Why Kalman Filters Work So Well

Summary

The video explains how a Kalman filter combines imperfect predictions and measurements to produce a more reliable estimate. Its central idea is to track not only the best estimate of a system’s state, but also how uncertain that estimate is. The filter uses this uncertainty to decide how much to trust each information source.

Core concepts

• State: A set of values describing a system at a particular moment. For a moving car, it might include position and velocity.
• Two imperfect information sources:
  • A model predicts how the system should change, but its errors can accumulate over time.
  • A sensor measures the real world, but its readings contain noise.
• Relying only on measurements can make an estimate jump around; relying only on a model can let it drift away from reality. A fixed average is also inadequate when the sources have different levels of reliability.
• The filter represents its estimate as a Gaussian distribution:
  • The center is the best estimate.
  • The variance describes uncertainty: a larger variance means less confidence.

How a single update works

The video uses a car-position example:

• The model predicts a position of 100 m, with variance 4.
• GPS reports 106 m, with variance 12.
• The difference between the prediction and the measurement is the innovation (or residual): 6 m.
• The Kalman gain determines how far to move the prediction toward the measurement:

[ K = \frac{P}{P+R} ]

Here, (P) is the prediction variance and (R) is the measurement variance, so:

[ K = \frac{4}{4+12} = \frac{1}{4} ]

• The estimate moves one-quarter of the way toward GPS:

[ 100 + \left(\frac{1}{4} \times 6\right) = \mathbf{101.5\ m} ]

• The updated variance falls to 3, lower than either the prediction’s variance of 4 or GPS’s variance of 12.

The gain is not an arbitrary weighting. Assuming the prediction and sensor errors are independent, it minimizes the uncertainty of the combined estimate. A more reliable measurement gets more weight; a noisier one gets less. The updated estimate’s precision—the inverse of its variance—is the sum of the two sources’ precisions.

The recurring predict–update cycle

A Kalman filter repeats two steps:

1. Predict
   • Use the system model to project the current state forward.
   • Increase uncertainty to account for unpredictable changes, represented as process noise.
2. Update
   • Compare the prediction with a new measurement.
   • Calculate the Kalman gain and adjust the estimate toward the measurement.
   • Reduce uncertainty based on the new information.

In the example, a filter starting with a very uncertain estimate initially relies heavily on GPS. As it gains confidence in the model, the gain decreases and eventually stabilizes. The balance between prediction and measurement is determined by the model’s and sensor’s noise levels.

Multiple dimensions

Real systems often require estimating several quantities at once, including quantities that sensors do not measure directly.

• A covariance matrix represents uncertainty in multiple state variables and how their errors relate.
• For a car, position and velocity errors become linked because predicted position depends on velocity.
• As a result, a GPS measurement of position can also help correct the velocity estimate, even though GPS did not directly measure speed.

The equations become matrix-based, but the basic process remains the same: predict the state and its uncertainty, then use measurements to update both.

Applications and limitations

• The video cites navigation, phones combining GPS with accelerometer and gyroscope data, and electric cars estimating battery charge.
• The basic Kalman filter assumes the system’s prediction and measurement relationships are linear. Nonlinear systems may require:
  • An extended Kalman filter, which approximates the system as linear near the current estimate.
  • An unscented Kalman filter, which tracks selected points through the nonlinear system.
  • Particle filters, which represent uncertainty using many samples rather than one Gaussian distribution.
• Incorrect noise assumptions can also cause failure. If the model is treated as more reliable than it really is, the filter may give measurements too little weight and become confidently wrong. Choosing realistic process- and sensor-noise values is therefore important.

Main takeaway: A Kalman filter continually tracks its estimate, its uncertainty, and the expected reliability of new information. It uses those uncertainties to combine sources appropriately, often producing a better estimate than either source can provide alone.

Speakers and sources featured

• Speaker: One unnamed narrator or voiceover; no additional speakers are identified.
• Information sources discussed: System-motion models or predictions; sensor measurements, including sextant and star observations and GPS; and, in examples, accelerometer and gyroscope readings.