Robust Variational Bayes UKF for GPS in Urban Canyons: Outlier Rejection with Student's t-Distribution
In urban canyons, GPS signals suffer from multipath propagation and signal blockage, causing outliers in pseudoranges. The classic Fusion UKF with Gaussian noise loses accuracy, while the Variational Bayes Fusion UKF models noise as a Student's t-distribution. Iterative estimation of the scale parameter automatically downweights anomalous measurements. Simulations showed over 50% reduction in positioning RMSE compared to the baseline UKF.
Robustness via Noise Model Replacement and Sensor Fusion
Classic Kalman filters assume Gaussian additive noise, but real-world GPS outliers have heavy tails. Switching to a Student's t-distribution weights measurements by residuals: large errors get low weight, preventing trajectory distortion.
The weight is computed as $w = \frac{\nu + m}{\nu + d^2}$, where $\nu$ is degrees of freedom, $m$ is measurement dimension, and $d^2$ is the squared Mahalanobis distance. With low $\nu$ (e.g., 2), weights drop quickly for outliers; with high $\nu$ ($\nu \to \infty$), it approaches the Gaussian case.
Sensor fusion pairs GPS with inertial data (IMU). In the VB approach, covariance matrices $R_i$ for each sensor are iteratively estimated as random variables, enabling dynamic trust reallocation.
During GPS anomalies, VB-UKF inflates $R_{GPS}$ variance, reducing its influence and relying on IMU until the signal stabilizes.
Safeguards in Implementation
# 1. Cap lambda at 1 to prevent model mismatch
if self.cap_lambda_to_1:
lamvec = np.minimum(lamvec, 1.0)
# 2. Prevent division by zero
lamvec = np.clip(lamvec, self.lambda_clip[0], self.lambda_clip[1])These clips prevent weights >1 (physically impossible) and NaNs from extreme outliers.
Computational Complexity and Trade-offs
Classic UKF uses 2n+1 sigma points for nonlinear approximation, with $O(n^3)$ complexity from matrix inversions. VB-UKF adds iterations for variational Bayesian inference to approximate the Student's t posterior with a Gaussian.
Variational inference minimizes KL divergence between the true posterior and approximation, converging in 5–10 iterations per step. It's a sweet spot: more efficient than Particle Filters ($O(N_p n^2)$, $N_p \gg 2n$), but more robust than plain UKF.
For real-time streaming on embedded systems with dynamics up to 100 Hz, VB-UKF is well-suited.
Simulation Setup: CTRV Trajectory
Dynamics model: Constant Turn Rate and Velocity (CTRV) for a maneuvering vehicle in an urban canyon:
- State: $[x, y, v, \psi, \dot{\psi}]^T$
- GPS measurements: $[x_{gps}, y_{gps}]$ with outliers
- IMU: accelerometer, gyroscope
Parameters:
- Nominal GPS noise: $\sigma_{gps} = 3$ m
- Outliers: jumps to 50–100 m at 5–10% probability
- $\nu = 4$ for balanced sensitivity
- Simulation: 1000 s, 10 Hz rate
Data structure: timestamped buffered measurements, 5D state vector.
Simulation Results
Trajectories and RMSE
On trajectories with turns and outliers, classic UKF drifts 20–50 m, while VB-UKF keeps errors <10 m. Positioning RMSE: 7.2 m (VB) vs 18.5 m (Fusion UKF).
NIS (Normalized Innovation Squared) for VB-UKF matches $\chi^2$ distribution, confirming correct noise modeling.
Weight Adaptation
VB mechanism dynamically downweights GPS when $d^2 > \nu$: plots show drop from 1 to 0.1 in 2–3 steps post-outlier, recovery in 5 steps.
Transient Response
After an outlier, VB-UKF recovers trajectory in 10–20 steps vs 50+ for baseline UKF. Adaptive $R$ smooths the transition.
Key Takeaways
- VB-UKF halves RMSE in GPS outliers using Student's t and iterative $R$ estimation.
- Dynamic sensor weighting without heuristics: GPS/IMU priority shifts Bayesian-style.
- Compute cost: 5–10 iterations/step, fine for real-time.
- Tune via $\nu$: 2–4 for aggressive robustness, >10 for Gaussian-like.
- NIS/NEES validate model consistency.
— Editorial Team
No comments yet.