B. IMU Pre-integration: Pre-integration uses the local body frame as reference so terms depend only on IMU measurements and biases, avoiding repeated propagation when optimization updates the starting states.Small bias changes are handled with first-order corrections; larger changes trigger re-propagation.

V. ESTIMATOR INITIALIZATION

VINS-Mono uses a loosely coupled vision–IMU initialization procedure because monocular scale is unobservable and initialization under motion makes direct fusion difficult. It recovers rough metric quantities by aligning IMU pre-integration with visual-only SfM, while initially ignoring accelerometer bias because it is difficult to observe.

  • Initialization challenges: Monocular visual-inertial initialization is difficult because camera scale is unobservable and stationary-start assumptions fail under motion.The system is highly nonlinear, and IMU measurement complications further increase initialization difficulty.
  • Initialization procedure: The method first uses loosely coupled vision–IMU fusion, leveraging visual-only SLAM or SfM to bootstrap relative-motion initial values.Visual initialization can use eight-point, five-point, or homography-based methods.
  • Initialization procedure: Aligning metric IMU pre-integration with visual-only SfM roughly recovers scale, gravity, and velocity for subsequent estimation.This alignment provides initial metric quantities before tightly coupled optimization.
  • Bias handling: The initial step ignores accelerometer bias because its coupling with gravity and the short initialization duration make it difficult to observe.Accelerometer-bias calibration is deferred to a more detailed treatment in prior work.

A. Sliding Window Vision-Only SfM

The initialization begins with vision-only structure-from-motion in a sliding window, estimating up-to-scale camera poses and feature positions before aligning the visual structure to metric scale.

  • Vision-only SfM estimates up-to-scale camera poses and feature positions as the first step of initialization.
  • A sliding window bounds computational complexity while evaluating feature correspondences between the latest frame and previous frames.
  • More than 30 tracked features and more than 20 rotation-compensated pixels indicate stable tracking and sufficient parallax for relative pose recovery.
  • Rough camera–IMU extrinsic parameters allow camera poses to be translated into the IMU body frame.
  • The scaling parameter aligns visual structure with metric scale and is key to successful initialization.

B. Visual-Inertial Alignment · 1) Gyroscope Bias Calibration: · 2) Velocity, Gravity Vector and Metric Scale Initialization:

Visual-inertial alignment first calibrates gyroscope bias from consecutive-frame visual and IMU rotation constraints, then initializes velocity, gravity, and metric scale. Gravity is represented with a 2-DOF parameterization, and a linear least-squares problem recovers the navigation quantities from up-to-scale monocular SfM.

  • 1) Gyroscope Bias Calibration:: The IMU pre-integration term is linearized with respect to gyroscope bias, and a window-wide cost function is minimized.B indexes all frames in the optimization window.
  • 1) Gyroscope Bias Calibration:: A first-order bias approximation yields an initial gyroscope-bias calibration, after which all IMU pre-integration terms are re-propagated.Re-propagation uses the newly estimated gyroscope bias.
  • 2) Velocity, Gravity Vector and Metric Scale Initialization:: After gyroscope-bias initialization, the method initializes velocity, the gravity vector, and metric scale as essential navigation states.The scale parameter converts monocular SfM from up-to-scale estimates to metric units.
  • 2) Velocity, Gravity Vector and Metric Scale Initialization:: Because gravity has known magnitude g ≈9.81m/s2, it is parameterized with 2 DOF on the tangent space of a sphere.The parameterization uses two orthogonal basis vectors, b1 and b2, around the current gravity estimate.
  • 2) Velocity, Gravity Vector and Metric Scale Initialization:: The method combines consecutive-frame constraints into a linear measurement model using visual SfM poses and frame time intervals.The visual quantities are obtained up to scale, while ∆tk denotes the interval between consecutive frames.
  • 2) Velocity, Gravity Vector and Metric Scale Initialization:: Solving the linear least-squares problem recovers body-frame velocities, gravity in the visual reference frame, and the scale parameter.These quantities are obtained for every frame in the window.

