Optimal Fusion of Sensors
This thesis deals with the problem of fusing and managing data concerning the state or identity of a given object. Focus is put on the challenges occurring within the field of mobile robot navigation. The main problem here will often be to keep track of the position and orientation of the robot within some global frame of reference using a wide variety of sensors providing odometric, inertial and absolute data concerning the robot and its surroundings.

Kalman filters have for a long time been widely used to solve this problem. However, when measurements are delayed or the mobile robot is inaccurately modelled some interesting problems arise. In the thesis different filter designs are evaluated and compared. A new method for dealing with delayed measurements by extrapolating these through the delay period is introduced and an augmented filter is developed that can reduce the effect of modelling errors due to inaccurately known system parameters. Further, a new method for determining the process noise matrix for Kalman filters on mobile robots is introduced and shown to be more robust towards modelling uncertainties than traditional methods. The method is based on the assumption that modelling errors constitute the most significant error source in the filter and requires a rough estimate of the size of the errors.

Finally, the problem of establishing the identity of objects using multiple sets of evidence is addressed. Here, methods for representing and fusing identity evidence are presented and evaluated, and through a case study some interesting points concerning the practical use of these methods are made. Not only the fusion but also the scheduling (or planning) of the sensors are treated and it is shown how a simple ad hoc sensor planning approach can outperform more complex ones.