Precise Train Positioning With Unscented Kalman Filter and Low-Cost Sensors

This contribution is embedded into the challenge of track fault localization with low-cost hardware. For precise localization on the track, with an accuracy of a few decimeters for separating overlapping errors, a high resolution trajectory is needed and therefore sensor fusion is used. The commonly used combination of sensors consists of Global Navigation Satellite Systems and Inertial Measurement Units. The steps of the Kalman filter for sensor fusion are covered and afterwards the Unscented transform is described. This transform is applied to the prediction step of the Kalman filter. The implemented filters are extended by an adaptive stochastic model that applies to the observations used in the update steps. The Error-state Kalman filter and the Unscented Kalman filter are compared with and without the adaptive stochastic model with respect to their resulting root-mean-square (RMS) values. It is observed that the applied adaptive stochastic model improves performance for both filters by a small margin of 2 to 3 cm down to an RMS of 0.26 m. Meanwhile the roll angle estimation achieves deviations down to 0.1∘. Both implemented filters achieve comparable results.