Location estimation for an autonomously guided vehicle using an augmented Kalman filter to autocalibrate the odometry

A Kalman filter using encoder readings as inputs and vision measurements as observations is designed as a location estimator for an autonomously guided vehicle (AGV). To reduce the effect of modelling errors an augmented filter that estimates the true system parameters is designed. The traditional way of reducing these errors is by fictitious noise injection in the filter model. The main problem with that approach however is that the filter does not learn about its bad model, it just puts more confidence in incoming measurements and less in the model. As a result the estimates will drift and the covariance grow rapidly between measurements causing these to be fused at a very high gain. This not only leads to a: very "bumpy" behavior of the estimates and a high sensitivity to measurement noise but will also lead to large estimation errors in the absence of measurements. The taken approach offers a better suppression of vision measurement noise and a better performance in the absence of vision measurements.