Navigation by Skylight Polarization via Rayleigh Model Optimization on a Road Vehicle

arXiv:2610.20797 · cs.RO, cs.SY, eess.SY · Submitted 2026-10-14 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "Navigation by Skylight Polarization via Rayleigh Model Optimization on a Road Vehicle".

Dev: The gist: This paper presents a skylight-polarization navigation method that estimates vehicle heading by globally optimizing a Rayleigh sky model against measurements from a division-of-focal-plane polarimetric camera.

Rosa: First, who's behind it and why it matters.

Paper summary: Rosa: So what we’re hearing is that this paper, "Navigation by Skylight Polarization via Rayleigh Model Optimization on a Road Vehicle," tackles the problem of getting absolute heading estimates when traditional sensors fail. Dev The authors propose a navigation method based on using the polarization pattern of skylight to estimate vehicle heading by optimizing a Rayleigh sky model against data from a division-of-focal-plane polarimetric camera.

Rosa: The framework they build combines several things: a pinhole camera model, that single-scattering Rayleigh polarization model, and a solar position algorithm. Taro This system is used to synthesize the expected full image angle of polarization distribution for candidate vehicle states.

Dev: They define an objective function using a degree-of-polarization-weighted root mean square error that compares their measured patterns with the simulated ones. Rosa That weighting is key because it lets the estimator use all available sky information while attenuating low confidence pixels caused by things like clouds, buildings, foliage, and other depolarizing occlusions.

Taro: The method frames heading estimation as an optimization problem evaluated over the full polarimetric image. Dev They use this optimization to compare measured and simulated polarization patterns.

Rosa: They implemented experimental evaluation using three urban road vehicle trajectories collected with a roof-mounted polarimetric camera and a high precision dual-antenna GNSS/INS reference system. Taro The results they got show sub-degree root mean square yaw error across all those trajectories, which is between zero point seven one eight degrees and zero point eight eight nine degrees.

Dev: Compared to representative feature-based and model-based polarization navigation methods, this proposed approach reduces the root mean square heading error by up to eighty-seven percent and demonstrates improved robustness when parts of the sky are blocked. Rosa So, in short, it’s a global optimization method that uses sky polarization data to get better heading estimates on the road.

Taro: It’s interesting because it doesn't rely on local features, which is where other methods often fail in urban canyons when things like buildings or tree canopies obscure the sky.

Conclusion: Rosa: So, looking at "Navigation by Skylight Polarization via Rayleigh Model Optimization on a Road Vehicle," the main thing we're seeing is this method for getting reliable absolute heading using skylight polarization measurements. Dev It tackles the problem of degraded magnetic, inertial, or GNSS references by globally optimizing a Rayleigh sky model against data from a division-of-focal-plane polarimetric camera.

Rosa: The authors show that they can combine a pinhole camera model, the single-scattering Rayleigh polarization model, and a solar position algorithm to synthesize the expected full image angle of polarization distribution for different vehicle states. Taro They use an objective function that weights the error by degree of polarization to compare measured and simulated patterns, which helps filter out noise from things like clouds or foliage.

Dev: Experimentally, they achieved sub-degree root mean square yaw error across three urban road vehicle trajectories, with errors ranging from zero point seven one eight degrees to zero point eight eight nine degrees when compared against a high precision GNSS/INS reference system. Rosa They found this method reduces the root mean square heading error by up to eighty-seven percent compared to other methods, and it shows improved robustness when parts of the sky are blocked.

Taro: What this means for someone driving is that you get a stable heading estimate even if your GPS signal drops or you’re in a tall city where features get blocked. Dev It suggests that by using the full polarization image and optimizing globally, you don't need to rely on finding one specific feature like the solar meridian.

Rosa: The authors also noted that because each frame of the trajectory measures a new image of the sky, there’s minimal coupling between adjacent state estimations, which makes it good for fusing with IMUs. Taro But they did flag a limitation: they mentioned that a local minimum at the antisolar meridian can cause challenges under high-noise environments.

Dev: So, while the performance is quite strong and robust to partial sky occlusion, you still have to be aware of those specific noise conditions where it might struggle. Rosa Overall, this paper presents a solid framework for absolute heading determination on road vehicles using polarization data.

