The Kalman Filter: A Practical Introduction

What Is a Kalman Filter?

The Kalman filter is an algorithm that fuses multiple sources of information to estimate the state of a dynamic system. It does this by combining predictions from a mathematical model with noisy measurements, producing an optimal estimate in the leastsquares sense when the underlying noise is Gaussian.

Why Use It?

• Realtime performance: The equations are recursive, requiring only the previous estimate and the newest measurement.
• Robustness to noise: It explicitly models measurement and process noise, weighing each source appropriately.
• Broad applicability: Navigation, finance, robotics, economics, and many engineering fields rely on it.

Key Concepts

State Vector

The state vector x contains all quantities needed to describe the system at a given instant (position, velocity, temperature, etc.).

Process Model

Often written as

x = Fx + Bu + w

where F is the statetransition matrix, B maps control inputs u, and w is process noise with covariance Q.

Measurement Model

Relates the hidden state to observable data:

z = Hx + v

H maps the state to the measurement space, and v is measurement noise with covariance R.

The TwoStep Cycle

1. Prediction (Time Update)
   • Predict the next state: x| = Fx| + Bu
   • Predict the error covariance: P| = FP|F + Q
2. Update (Measurement Update)
   • Compute the Kalman gain: K = P|H(HP|H + R)
   • Correct the state estimate: x| = x| + K(z Hx|)
   • Update the covariance: P| = (I KH)P|

Simple Example: OneDimensional Position Tracking

Assume a vehicle moves along a straight line. The state vector contains position p and velocity v:

x = [p; v]

With a sampling interval t, the transition matrix is

F = [[1, t],     [0, 1]]

If a GPS sensor provides noisy position measurements z = p + noise, the measurement matrix is

H = [1, 0]

Choosing appropriate covariances Q and R, the filter continuously corrects the velocity estimate even though the sensor never measures velocity directly.

Extensions and Variants

• Extended Kalman Filter (EKF): Linearizes nonlinear models by using Jacobians.
• Unscented Kalman Filter (UKF): Propagates a set of sigma points through the nonlinear functions, often yielding better performance than EKF.
• Information Filter: Works with the inverse covariance (information matrix), useful when many measurements share the same state.
• Ensemble Kalman Filter (EnKF): Uses MonteCarlo ensembles, popular in meteorology and largescale geophysical modeling.

Implementation Tips

1. Start with a simple linear model. Validate the filter with synthetic data where the true state is known.
2. Make sure Q and R reflect realistic uncertainties. Overconfident covariances cause divergence.
3. Monitor the innovation (measurement residual). Its covariance should match the predicted value; large discrepancies hint at model mismatch.
4. When dealing with poorly conditioned matrices, use numerically stable forms such as the Joseph covariance update.
5. For highdimensional problems, consider squareroot filters to preserve positive definiteness.

Common Pitfalls

• Incorrect noise modeling: Assuming too low process noise can make the filter ignore new measurements.
• NonGaussian noise: The optimality of the Kalman filter relies on Gaussian assumptions; heavytailed noise may require robust alternatives.
• Discretetime vs. continuoustime: Mixing continuous equations with discrete updates leads to instability.

Further Reading

Wikipedia Kalman filter
Kalman Filter website tutorials and examples
A survey of Kalman filtering and its extensions