Pith. sign in

REVIEW 3 major objections 4 minor 26 references

Observer Design for Optical Flow-Based Visual-Inertial Odometry with Almost-Global Convergence

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Optical-flow and IMU fusion can be made almost-globally stable by a two-stage observer.

desk verdict Solid observer cascade with clean proofs, but the almost-global result does not cover the optical-flow pre-observer, whose convergence is unproven. read the letter →

arxiv 2508.21163 v1 pith:WFNNVNSX submitted 2025-08-28 cs.RO cs.SYeess.SY

Tarek Bouazza , Soulaimane Berkane , Minh-Duc Hua , Tarek Hamel This is my paper · ORCID
classification cs.ROcs.SYeess.SY MSC 93B0793C1093D15
keywords opticalflowvisual-inertialodometrynonlinearobserverRiccatiattitudeestimationcomplementaryfilterpersistentexcitationunitspheregradientdescent
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes a way to do visual-inertial odometry with a monocular camera by treating optical flow only as a source of velocity direction, not as a feature tracker or map builder. The authors' central claim is that a two-stage observer can turn that direction measurement, combined with gyro and accelerometer data, into estimates of body-frame velocity, gravity direction, and attitude. The first stage is a Riccati observer proven globally exponentially stable under a persistent-excitation condition on the vehicle's motion; the second stage is a complementary filter on SO(3) proven almost-globally stable, meaning attitude is recovered for every initial condition except a measure-zero set. A gradient-descent algorithm on the unit sphere computes the velocity direction from sparse optical flow. The stability theorems treat that computed direction as exact; the iterative algorithm's own convergence is demonstrated in simulation, not in the proof.

What carries the argument

The load-bearing object is the projection output y = -π_{η_v} v̂_B = C(t)x, which converts the unit-vector optical-flow measurement η_v into a linear output for the six-dimensional error system (velocity error plus gravity error). The rotated form Cbar(t)=Σ(t)H with Σ=π_{η_I_v} makes uniform observability equivalent to a persistent-excitation condition on the inertial velocity direction, and the Riccati gain K=PC^T D then supplies global exponential stability. The second load-bearing object is the SO(3) complementary filter whose correction mixes the estimated gravity direction with an optional magnetometer direction; its Lyapunov function L(g̃)=|g|²-gᵀg̃ shows the only non-attracting equili

What would settle it

Construct a sparse optical-flow configuration, such as four landmarks, for which the cost function C on S² has a local minimum away from the true η_v, initialize the gradient descent there, and drive the vehicle along a persistently exciting trajectory. If the pre-observer converges to the wrong direction and the Riccati stage then fails to recover velocity and gravity, the weak link is confirmed. A complementary test is a trajectory with constant inertial velocity direction, violating (14), to see whether the velocity/gravity estimates still converge despite the failed excitation assumption.

Watch

Extended reading notes

Core claim

The central claim is that the six-dimensional error system formed by body-frame velocity error and body-frame gravity error becomes uniformly observable whenever the inertial-frame velocity direction is persistently exciting, and that a Riccati observer with output y = -π_{η_v} v̂_B is therefore globally exponentially stable. Because the gravity estimate z converges independently of attitude, it can serve as a time-varying reference direction for a complementary filter on SO(3): reduced attitude converges to (g,0) almost globally without a magnetometer, and full attitude to (I3,0) almost globally when the magnetometer is added. The proof converts the time-varying observability problem into a

Load-bearing premise

The stability proofs assume the gradient-descent pre-observer has already produced the exact velocity direction η_v; if that iterative algorithm stalls or converges to a wrong local minimum, the whole cascade receives a corrupted input and the guarantees no longer apply.

