Unscented Kalman Filter
- Introduced by
- Simon Julier
- Year
- 1997
- Type
- algorithm
- Family
- Nonlinear Estimation
What does it do?
The Unscented Kalman Filter (UKF) is a nonlinear state estimation algorithm that approximates nonlinear systems without requiring explicit Jacobian computation.
Introduced by Julier and Uhlmann in 1997, the UKF uses the unscented transform—a deterministic method to capture mean and covariance statistics through a carefully chosen set of sample points (sigma points)—making it more accurate than the Extended Kalman Filter for highly nonlinear systems while avoiding the computational burden of derivative calculations.
The intuition behind the method
The key idea of the UKF is to replace explicit linearization (used in the EKF) with a statistical sampling approach. Rather than computing Jacobians, the UKF selects a minimal set of deterministically chosen points (sigma points) around the current state estimate, propagates each through the nonlinear function, and recovers the mean and covariance of the results. Because the nonlinear function is evaluated at multiple points rather than approximated by its gradient, the UKF captures nonlinearity more accurately. For a Gaussian distribution, this approach is exact to the third moment (skewness), compared to the EKF which is only first-order accurate.
History
Julier and Uhlmann introduced the unscented transform in 1997 as a general technique for propagating mean and covariance through nonlinear functions. Wan and Van der Merwe developed the full UKF algorithm in 2000 and demonstrated superior performance over EKF in benchmark problems. Sarkka's textbook (2013) provided comprehensive treatment including adaptive and distributed UKF variants. The UKF has become standard in robotics and aerospace, especially for attitude estimation (quaternion dynamics are highly nonlinear).
When to use it
Use the UKF when you have a strongly nonlinear system and want accuracy better than the EKF without computing Jacobians. It is ideal for robotics (especially 3D attitude estimation), navigation with nonlinear sensor models (radar, bearing angles), and systems where linearization is a poor approximation. The UKF is more computationally expensive than EKF (by a small factor), so avoid it only if computation is severely constrained or system nonlinearity is mild (EKF sufficient).
Limitations and considerations
- Slightly more computationally expensive than EKF (2n+1 function evaluations per time step instead of Jacobian).
- Still assumes Gaussian distributions; non-Gaussian noise can cause filter divergence.
- Sigma point propagation requires storing and manipulating more matrices than EKF.
- No universally optimal choice of alpha and beta parameters; system-dependent tuning may be needed.
- Can still diverge if initial estimate is far from truth or model mismatch is severe.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- More accurate than EKF for highly nonlinear systems; captures third-order moments without explicit Jacobians.
- No need to compute or code Jacobian matrices; reduces implementation errors and improves robustness.
- Minimal tuning required; sigma point placement is determined by well-established formulas.
- Provably convergent for a wider class of nonlinear systems compared to EKF.
- Works well with non-Gaussian noise if the noise statistics can be approximated by a Gaussian mixture.
Applications
- Robotics: 3D attitude estimation from accelerometer and gyroscope, mobile robot SLAM with nonlinear sensor models.
- Aerospace: Aircraft attitude and heading estimation, spacecraft attitude determination from star tracker and gyro measurements.
- Autonomous vehicles: GPS/INS fusion with nonlinear dynamics, sensor fusion from radar, lidar, and camera.
- Navigation: Inertial navigation system error correction, bearings-only target localization.
- Power systems: Nonlinear state estimation in power grids, harmonic detection in AC power systems.
How it works
How it works
Set the initial state estimate x_hat_0 and covariance P_0; also set UKF parameters (α, β, κ) that control sigma point spacing.
Generate 2n+1 sigma points from the current estimate and covariance, where n is the state dimension. Points are placed at the mean and at locations sqrt(λ + n) standard deviations away.
Pass each sigma point through the nonlinear motion model to get predicted sigma points; compute predicted mean and covariance from the propagated points.
Pass predicted sigma points through the measurement function to get predicted measurements; compute Kalman gain and update state and covariance using the standard update equations.
Common pitfalls
- Using default UKF parameters (α, β, κ) without understanding their effect; these control sigma point spread and higher-order moment approximation.
- Over-specifying measurement noise covariance R to suppress oscillations instead of improving the model.
- Ignoring the curse of dimensionality; UKF scales as O(n^3) in state dimension, so high-dimensional state spaces (>50D) require approximations.
- Assuming UKF is robust to all nonlinearities; it still assumes the system can be well-approximated by a Gaussian at each step.
Frequently asked questions
What are sigma points and how are they chosen?
Sigma points are a deterministically chosen set of sample points that capture the mean and covariance of a Gaussian distribution. For n-dimensional state, 2n+1 sigma points are placed at: x and x ± sqrt((λ+n) P), where λ = α^2(n+κ)−n. The parameters α and κ control spread: α (typically 0.001-1) controls distance from mean, κ is a secondary scaling parameter (often 0).
How does UKF compare to EKF in terms of accuracy and computation?
UKF is typically more accurate for nonlinear systems (third-order vs. first-order moment matching). Computationally, UKF requires 2n+1 evaluations of the nonlinear function per step, while EKF requires one function evaluation plus one Jacobian evaluation. For low-dimensional systems (n<20), the overhead is small; for high-dimensional systems, EKF can be faster despite lower accuracy.
Can I use UKF with multimodal (non-Gaussian) distributions?
Standard UKF assumes Gaussian distributions and will perform poorly with multimodal distributions. For multimodal posteriors, use a Particle Filter, which represents the distribution as a weighted set of particles. Alternatively, use the Gaussian Mixture UKF, which maintains multiple Gaussian components.
How do I tune the UKF parameters α, β, and κ?
Standard choice: α = 0.001 (places sigma points very close to mean), κ = 0, β = 2 (optimal for Gaussian). For non-Gaussian noise, try β = 2 (Gaussian optimal), larger β emphasizes outer sigma points. For systems with strong nonlinearity, increase α (e.g., 0.1) to spread sigma points wider. Empirically test on simulated data.
Related methods
Sources
Julier, S. J., & Uhlmann, J. K. (1997). A new method for the nonlinear transformation of means and covariances in filters and estimators. IEEE Transactions on Automatic Control, 45(3), 477-482.
Open publication referenceCopied.Wan, E. A., & Van Der Merwe, R. (2000). The unscented Kalman filter for nonlinear estimation. Proceedings of the IEEE 2000 Adaptive Systems for Signal Processing, 153-158.
Open publication referenceCopied.Sarkka, S. (2013). Bayesian Filtering and Smoothing. Cambridge University Press.
Open DOI referenceCopied.