3) Gravity Refinement: … A. Formulation

The method refines gravity under a known-magnitude constraint, completes metric initialization, and then formulates tightly coupled monocular VIO as a sliding-window visual-inertial bundle adjustment. The nonlinear estimator combines state, feature, prior, IMU, and visual terms for maximum-posteriori state estimation.

  • 3) Gravity Refinement:: Gravity refinement reduces the gravity direction to 2 DOF by constraining its known magnitude and re-parameterizing it on the tangent space.The gravity vector is represented using a fixed-magnitude direction plus two tangent-space displacement variables.
  • 3) Gravity Refinement:: The tangent-space basis vectors b1 and b2 are obtained with cross-product operations, while their displacements are jointly optimized with other state variables until ĝ converges.The procedure substitutes the refined gravity parameterization into the initialization equations and iterates on w1 and w2.
  • 4) Completing Initialization:: After gravity refinement, the world-to-camera rotation is obtained by rotating gravity to the z-axis, and all initialized variables are transformed into the world frame.Body velocities are also rotated into the world frame.
  • 4) Completing Initialization:: Visual structure is scaled to metric units, completing initialization before the resulting metric values are supplied to tightly coupled monocular VIO.The completed initialization provides the inputs for the subsequent tightly coupled estimator.
  • VI. TIGHTLY-COUPLED MONOCULAR VIO: The tightly coupled monocular VIO performs high-accuracy, robust state estimation using a sliding-window formulation after estimator initialization.The sliding window formulation is illustrated in Fig. 3.
  • A. Formulation: Each sliding-window IMU state contains position, velocity, orientation, acceleration bias, and gyroscope bias, while each feature is represented by inverse depth from its first observation.The window also tracks keyframes and features, with λ_l denoting the inverse depth of feature l.
  • A. Formulation: The estimator uses visual-inertial bundle adjustment to minimize prior and Mahalanobis-norm measurement residuals for maximum-posteriori estimation.Residuals cover IMU and visual measurements; marginalization supplies prior information, and Ceres Solver solves the nonlinear problem.

B. IMU Measurement Residual · C. Visual Measurement Residual

The estimator defines pre-integrated IMU residuals between consecutive frames, incorporating bias correction, and models visual residuals on a unit sphere with tangent-plane projection. This formulation supports varied camera optics while representing the visual residual’s two degrees of freedom.

  • B. IMU Measurement Residual: The IMU residual is defined from measurements between two consecutive frames in the sliding window using the stated IMU measurement model.The residual is built for pre-integrated IMU measurements associated with consecutive frames b_k and b_{k+1}.
  • B. IMU Measurement Residual: Quaternion vector-part extraction and a three-dimensional error-state term are used in representing the IMU residual errors.The supplied notation identifies xyz as quaternion vector-part extraction and δθ_bk+1 as a three-dimensional error term.
  • B. IMU Measurement Residual: Pre-integrated accelerometer and gyroscope measurements are used within each inter-frame interval, with both sensor biases included for online correction.The pre-integrated terms use noisy accelerometer and gyroscope measurements, while accelerometer and gyroscope biases appear in the residual terms.
  • C. Visual Measurement Residual: Unlike generalized-image-plane reprojection errors, the visual measurement residual is defined on a unit sphere.A camera observation is represented as a unit ray connecting the surface of the unit sphere.
  • C. Visual Measurement Residual: The residual for a feature observed first in image i is defined for its observation in image j.The formulation distinguishes the feature’s first observation from its subsequent observation.
  • C. Visual Measurement Residual: The unit-sphere formulation is illustrated separately from the marginalization strategy, which converts marginalized measurements into a prior or removes frame-associated visual measurements.For non-keyframes, pre-integrated inertial measurements are retained even when the frame and visual measurements are removed.
  • C. Visual Measurement Residual: Pixel locations are back-projected into unit vectors using camera intrinsics, and the two-dimensional visual residual is projected onto a tangent plane.Two orthogonal bases, b1 and b2, span the tangent plane used for this projection.
  • C. Visual Measurement Residual: The visual residual’s covariance is represented as the standard covariance of a fixed length in the tangent space.The supplied passage identifies this covariance representation for the tangent-space formulation.

