Abstract
The widely-used Extended Kalman Filter (EKF) provides a straightforward recipe to estimate the mean and covariance of the state given all past measurements in a causal and recursive fashion. For a wide variety of applications, the EKF is known to produce accurate estimates of the mean and typically inaccurate estimates of the covariance. For applications in visual inertial localization, we show that inaccuracies in the covariance estimates are systematic, i.e. it is possible to learn a nonlinear map from the empirical ground truth to the estimated one. This is demonstrated on both a standard EKF in simulation and a Visual Inertial Odometry system on real-world data.
Introduction
Extended Kalman Filters (EKFs) are the workhorse of robotic localization, providing both state estimates and their associated uncertainty (covariance). However, EKFs are notoriously "overconfident," often yielding uncertainty estimates significantly smaller than actual errors. This miscalibration stems from the linearization of complex nonlinear dynamics and the assumption of Gaussian noise, which frequently fails in vision-based tasks. In safety-critical applications like autonomous driving, underestimating error bounds can lead to catastrophic failures. This paper demonstrates that EKF miscalibration is not random but systematic, allowing for correction via efficient neural networks.
Methods
The authors propose a "post-hoc" calibration framework that learns a nonlinear mapping to correct the output of an existing EKF. Key technical components include:
- Low-Dimensional Inputs: By using only the estimated state ($\hat{x}$) and covariance ($\hat{P}$) as inputs, the authors employ extremely small and fast feedforward neural networks, making them suitable for real-time robotics.
- Calibration Metric: The primary metric is the L2 divergence between the empirical distribution of the Normalized Estimation Error Squared (NEES) and the theoretical $\chi^2$ distribution. A perfectly calibrated filter would yield an L2 divergence of zero.
- Ergodicity Assumption: To address the impossibility of Monte-Carlo simulations in real-world settings, the authors assume ergodicity—that temporal error windows represent the statistical population—enabling the calculation of "pseudo-ground-truth" covariance from single trajectories.
- Model Comparison: The study evaluated four models of increasing complexity, ranging from simple constant scaling to state-dependent neural networks.
Results and Discussion
The approach was validated across three datasets: a simulated linear Kalman filter, a 2D Dubin’s car model, and the real-world TUM Visual Inertial (VI) dataset.
Key findings include:
- Systematic Miscalibration: The neural network approaches significantly outperformed traditional scaling. In 2D simulations, divergence from ground truth was reduced by up to 90.9%. On the TUM VI dataset, a 97.8% reduction was achieved.
- Improved Error Bounds: In the uncalibrated VIO system, only ~11% of errors fell within the 1-$\sigma$ bound. Post-calibration, this improved to ~70-76%, aligning with theoretical Gaussian expectations.
- Covariance as a Sufficient Signal: Interestingly, models using only the estimated covariance $\hat{P}$ performed nearly as well as those including the state $\hat{x}$, suggesting that $\hat{P}$ contains most of the information needed to diagnose its own miscalibration.
Conclusion
This work proves that EKF uncertainty miscalibration in visual-inertial systems is systematic and learnable. By applying a post-hoc neural correction, researchers can significantly improve the reliability of localization without redesigning the underlying filter. Future research will investigate "in-the-loop" calibration and the potential of Recurrent Neural Networks (RNNs) to leverage temporal context for even greater accuracy.
Summary updated