Editorial extensions

If this is right

  • Body-frame velocity and gravity direction can be estimated without knowing attitude, so translation estimation decouples from rotation estimation.
  • A camera-IMU pair without a magnetometer can recover pitch and roll almost globally; adding a magnetometer recovers yaw as well.
  • Position converges up to an unknown inertial offset and, without a magnetometer, an unknown yaw rotation, making the unobservable directions explicit.
  • Under persistent excitation of the velocity direction, convergence in the velocity/gravity stage is exponential, not merely asymptotic.
  • The number of gradient-descent iterations and the number of tracked features give an explicit accuracy-versus-computation trade-off for the velocity-direction stage.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A certified global solver for the S² minimization would remove the one unproved link and make the cascade's convergence guarantee unconditional.
  • The Riccati stage is measurement-agnostic: any unit-vector direction measurement derived from body-frame velocity, such as a bearing to a known beacon, could replace the optical-flow direction without changing the stability proof.
  • The persistent-excitation condition (14) could be monitored online; when it fails, freezing or reweighting the Riccati correction would keep estimates bounded until excitation resumes.
  • The explicit unobservable directions (inertial offset and yaw) suggest that loop closure or range measurements are the natural next sensors to add, rather than additional filtering of the same measurements.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes a cascaded visual-inertial odometry architecture. A Riccati observer on R^6 estimates body-frame velocity and gravity from IMU measurements and a velocity-direction measurement, with a global exponential stability (GES) claim under a persistence-of-excitation condition on the inertial-frame velocity direction. A complementary filter on SO(3) then estimates attitude from the gravity estimate and, optionally, a magnetometer, yielding an almost-globally asymptotically stable (AGAS) cascade. A separate pre-observer stage estimates the velocity direction from sparse optical flow by solving a constrained least-squares problem on S^2 via Riemannian gradient descent. The theoretical claims are established for an ideal velocity-direction input; the pre-observer is validated only in simulation.

Significance. If the stability results hold, the paper offers a clean geometric alternative to EKF-based VIO, with global convergence guarantees for velocity, gravity, and attitude. The derivation of the Riccati observer from error dynamics and the PE condition is self-contained and elegant (Lemma 2), and the paper is honest about the PE condition on the inertial-frame velocity direction. The separation of the attitude problem from the translational estimation problem is a nice structural contribution. However, the end-to-end claim of an optical-flow-based VIO architecture with almost-global convergence is currently not supported by the analysis, because the pre-observer that produces the velocity-direction input has no convergence or robustness guarantees and is not included in the stability theorems. This gap is substantial but appears addressable.

major comments (3)
  1. [Section V-B, Eqs. (26)-(28), and Theorem 1] The stability theorems are conditioned on an exact velocity-direction measurement: C(t) in (10) and the PE condition (14) use the true η_v. In the actual architecture, η_v is produced by the iterative algorithm (26)-(28), which minimizes the nonconvex cost (25) on S^2. The manuscript gives no convergence proof for this algorithm, no characterization of its critical points, no finite-N error bound, and no robustness analysis under noise or degenerate landmark geometry. If the pre-observer converges to a wrong direction, the Riccati observer sees a perturbed C(t), so the GES of x and the AGAS claim of Theorem 1 do not apply. This is the load-bearing gap between the theoretical contribution and the claimed optical-flow-based VIO architecture. The authors should either provide a rigorous analysis of the pre-observer (global convergence or a quantified error bound) and a robustness analysis o
  2. [Theorem 1 proof, Eq. (19)] The cascade argument relies on [3, Proposition 2] and [2, Theorem 2] but does not verify their hypotheses. In particular, the perturbation O(x) in (19a) is introduced with the definition of O(x) in Section II-A as merely 'bounded for bounded x and tends to 0 as x → 0'; this is likely too weak for the cited cascade theorems, which typically require a class-K or linear-growth bound on the interconnection term and the existence of a suitable Lyapunov function for the nominal AGAS subsystem. Since the AGAS conclusion is the paper's headline, the proof must show explicitly that the interconnection satisfies the required ISS property.
  3. [Section V-B, cost (25) and gradient (26)-(28)] The minimization of C on S^2 is nonconvex, and the algorithm is singular whenever |π_{b_i} η̂_v| = 0 or |s_i| = 0; the sign correction β_k in (28) also becomes undefined if β_k = 0. The paper provides no conditions on landmark geometry or motion that guarantee the global minimum is isolated and attractive, nor a proof that the Riemannian gradient descent avoids local minima. The simulations are encouraging but do not substitute for analysis, especially because the observer theorems assume the pre-observer output is the true η_v.
minor comments (4)
  1. [Section II-A] The notation O(x) is nonstandard: it is used as 'bounded for bounded x and tends to 0 as x → 0', rather than the usual 'bounded by C|x|'. Please clarify or use class-K functions, since the distinction matters in Theorem 1.
  2. [Lemma 2 proof] After applying Lemma 1, the lower bound is written as \bar W ≥ \bar μ I_3, but \bar W is a 6×6 matrix. The bound should be stated with the correct dimension (e.g., c\bar μ I_6). Also, the boundedness of the Riccati solution P(t), which is needed for the GES conclusion, is asserted without proof or a precise citation; please state the standard result.
  3. [Eq. (28)] The sign correction sign(β_k) is heuristic. Please provide a more formal explanation of the hemisphere ambiguity and discuss what happens when β_k = 0 or when the denominator in (28) vanishes.
  4. [Section VI-A] The simulation corrupts bearing vectors with Gaussian noise on the components. Please clarify whether the bearing vectors are renormalized to S^2 and whether the optical flow model (22) uses the noisy or the true bearings.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the core observer derivation is self-contained; the unproven pre-observer convergence is a missing proof, not a circular reduction.

