Unscented Kalman Filter (UKF)

The core innovation of the UKF is the unscented transform, a deterministic sampling technique. It selects a minimal set of sigma points that capture the true mean and covariance of a probability distribution. These points are then propagated through the nonlinear system and measurement functions. The transformed points are used to compute the new mean and covariance, providing a more accurate approximation than a first-order Taylor series linearization used by the EKF.

• Deterministic Sampling: Unlike Monte Carlo methods, sigma points are chosen via a fixed rule.
• Captures Moments: Designed to accurately capture the first two moments (mean, covariance) of the distribution.
• No Jacobians Required: Eliminates the need to derive and compute complex Jacobian matrices.

Sigma points are a carefully chosen set of sample points that represent the prior state distribution. For an n-dimensional state vector with mean (\mathbf{x}) and covariance (\mathbf{P}), a typical set consists of (2n+1) points.

• Central Point: One point at the mean.
• Symmetric Points: (2n) points symmetrically spread along the principal axes of the covariance ellipse, scaled by a parameter (\sqrt{(n+\lambda)}).
• Weighting: Each sigma point is assigned a weight for mean and covariance calculation. These weights control the spread of points and incorporate prior knowledge about the distribution (e.g., kurtosis).

The UKF excels in systems with highly nonlinear dynamics or measurement models where the EKF's linearization can introduce significant errors or cause divergence. By propagating sigma points through the true nonlinear function, it better approximates the posterior distribution.

Common Applications:

• Aerospace: Aircraft attitude estimation (highly nonlinear kinematics).
• Robotics: SLAM and odometry for non-holonomic robots.
• Automotive: Vehicle state estimation during aggressive maneuvers.
• Finance: Modeling stochastic volatility.

A major practical advantage is that the UKF is derivative-free. It does not require the analytical derivation of Jacobian matrices for the system's process (f) and observation (h) models.

• Reduced Development Complexity: Engineers implement the actual nonlinear functions, not their derivatives.
• Improved Robustness: Avoids errors from incorrect Jacobian calculations.
• Easier Maintenance: System models can be changed without needing to re-derive linearizations.

This makes the UKF particularly attractive for complex, black-box, or rapidly prototyped systems.

The UKF tends to produce more consistent covariance estimates than the EKF. The estimated covariance from the UKF more accurately reflects the true error in the state estimate because it captures the nonlinear transformation's effect on uncertainty.

• Better Uncertainty Quantification: Crucial for sensor fusion gating, resource allocation, and risk-aware planning.
• Improved Data Association: More accurate Mahalanobis distance calculations for associating new measurements to existing tracks.
• Tuned Performance: Parameters like the primary scaling parameter alpha allow tuning between sampling spread and higher-order moment accuracy.

The UKF's superior accuracy comes with a computational cost. It requires evaluating the nonlinear function (2n+1) times per filter step, compared to the EKF's single evaluation plus Jacobian computation.

• Complexity: (O(n^3)) for covariance updates, similar to EKF, but with a larger constant factor.
• When to Choose UKF: Ideal when nonlinearities are severe, derivatives are hard to compute, or model fidelity is critical.
• When EKF Suffices: For mildly nonlinear systems or in severely resource-constrained environments (e.g., tinyML), the EKF may be preferred for its speed.

The UKF is often seen as a sweet spot between the simple-but-fragile EKF and the powerful-but-expensive Particle Filter.

The foundational recursive algorithm for optimal linear state estimation. It operates in a two-step cycle:

• Prediction: Projects the current state and its uncertainty forward using a linear process model.
• Update: Corrects the prediction with a new sensor measurement, weighting the correction by the relative certainty (covariance) of the prediction and the measurement. It assumes Gaussian-distributed noise and a linear system model, providing the theoretical basis for the UKF and EKF.

The primary nonlinear predecessor to the UKF. The EKF handles nonlinearities by linearizing the system's process and observation models around the current state estimate using a first-order Taylor series expansion. This Jacobian-based linearization can introduce significant errors for highly nonlinear systems or with large uncertainties, a limitation the UKF was designed to address. It remains widely used for moderately nonlinear problems.

A sequential Monte Carlo method for state estimation in nonlinear and non-Gaussian systems. It represents the probability distribution with a large set of random samples (particles), each with an associated weight.

• Propagates particles through the nonlinear model.
• Updates weights based on sensor likelihood.
• Resamples to focus computation on high-probability regions. It is more flexible than the UKF for arbitrary distributions but is computationally expensive, often requiring thousands of particles for high-dimensional states.

The general probabilistic framework that encompasses the Kalman, UKF, EKF, and Particle filters. It recursively estimates the posterior probability distribution of a system's state given all previous measurements: P(state | measurements). All specific filters are implementations of this framework, making different trade-offs (e.g., Gaussian assumption vs. computational cost) to solve the Bayes filter equations in practice.

The overarching engineering discipline of combining sensory data from disparate sources (e.g., camera, lidar, IMU, radar) to produce estimates with greater accuracy, completeness, and reliability than is possible from any single sensor. The UKF is one algorithmic tool within this discipline. Fusion architectures include:

• Centralized: Raw data fused at a central node.
• Decentralized: Local estimates fused across a network.
• Hybrid: A mix of both approaches.

The core problem the UKF solves: inferring the hidden variables (state) of a dynamic system from a sequence of noisy, indirect observations. The state typically includes quantities like position, velocity, orientation, and sensor biases. It is a fundamental requirement for autonomous navigation, target tracking, and process control. The UKF provides a computationally efficient method for nonlinear state estimation when the noise can be approximated as Gaussian.