MarsCoT: Fiducial Marker Pose Uncertainty Estimation by Propagating Corner Uncertainty Through the Unscented Transform

초록

Square fiducial markers, such as AprilTags and ArUco, are widely used in computer-vision tasks that require six-degree-of-freedom (6-DoF) pose estimation from a single image. However, existing systems can estimate pose but lack the ability to quantify pose uncertainty. This uncertainty information is crucial for evaluating pose reliability and for providing uncertainty information to probabilistic algorithms such as Kalman filters and factor graphs. Current uncertainty estimators rely on computationally expensive Monte Carlo (MC) simulation, experimental methods requiring extensive data collection, or analytical approaches with unstable Jacobian computations. This article presents a Jacobian-free analytical framework that augments any square-marker pose estimator with real-time 6-DoF pose uncertainty. Our method derives corner uncertainties from local image features and propagates these corner uncertainties through the nonlinear PnP algorithm using the unscented transform (UT) to estimate pose uncertainty. We validate our approach on a diverse synthetic dataset and through real-world experiments. The corner uncertainty estimates are validated using chi-square tests. The pose uncertainty estimates demonstrate close agreement with the MC reference via low Jensen–Shannon divergence (JSD) and are validated through confidence interval (CI) analysis. The framework operates at > 166 Hz, enabling seamless integration into real-time applications.

키워드