Ben Potter, Eslam Mounier, Aboelmagd Noureldin, Yahia Antar, Muhammad Alam

Queen’s University · Royal Military College of Canada

cs.RO, cs.SY, eess.SY

Submitted: 2026-10-14

Updated: 2026-10-14

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 61/100

The gist: The gist: This paper presents a skylight-polarization navigation method that estimates vehicle heading by globally optimizing a Rayleigh sky model against measurements from a division-of-focal-plane

Key concepts

Pinhole Camera Model (PCM): This is a geometrical model that tracks how light rays from pixels on the camera sensor travel to points in the sky. It uses principles of optics to relate a pixel's location on the sensor to a specific direction in 3D space, helping map where each part of the image corresponds to in reality.
Rayleigh Polarization Model: This model describes how sunlight becomes polarized after scattering off atmospheric particles. It predicts the expected angle and degree of polarization based on geometry related to the sun's position. This allows the system to simulate what a perfect, unoccluded sky image should look like for a given vehicle state.
Degree-of-Polarization Weighted RMSE: This is the mathematical objective function used to compare what was actually measured from the camera with what the model predicted. It assigns different importance (weights) to each pixel based on its measured polarization degree, which helps filter out noisy or low-confidence data caused by clouds or buildings.
Solar Position Algorithm (PSA): This algorithm accurately calculates the direction of the sun in real-time using location and time data. It is crucial because it provides the necessary input for the Rayleigh model to correctly predict how sunlight will interact with the atmosphere and be observed by a camera at any given moment.

Terminology

Objective Function — 1× in this paper · explained in 2 episodes
Robustness — 1× in this paper · explained in 5 episodes

Summary

The gist: This paper presents a skylight-polarization navigation method that estimates vehicle heading by globally optimizing a Rayleigh sky model against measurements from a division-of-focal-plane polarimetric camera.

How it works

The proposed framework combines a pinhole camera model, a single-scattering Rayleigh polarization model, and a solar position algorithm to synthesize the expected full-image angle-ofpolarization distribution for candidate vehicle states<ref:2610.20797#pg2> The objective function is defined as a degree-ofpolarization-weighted root-mean square error that compares measured and simulated polarization patterns<ref:2610.20797#pg6> This weighting naturally attenuates low-polarization noise from depolarizing cloud cover and solid occlusions<ref:2610.20797#pg6> The method formulates heading estimation as an optimization problem evaluated over the full polarimetric image<ref:2610.20797#pg4>

Key Components and Theory

The theoretical foundation involves several interconnected models to bridge the gap between sensor measurements and absolute heading determination<ref:2610.20797#pg4>

