50.1 Why Kalman filters?

Source: rlabbe

We work with 2 sources of data:

• Sensor measurements
• Our own predictions (based on knowledge of system behaviour)

Sensors are noisy, don’t give perfect information

• Simple solution: to average readings
• However, this doesn’t work when
  • the sensor is too noisy
  • data collection not possible

The prediction, however, is also susceptible to noise (the world is noisy, outside/unaccounted for influences).

In short: “Knowledge is uncertain, and we [must] alter our beliefs based on the strength of the evidence.”

Note: the Kalman filter math is based on an idealised model of the world (errors in sensor measurement are rarely truly Gaussian).

• Kalman filter equations assume normal distribution of noise
• Performance is suboptimal if this assumption is not true