D. Marginalization … G. Failure Detection and Recovery

The system controls optimization cost through selective marginalization and lightweight state-estimation modes, while detecting failures and restarting initialization with a new pose-graph segment. These mechanisms support computational efficiency, high-rate outputs, and recovery from challenging conditions.

  • D. Marginalization: Selective marginalization removes IMU states and features from the sliding window while converting their measurements into a prior to bound optimization complexity.The scheme uses the Schur complement to construct and add the new prior to the existing prior.
  • D. Marginalization: Keyframe-aware marginalization retains spatially separated keyframes for sufficient parallax and preserves excitation-rich accelerometer measurements while maintaining system sparsity.For non-keyframes, visual measurements are discarded but connecting IMU measurements are retained.
  • D. Marginalization: Marginalization fixes early linearization points and can produce suboptimal estimates, but its negative impact is considered noncritical because small VIO drift is acceptable.This limitation is explicitly acknowledged as a tradeoff of the marginalization procedure.
  • E. Motion-only Visual-Inertial Bundle Adjustment for Camera-Rate State Estimation: 30 Hz camera-rate outputs are obtained on computationally constrained devices through lightweight motion-only visual-inertial bundle adjustment.The method optimizes poses and velocities of a fixed number of latest IMU states while treating feature depth, extrinsics, biases, and older states as constants, using all visual and inertial measurements.
  • F. IMU Forward Propagation for IMU-Rate State Estimation: Latest VIO estimates are directly propagated with recent IMU measurements to provide IMU-rate state estimates for closed-loop state feedback.IMU measurements arrive at a much higher rate than visual measurements, enabling this propagation between image updates.
  • G. Failure Detection and Recovery: Failure detection monitors unusual estimator outputs because violent illumination changes and severely aggressive motions can still cause VIO failure.The detector is an independent module within the system.
  • G. Failure Detection and Recovery: Failures are flagged by too few tracked features, large position or rotation discontinuities, or large changes in estimated biases or extrinsic parameters.These criteria cover feature availability, output continuity, and parameter-estimation changes.
  • G. Failure Detection and Recovery: After failure detection, the system returns to initialization and creates a new, separate pose-graph segment after successful monocular VIO initialization.Recovery therefore restarts estimation rather than continuing the failed segment.

VII. RELOCALIZATION … A. Adding Keyframes into the Pose Graph

VINS-Mono addresses accumulated position and yaw drift through tightly coupled loop detection, feature-based relocalization, and globally consistent four-degree-of-freedom pose-graph optimization. Keyframes marginalized from the sliding window are incorporated into the pose graph as vertices.

  • VII. RELOCALIZATION: The relocalization module eliminates accumulated drift in global position and yaw by integrating loop detection with monocular visual-inertial odometry.The sliding-window and marginalization scheme bounds computation complexity but introduces drift in x, y, z, and gravity-direction yaw.
  • A. Loop Detection: DBoW2 detects revisited places using BRIEF-described corners and returns loop-closure candidates after temporal and geometrical consistency checks.In addition to VIO corner features, 500 additional corners are detected and described to improve loop-detection recall.
  • A. Loop Detection: Because monocular VIO makes roll and pitch observable, loop detection does not require rotation-invariant features such as ORB.This contrasts with reliance on rotation-invariant features in ORB SLAM.
  • B. Feature Retrieval: Feature retrieval establishes loop connections through BRIEF matching followed by 2D-2D fundamental-matrix and 3D-2D PnP RANSAC outlier rejection.A candidate is accepted as a correct loop detection when its inlier count exceeds a specified threshold.
  • C. Tightly-Coupled Relocalization: Tightly coupled relocalization aligns the current sliding window with past poses by jointly optimizing IMU, local visual, and retrieved loop correspondences.Loop-closure-frame poses remain constant, so the optimized state dimension is unchanged; multiple closures provide multi-view constraints for higher accuracy and smoother results.
  • VIII. GLOBAL POSE GRAPH OPTIMIZATION: Pose-graph optimization registers past poses into a globally consistent configuration after relocalization, estimating drift only in x, y, z, and yaw.Roll and pitch are ignored because the visual-inertial setup makes them drift-free and fully observable.
  • A. Adding Keyframes into the Pose Graph: Each keyframe marginalized from the sliding window is added to the pose graph as a vertex connected to other vertices by two edge types.