** Pinhole Camera Model (PCM): The PCM is implemented as a geometrical model for tracing the path of a light ray from a sensor pixel onto a sky point<ref:2610.20797#pg5> It defines the relationship between each sensor location and its corresponding ray in the camera body frame<ref:2610.20797#pg5> The PCM uses an optical axis, principal axis, or principal ray defined by the Zb axis<ref:2610.20797#pg5> The forward projection and inverse backward projection are used to trace the pixel coordinate through the lens into a direction p b using equations (7) and (8)<ref:2610.20797#pg5> The transformation between the b-frame and the n-frame, denoted Cnb, is determined by the orientation of the present state estimate<ref:2610.20797#pg5> The transformation from the sensor frame to the ray frame involves approximating an angle theta i from Xs i to Xr i with some error epsilon i<ref:2610.20797#pg5> This approximation is exact in the level case where the image sensor is aligned with the horizon, but an error term epsilon i > 0 exists in the tilt case<ref:2610.20797#pg5> The measured AoP from the s-frame is transformed to the r-frame using equation (14)<ref:2610.20797#pg5> The DoP of the ray is determined by a formula involving the scattering angle γ, defined as cos γ = −→OS · −→OP<ref:2610.20797#pg5> The Rayleigh model predicts the AoP and DoP based on geometry in relation to the sun, where light after scattering is orthogonal to the plane defined by O, P, and S<ref:2610.20797#pg5> The solar position algorithm (PSA) determines the direction of the solar vector using location and time inputs<ref:2610.20797#pg5> The PSA algorithm is popular for open-loop computation of the solar vector due to its balance between accuracy and computational complexity<ref:2610.20797#pg5> The Rayleigh model is implemented because it does not depend on accurate modelling of the neutral points, and prior work has shown that it matches measurements remarkably well<ref:2610.20797#pg4> The objective function compares the simulated AoP distribution with real measurements, weighted by the measured DoP to increase invariance to noisy AoP measurements<ref:2610.20797#pg6> The objective function is defined as J(s) = s P d i [a r i − aˆ r i (s)]2 / N P d i<ref:2610.20797#pg6> The expected AoP is computed using the Rayleigh model and a camera model, simulating the expected full-image AoP distribution for candidate vehicle states<ref:2610.20797#pg4> The measurement with a polarization camera involves measuring intensity at orthogonal linear polarization angles (I0, I45, I90, and I135) to derive the Stokes vector<ref:2610.20797#pg5> The AoP is derived from the Stokes vector using a formula involving S2 and S1<ref:2610.20797#pg5> The DoP is derived from the Stokes vector using a formula involving S0, S2, and S3<ref:2610.20797#pg5> The transformation between the s-frame and the r-frame involves finding the pixel coordinate of the zenith vector U b by applying equation (12)<ref:2610.20797#pg5> This process identifies the pixel on the image sensor that observes the zenith location<ref:2610.20797#pg5> The transformation is approximated using equation (13) which involves an angle theta i from Xs i to Xr i with some error epsilon i<ref:2610.20797#pg5> This approximation is reasonable for the application domain, as the absolute value of epsilon is relatively small<ref:2610.20797#pg5> The measurement with a polarization camera captures polarization information by superimposing a repeating pattern of 2x2 micropolarizer arrays onto the pixel matrix<ref:2610.20797#pg5> Each instance of the micro-polarizer array on the image sensor is called a superpixel<ref:2610.20797#pg5> The Stokes vector is derived from intensity measurements for a single superpixel following equation (9)<ref:2610.20797#pg5> The AoP measured by a DoFP polarimeter is taken implicitly with reference to the 0◦ polarizer in each superpixel<ref:2610.20797#pg5> The error term epsilon i comes from simulating only the optical path of the zenith, rather than the optical paths for every pixel position<ref:2610.20797#pg5> This approach is motivated by the fact that using the zenith position can provide an exact transform in the level case without exhaustive simulation<ref:2610.20797#pg5> The DoP weighting allows for strategic weighting of each pixel, enabling estimation to use as much of the available polarization information as possible<ref:2610.20797#pg6> This granularity allows the estimation to rely on pixels with a large weight factor, while still considering pixels which may be excluded by a traditional binary threshold<ref:2610.20797#pg6> The camera model uses the orientation of the state estimate to map each pixel onto a corresponding sky point<ref:2610.20797#pg5> The simulation is based exclusively on the state (i.e., position, time, orientation) of a virtual polarization camera in the horizontal coordinate system<ref:2610.20797#pg4> The state estimate is composed of a position, orientation, and time<ref:2610.20797#pg5> The objective function evaluates the difference between the predicted AoP of each pixel against real measurements<ref:2610.20797#pg6> The camera model uses a focal length of 8 mm in its implementation<ref:2610.20797#pg5> The real polarimeter has a pixel size of 3.45 µm in a 2048x2448 array, which corresponds to a superpixel size 6.9 µm in a 1024x1224 array<ref:2610.20797#pg6> The method achieves up to an 87 % decrease in the root mean squared error (RMSE) of the heading estimate compared with traditional HT-based methods<ref:2610.20797#pg6> The method exhibits a drift-free property because each frame of the trajectory measures a new image of the sky, leading to minimal coupling between adjacent state estimations<ref:2610.20797#pg6> This property makes our method ideal for the application of heading constraints on fused sensor systems with IMUs<ref:2610.20797#pg6> The proposed method achieves sub-degree RMS yaw error across all trajectories, ranging from 0.718◦ to 0.889◦<ref:2610.20797#pg6> Compared with representative feature-based and model-based polarization navigation methods, the proposed approach reduces RMS heading error by up to 87 % and demonstrates improved robustness to partial sky occlusion<ref:2610.20797#pg6> The method is robust against heading ambiguity because it does not rely on any single feature of the angle of polarization distribution<ref:2610.20797#pg6> However, a local minima at the antisolar meridian may cause challenges under high-noise environments<ref:2610.

Improvements for AI systems

Improve heading estimation robustness by utilizing a degree-of-polarization-weighted root mean squared error objective function to attenuate low-confidence pixels caused by clouds, buildings, foliage, and other depolarizing occlusions. This allows the estimator to use all available sky information while attenuating low-confidence pixels, leading to a demonstrated reduction in RMS heading error by up to 87% compared to existing methods.
Enhance absolute heading determination under non-horizontal camera orientations by explicitly accounting for measurement tilts through a formulated coordinate transformation between the DoFP measurement system and the ray coordinate systems, documenting the errors incurred from such tilted measurements in vehicle-mounted schemes. This directly addresses systematic errors that are often overlooked in literature.
Develop a navigation system capable of drift-free absolute heading estimation by combining a pinhole camera model (PCM) with single-scattering Rayleigh sky theory and an accurate solar position algorithm to synthesize the expected full-image angle-ofpolarization distribution for candidate vehicle states. This framework allows the estimator to operate successfully even when traditional references like GNSS or IMUs are degraded.
Improve computational efficiency in heading estimation by extending the optimization with advanced search strategies, such as Riemannian gradient descent, instead of a simple grid search. This strategy is intended to reduce the number of objective-function evaluations required for each estimate by only evaluating promising states, thereby mitigating the current computational expense of repeated full-image simulations.
Increase resilience to partial sky occlusion by implementing a mechanism that allows for strategic weighting of each pixel in the objective function instead of binary thresholding. This granularity permits the estimation to rely on pixels with a large weight factor, while still considering pixels which may be excluded by a traditional binary threshold, allowing for more nuanced data utilization.