Thesis subject

Sensor fusion for increasing accuracy of localisation by means of dead reckoning

Poultrybot currently uses a particle filter for localisation, where dead reckoning is used to make a prediction of the location and a laser scanner to correct this prediction. However, it is expected that inside a poultry house, relying on the laser scanner only is not sufficient, so that a more accurate prediction was desired. Therefore, the aim of this work was to increase the accuracy of the prediction based on dead reckoning.

From literature, several approaches were found to improve data accuracy, and a Kalman filter was implemented to combine estimates from wheels, Xsens IMU and control inputs. For this purpose, all data was first converted into a rotation and a displacement. Next, the major feature of this Kalman filter was that a threefold loop was used to integrate all data, starting from the control input, and adding movement of front and rear wheels, as well as some input of the IMU.
This method was evaluated on 3 types of datasets, varying from straight lines via rectangular paths up to advanced trajectories. It was shown that accuracy was lowered with about 50%, and deviation of the real path stayed within 0.5 meter. This result met the required accuracy for the prediction of the particle filter, but is expected to improve further if more accurate data is used.