The Bayesian methodology is able to deal with a number of challenges in object tracking, especially with uncertainties in the system dynamics and sensor characteristics. However, model complexities can result in non-analytical expressions which require computationally cumbersome approximate solutions. In this thesis computationally efficient approximate methods for object tracking with complex models are developed.
One such complexity is when a large group of objects, referred to as a crowd, is required to be tracked. A crowd generates multiple measurements with uncertain origin. Two solutions are proposed, based on a box particle filtering approach and a convolution particle filtering approach. Contributions include a theoretical derivation for the generalised likelihood function for the box particle filter, and an adaptive convolution particle filter able to resolve the data association problem without the measurement rates. The performance of the two filters is compared over a realistic scenario for a large crowd of pedestrians.
Extended objects also generate a variable number of multiple measurements. In contrast with point objects, extended objects are characterised with their size or volume. Multiple object tracking is a notoriously challenging problem due to complexities caused by data association. An efficient box particle filter method for multiple extended object tracking is proposed, and for the first time it is shown how interval based approaches can deal efficiently with data association problems and reduce the computational complexity of the data association. The performance of the method is evaluated on real laser rangefinder data.
Advances in digital sensors have resulted in systems being capable of accumulating excessively large volumes of data. Three efficient Bayesian inference methods are developed for object tracking when excessively large numbers of measurements may otherwise cause standard algorithms to be inoperable. The underlying mechanics of these methods are adaptive subsampling and the expectation propagation algorithm.
