Kalman Filter

Kalman Filter
💡No image available
Overview

The Kalman filter is an algorithm for estimating the state of a dynamic system from a series of incomplete and noisy measurements. It is widely used in engineering, navigation, and signal processing because it produces optimal estimates under well-known statistical assumptions. The method was developed by Rudolf E. Kálmán and is closely related to Bayesian inference and state-space models.

Overview

A dynamic system is typically modeled as having an unobserved (hidden) state that evolves over time according to a process model, while observations provide indirect and noisy information about that state. In the standard formulation, the system is expressed in a linear form with additive Gaussian noise, a setting sometimes summarized as Gaussian noise. The Kalman filter maintains an estimate of the current state along with its uncertainty, updating both as new measurements arrive.

The algorithm operates recursively: it alternates between a prediction (time update) step and a correction (measurement update) step. This structure allows real-time estimation without storing all past data. The same conceptual framework underlies many modern tracking and control pipelines, including control theory applications and guidance systems that estimate position and velocity from sensor readings.

Mathematical formulation

In the discrete-time linear Gaussian model, the system state (x_k) evolves as [ x_k = F_k x_{k-1} + B_k u_k + w_k, ] and measurements (z_k) are related to the state as [ z_k = H_k x_k + v_k, ] where (F_k) and (H_k) are matrices defining the dynamics and measurement model, (u_k) is a known input, and (w_k) and (v_k) represent process and measurement noise, respectively. When these noises are independent and Gaussian, the posterior distribution of the state remains Gaussian, so the filter can represent it using only the mean and covariance.

The update equations use the Kalman gain to balance trust between the predicted state and the new measurement. The gain depends on both the predicted uncertainty and the measurement noise level. This mechanism parallels ideas from least squares estimation, but it is performed sequentially and accounts for uncertainty propagation through the model.

Properties and optimality

Under the linear Gaussian assumptions, the Kalman filter yields the minimum mean square error estimate of the state, meaning it minimizes the expected squared difference between the estimated and true states. This optimality is often discussed in terms of conditional expectation because the filter computes the mean of the posterior distribution given all measurements up to the current time.

The filter’s uncertainty quantification is central: the covariance matrix not only tracks estimation error but also informs future updates. Because the covariance evolves according to the model and noise statistics, the Kalman filter can detect when measurements are likely to be unreliable and automatically weight them less.

Variants and extensions

When the system dynamics or measurement models are nonlinear, the standard Kalman filter may not apply directly. Common extensions include the extended Kalman filter (EKF), which linearizes nonlinear functions around the current estimate, and the unscented Kalman filter (UKF), which uses a deterministic sampling strategy to propagate uncertainty through nonlinear transformations. These methods are used in nonlinear estimation problems such as robot localization and navigation.

Another important generalization is the Kalman smoothing approach, which estimates past states using information from future measurements. There is also the information filter form, which reformulates the update in terms of the inverse covariance and can be advantageous in certain computational settings. For scenarios involving outliers or non-Gaussian noise, robust alternatives may replace or augment the classical assumptions.

Applications

Kalman filtering is used wherever estimating hidden variables from noisy sensors is required. In navigation, it supports fusion of data from instruments such as accelerometers and gyroscopes, improving estimates of attitude, position, and velocity. In signal processing, it is used for tracking time-varying signals and filtering noise in dynamic systems modeled with state variables. In control engineering, it is frequently paired with controllers to form estimation–control loops that rely on filtered state feedback.

The method also appears in large-scale systems and distributed settings, including sensor networks and multi-agent estimation, where different nodes contribute partial measurements. For such setups, variants of Kalman filtering have been developed to handle communication constraints and evolving network topology, drawing on ideas from distributed computing.