What is Unscented Kalman Filter?
An Unscented Kalman Filter is a nonlinear Gaussian state estimator that propagates deterministically weighted sigma points through transition and observation functions to approximate posterior moments.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Represent moments with deterministic sigma points
For an n-dimensional mean m and covariance P, the common scaled Unscented Transform forms 2n+1 points from columns of a matrix square root of (n+lambda)P. Separate mean and covariance weights reconstruct the first two moments before and after a nonlinear map. These are quadrature points, not random particles or posterior samples.
Julier and Uhlmann's 1997 paper introduced this derivative-free nonlinear extension. The exact scaling and weight convention must be documented because multiple Sigma Point parameterizations exist and can behave differently.
Propagate transition and measurement moments
Prediction sends every Sigma Point through f, computes the weighted predicted mean and covariance, and adds process-noise covariance. Measurement update usually regenerates points from that predicted Gaussian, sends them through h, and computes predicted measurement covariance plus the state-measurement cross-covariance. Their ratio defines the Kalman-style gain.
For non-additive noise, augmenting the state with noise variables is one valid construction, but it increases Sigma Point count. Angles, quaternions, positive quantities, and constrained states need geometry-aware means, residuals, and retractions; Euclidean averaging can produce invalid states.
Audit scaling, covariance, and approximation error
The AHRS UKF documentation shows the scaled parameters alpha, beta, and kappa, Cholesky-based points, and distinct mean and covariance weights. Small alpha can produce large-magnitude positive and negative weights; mathematically valid weights still demand stable square-root and covariance handling.
Track innovation whiteness, normalized innovation squared, state error, positive definiteness, Cholesky failures, sensitivity to scaling parameters, and runtime. Compare UKF with EKF and Particle Filter under identical models and seeds. Better moment propagation on one nonlinear example does not establish better calibration, robustness, or cost for another.
Key Characteristics
- Approximates one Gaussian distribution with deterministic Sigma Points
- Propagates points directly through nonlinear model functions
- Avoids analytic transition and measurement Jacobians
- Reconstructs predicted means, covariances, and cross-covariances
- Uses configurable Sigma Point scaling and weighting rules
- Still approximates multimodal, constrained, or strongly non-Gaussian states poorly
Common Use Cases
- Nonlinear navigation and sensor fusion without reliable Jacobians
- Attitude estimation with geometry-aware Sigma Point handling
- Tracking with nonlinear transition and observation functions
- Embedded Gaussian filtering when Particle Filters are too costly
- Benchmarking first-order versus deterministic-moment approximations
Example
Loading code...Frequently Asked Questions
Are UKF Sigma Points random samples?
No. They are deterministic weighted quadrature points chosen to match a mean and covariance. They do not form an empirical posterior population and should not be interpreted like Particle Filter particles.
Does a UKF produce an exact nonlinear posterior?
No. It approximates transformed moments and then retains a single Gaussian state summary. Nonlinear transformations can create skewed, constrained, heavy-tailed, or multimodal distributions that one mean and covariance cannot represent.
How do alpha, beta, and kappa affect a UKF?
They determine Sigma Point spread and weights in the scaled Unscented Transform. Their interpretation depends on the chosen convention; poor combinations can create extreme weights, numerical cancellation, or points outside the meaningful state domain.
Is an Unscented Kalman Filter always better than an EKF?
No. A UKF avoids Jacobians and may propagate moments better under some nonlinearities, but it needs repeated model evaluations and covariance factorizations. A well-parameterized EKF can be cheaper and equally or more reliable for locally linear systems.
When should a Particle Filter be used instead of a UKF?
Use Particle Filter comparisons when the posterior may be multimodal, strongly skewed, truncated, or governed by non-Gaussian likelihoods. Particle methods add Monte Carlo and resampling costs, so the decision should follow measured accuracy and runtime.