1) Sequential Edge: … A. Dataset Comparison

The system builds a 4-DOF pose graph from VIO-based sequential edges and relocalization-based loop closures, then optimizes and manages it for scalable operation. Experiments on EuRoC datasets show strong performance, including accurate loop-closure results, rapid initialization, and system-level advantages over OKVIS.

  • 1) Sequential Edge:: Sequential edges connect each newly marginalized keyframe to previous keyframes using relative position and yaw angle taken directly from VIO.Each edge represents a relative transformation within the local sliding window.
  • 2) Loop Closure Edge:: Loop closure edges connect keyframes with detected loop connections using 4-DOF relative pose transforms obtained from relocalization.The loop closure transform follows the same definition as the pose-graph relative pose.
  • B. 4-DOF Pose Graph Optimization: The pose graph jointly optimizes sequential and loop closure edges while applying a Huber norm only to loop closures to reduce wrong-loop effects.Sequential edges use no robust norm because VIO already provides outlier rejection.
  • B. 4-DOF Pose Graph Optimization: Asynchronous pose-graph optimization and relocalization threads enable relocalization with the latest optimized graph or the existing configuration.Relocalization does not need to wait for the current graph optimization to finish.
  • C. Pose Graph Management: A downsampling process limits pose-graph growth by retaining loop-constrained keyframes and removing redundant nearby or similarly oriented neighbors.The probability of removal is proportional to spatial distance according to the provided passage.
  • IX. EXPERIMENTAL RESULTS: The evaluation includes public-dataset comparison, numerical accuracy analysis, repetitive indoor scenes, and a large-scale experiment for long-time practicability.The paper reports three experiments and two applications for evaluating VINS-Mono.
  • A. Dataset Comparison: EuRoC evaluation uses only the left camera from stereo sequences with synchronized 200 Hz IMU measurements and ground-truth states, despite large bias and illumination changes.The images are WVGA monochrome frames captured at 20 FPS, and the datasets include VICON and Leica MS50 ground truth.
  • A. Dataset Comparison: Against OKVIS on MH_03_median and MH_05_difficult, VINS-Mono with loop closure has the smallest translation error and performs well even on the challenging V1_03_difficult sequence.VINS-Mono initializes quickly in V1_03_difficult, while pure-VIO accuracy is similar between VINS-Mono and OKVIS and system-level performance favors VINS-Mono.

B. Indoor Experiment … 2) Go around campus:

VINS-Mono is evaluated in indoor, mixed indoor–outdoor, and large-scale campus experiments using handheld camera–IMU sensing. Relocalization and loop closure substantially reduce drift and preserve accurate, stable trajectories in challenging and long-duration environments.