full rationale

The derivation chain is self-contained. The Riccati observer is constructed from the body-frame kinematics (5)-(7), error dynamics (9), output equation (10) with C(t)=[π_{η_v} 0], and the correction/gain structure (11)-(12). Lemma 2 proves uniform observability via the transition-matrix change ar R(t)=I⊗R(t), the factorization ar C(t)=Σ(t)H, Kalman observability of (H,ar A), and the PE condition (14), using Lemma 1 from [9]. No parameter is fitted to reproduce the stability conclusion, and the PE condition is an input assumption rather than a consequence of the theorem. Theorem 1's AGAS proof is a cascade argument: the attitude subsystem for x=0 is analyzed with a Lyapunov function for the reduced-attitude case, and the magnetometer case imports the equilibrium characterization and instability from [17] and [24]. These are published, parameter-free theorems (though with author overlap) and do not assume the cascade result; they are independent support under the stated rules. The velocity-direction pre-observer (26)-(28) has no convergence proof—this is a genuine limitation and a correctness risk, but it is not circular because the stability theorems do not define their inputs in terms of the algorithm's output; they assume η_v as an exact measurement. Thus no step reduces, by the paper's own equations, to its own inputs. Score 1 reflects only the presence of minor non-load-bearing self-citations.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central stability claims rest on standard observer and cascade theorems (most from the authors' own prior work) and on the PE and boundedness assumptions. No new physical entities are introduced; the estimated gravity z is an auxiliary state.

free parameters (5)
  • Riccati weight S = 30*I6 in simulation
    Symmetric positive definite matrix in the Riccati equation (12). The stability theorem only requires bounded SPD; the concrete value is a tuning choice.
  • Riccati output weight D = 5*pi_eta_v in simulation
    SPD matrix in (12). Tuning choice, not fitted to output.
  • Attitude gains kz, km = kz=1, km=0.1 in simulation
    Positive constants in correction (17); stability holds for any kz>0, km>=0.
  • Gradient step size kappa = 5
    Step size in (27). No convergence analysis depends on its value; selected for simulation.
  • Iterations N = 3, 10, 20 in simulations
    Number of gradient steps per time step; accuracy/computation trade-off.
assumptions (6)
  • domain assumption Assumption 1: stationary landmarks, known camera intrinsics
    Used in Section V-A to derive the spherical optical flow model (22).
  • domain assumption Persistent excitation condition (14) on the inertial velocity direction
    Needed in Lemma 2 for uniform observability of (A,C).
  • domain assumption Boundedness of vB and omega
    Used in Lemma 2 to invoke Lemma 1 and to ensure bounded signals.
  • standard math Lemma 1 from [9] (Hamel and Samson)
    Using uniform observability from factorization C=Sigma H; accepted as a theorem from the literature.
  • standard math Cascade stability theorems [2, Theorem 2] and [3, Proposition 2]
    To conclude AGAS of the interconnection from GES of the low subsystem and AGAS/AGIS of the high subsystem.
  • standard math Attitude observer equilibrium structure from [17] and [24, Theorem 6.1]
    Used in Theorem 1 to identify stable and unstable equilibrium sets for the complementary filter.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Observer Design for Optical Flow-Based Visual-Inertial Odometry with Almost-Global Convergence." pith.science (2026). https://pith.science/paper/WFNNVNSX