  • B. Indoor Experiment: The indoor experiment uses a handheld monocular camera at 20Hz and IMU at 100Hz in a laboratory with pedestrians, low light, texture-less areas, glass, and reflections.These conditions test robustness under common visual and sensing challenges.
  • B. Indoor Experiment: During indoor circling, OKVIS and VINS-Mono without loop closure accumulate significant drift in x, y, z, and yaw, whereas relocalization and loop closure eliminate these drifts.The comparison is shown between OKVIS, VINS-Mono without loop closure, and VINS-Mono with relocalization and loop closure.
  • 1) Go out of the lab:: The mixed indoor–outdoor experiment follows a more-than-700-meter route lasting approximately ten minutes through the laboratory, stairs, playground, and building return path.The route begins and ends at the same laboratory seat.
  • 1) Go out of the lab:: VINS-Mono maintains a clear staircase trajectory where OKVIS suffers unstable feature tracking and bad estimation during the ascent.The loop-closure trajectory is aligned with Google Map to verify its accuracy.
  • 1) Go out of the lab:: 13.80, -5.26, 7.23m is OKVIS’s final drift, compared with 0.88% for VINS-Mono without loop closure and [-0.032, 0.09, -0.07]m after loop correction.The corresponding VINS-Mono drift without loop closure is smaller than OKVIS’s 2.36%.
  • 2) Go around campus:: The campus dataset covers a 5.62km path across approximately 710m in length, 240m in width, and 60m in height changes, with 25Hz images and 200Hz IMU data lasting 1 hour and 34 minutes.It is described as a large-scale test of VINS-Mono’s stability and durability.
  • 2) Go around campus:: Using a 2000-keyframe database on an Intel i7-4790 CPU at 3.60GHz, VINS-Mono produces an almost drift-free trajectory compared with Google Map.The configuration is intended to provide sufficient loop information while maintaining real-time performance.

D. Application I: Feedback Control of an Aerial Robot · E. Application II: Mobile Device · X. CONCLUSION AND FUTURE WORK

VINS-Mono demonstrates accurate, robust operation in autonomous aerial control and mobile augmented-reality applications, while the conclusion identifies observability, calibration, and dense mapping as future research directions.

  • D. Application I: Feedback Control of an Aerial Robot: The aerial platform used a 752×480 forward-looking global-shutter camera with a 190-degree fisheye lens, DJI A3 flight controller, and Intel i7-5500U onboard CPU.The traditional pinhole camera model was unsuitable for the large field of view.
  • D. Application I: Feedback Control of an Aerial Robot: VINS-Mono provides real-time position feedback for an aerial robot tracking a figure-eight trajectory four times with loop closure disabled.The experiment used a 1.0-meter-radius figure-eight pattern and four surrounding obstacles.
  • D. Application I: Feedback Control of an Aerial Robot: 0.29% position drift was achieved over a 61.97 m aerial trajectory, with final drift [0.08, 0.09, 0.13] m against OptiTrack ground truth.The reported translation and rotation errors are shown in Fig. 23.
  • E. Application II: Mobile Device: VINS-Mono was ported to iPhone 7 Plus as VINS-Mobile and compared against the Google Tango device in a mobile augmented-reality application.The iPhone used 30 Hz, 640 × 480 images and 100 Hz IMU data from its built-in six-axis gyroscope and accelerometer.
  • E. Application II: Mobile Device: 4-DOF pose graph optimization eliminates x, y, z, and yaw drift when loops are detected during indoor and outdoor motion.The method was applied while walking between two devices at a normal pace.
  • E. Application II: Mobile Device: After traveling about 264 m, VINS-Mono returned to the start point while Tango drifted during the final lap, with total trajectory drift eliminated by pose graph optimization.VINS-Mono remained operational despite a challenging moving-door case involving Tango yaw jumps and potential estimator failures.
  • E. Application II: Mobile Device: The mobile experiment shows that the proposed method runs on general mobile devices and demonstrates robustness, although Tango is more accurate for local state estimates.The authors describe the application as having potential for comparison with specially engineered devices.
  • X. CONCLUSION AND FUTURE WORK: The conclusion presents VINS-Mono as a robust and versatile monocular visual-inertial estimator with open-source implementations and solutions spanning initialization, calibration, odometry, relocalization, and global optimization.Future work targets online observability evaluation and motion planning, broad consumer-device calibration, calibration-quality identification, and dense mapping.
Loading 1708.03852v1…