@misc{pith2026250821163,
  author       = {Pith},
  title        = {Pith review of: Observer Design for Optical Flow-Based Visual-Inertial Odometry with Almost-Global Convergence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WFNNVNSX}},
  note         = {Machine review of arXiv:2508.21163}
}
abstract

This paper presents a novel cascaded observer architecture that combines optical flow and IMU measurements to perform continuous monocular visual-inertial odometry (VIO). The proposed solution estimates body-frame velocity and gravity direction simultaneously by fusing velocity direction information from optical flow measurements with gyro and accelerometer data. This fusion is achieved using a globally exponentially stable Riccati observer, which operates under persistently exciting translational motion conditions. The estimated gravity direction in the body frame is then employed, along with an optional magnetometer measurement, to design a complementary observer on $\mathbf{SO}(3)$ for attitude estimation. The resulting interconnected observer architecture is shown to be almost globally asymptotically stable. To extract the velocity direction from sparse optical flow data, a gradient descent algorithm is developed to solve a constrained minimization problem on the unit sphere. The effectiveness of the proposed algorithms is validated through simulation results.

Figures

Figures reproduced from arXiv: 2508.21163 by the authors.

Figure 1
Figure 1. Diagram of the proposed optical flow-based VIO architecture [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Estimated velocity direction with N = 3 iterations per time step using 4 optical flow measurements (red) and 8 measurements (black), N = 10 with 4 measurements (green), and N = 20 with 8 measurements (magenta), compared to the true velocity direction (blue). 0 2 4 6 8 10 12 14 16 18 20 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 N=3 + 4pts N=3 + 8pts N=10 + 4pts N=20 + 8pts 0 0.1 0.2 0.3 0.4 0.5 0 0.5 1 [PITH_FULL_IMAGE:… view at source ↗
Figure 3
Figure 3. Estimation errors 1−ηˆ⊤ηv for N = 3 iterations per time step using 4 optical flow measurements (red) and 8 measurements (black), N = 10 with 4 measurements (green), and N = 20 with 8 measurements (magenta) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Estimated and true body-frame linear velocity components, and [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Estimated and true body-frame gravity components, and observer [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Full attitude estimation error ∥I3 − RRˆ ⊤∥, and reduced attitude (gravity direction) estimation error 1 − gˆ⊤g/|g| 2 . of velocity estimation under persistently exciting motion, along with a nonlinear complementary filter on SO(3) that provides an estimate of the atti…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 24 canonical work pages

  1. [24]

    Synchronous Observer Design for Inertial Navigation Systems with Almost-Global Convergence

    P. van Goor, T. Hamel, and R. Mahony. Synchronous observer design for inertial navigation systems with almost-global convergence. arXiv preprint arXiv:2311.02234, 2023

    work page Pith review arXiv 2023

  2. [1]

    Andrade-Cetto and A

    J. Andrade-Cetto and A. Sanfeliu. The effects of partial observability in slam. In IEEE International Conference on Robotics and Automa- tion, 2004. Proceedings. ICRA’04. 2004 , volume 1, pages 397–402. IEEE, 2004

    work page 2004

  3. [2]

    D. Angeli. An almost global notion of input-to-state stability. IEEE Transactions on Automatic Control , 49(6):866–874, 2004

    work page 2004

  4. [3]

    Angeli and L

    D. Angeli and L. Praly. Stability robustness in the presence of exponentially unstable isolated equilibria. IEEE Transactions on Automatic Control, 56(7):1582–1592, 2010

    work page 2010

  5. [4]

    Barrau and S

    A. Barrau and S. Bonnabel. An ekf-slam algorithm with consistency properties. arXiv preprint arXiv:1510.06263 , 2015

    arXiv 2015

  6. [5]

    Bloesch, S

    M. Bloesch, S. Omari, M. Hutter, and R. Siegwart. Robust visual inertial odometry using a direct ekf-based approach. In 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS) , pages 298–304. IEEE, 2015

    work page 2015

  7. [6]

    Delmerico and D

    J. Delmerico and D. Scaramuzza. A benchmark comparison of monoc- ular visual-inertial odometry algorithms for flying robots. In 2018 IEEE international conference on robotics and automation (ICRA) , pages 2502–2509. IEEE, 2018

    work page 2018

  8. [7]

    Fornasier, P

    A. Fornasier, P. van Goor, E. Allak, R. Mahony, and S. Weiss. Msceqf: A multi state constraint equivariant filter for vision-aided inertial navigation. IEEE Robotics and Automation Letters , 9(1):731–738, 2023

    work page 2023

Show all 26 references
  1. [8]

    Forster, M

    C. Forster, M. Pizzoli, and D. Scaramuzza. Svo: Fast semi-direct monocular visual odometry. In 2014 IEEE international conference on robotics and automation (ICRA) , pages 15–22. IEEE, 2014

    2014

  2. [9]

    Hamel and C

    T. Hamel and C. Samson. Position estimation from direction or range measurements. Automatica, 82:137–144, 2017

    2017

  3. [10]

    B. K. Horn and B. G. Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185–203, 1981

    1981

  4. [11]

    Jaegle, S

    A. Jaegle, S. Phillips, and K. Daniilidis. Fast, robust, continuous monocular egomotion computation. In 2016 IEEE International Conference on Robotics and Automation (ICRA) , pages 773–780. IEEE, 2016

    2016

  5. [12]

    Leutenegger, P

    S. Leutenegger, P. Furgale, V . Rabaud, M. Chli, K. Konolige, and R. Siegwart. Keyframe-based visual-inertial slam using nonlinear optimization. Proceedings of Robotis Science and Systems (RSS) 2013, 2013

    2013

  6. [13]

    B. D. Lucas and T. Kanade. An iterative image registration technique with an application to stereo vision. In IJCAI’81: 7th international joint conference on Artificial intelligence , volume 2, pages 674–679, 1981

    1981

  7. [14]

    Y . Ma, S. Soatto, J. Koˇseck´a, and S. Sastry. An invitation to 3-d vision: from images to geometric models , volume 26. Springer, 2004

    2004

  8. [15]

    Mahony, P

    R. Mahony, P. Corke, and T. Hamel. Dynamic image-based visual servo control using centroid and optic flow features. 2008

    2008

  9. [16]

    Mahony and T

    R. Mahony and T. Hamel. A geometric nonlinear observer for simultaneous localisation and mapping. In 2017 IEEE 56th Annual Conference on Decision and Control (CDC), pages 2408–2415. IEEE, 2017

    2017

  10. [17]

    Mahony, T

    R. Mahony, T. Hamel, and J.-M. Pflimlin. Nonlinear complementary filters on the special orthogonal group. IEEE Transactions on auto- matic control, 53(5):1203–1218, 2008

    2008

  11. [18]

    A. I. Mourikis and S. I. Roumeliotis. A multi-state constraint kalman filter for vision-aided inertial navigation. In Proceedings 2007 IEEE international conference on robotics and automation , pages 3565–

    2007

  12. [19]

    R. A. Newcombe, S. J. Lovegrove, and A. J. Davison. Dtam: Dense tracking and mapping in real-time. In 2011 international conference on computer vision , pages 2320–2327. IEEE, 2011

    2011

  13. [20]

    T. Qin, P. Li, and S. Shen. Vins-mono: A robust and versatile monoc- ular visual-inertial state estimator. IEEE transactions on robotics , 34(4):1004–1020, 2018

    2018

  14. [21]

    Scaramuzza and F

    D. Scaramuzza and F. Fraundorfer. Visual odometry [tutorial]. IEEE robotics & automation magazine , 18(4):80–92, 2011

    2011

  15. [22]

    Scaramuzza and Z

    D. Scaramuzza and Z. Zhang. Visual-inertial odometry of aerial robots. arXiv preprint arXiv:1906.03289 , 2019

    1906 arXiv

  16. [23]

    C. V . Stewart. Robust parameter estimation in computer vision. SIAM review, 41(3):513–537, 1999

    1999

  17. [25]

    van Goor and R

    P. van Goor and R. Mahony. Eqvio: An equivariant filter for visual- inertial odometry. IEEE Transactions on Robotics , 39(5):3567–3585, 2023

    2023

  18. [26]

    Zhang and C

    T. Zhang and C. Tomasi. Fast, robust, and consistent camera motion estimation. In Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No PR00149) , volume 1, pages 164–170. IEEE, 1999

    1999

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.