| In this thesis, a simple yet effective technique is presented for increasing the accuracy of multi-target tracking algorithms with a focus on sequential Monte-Carlo implementations of random finite set-based approaches. This technique, referred to throughout this work as an interactive likelihood, exploits the spatial information that exists in any given measurement, reducing the need for data association and allowing for more target interaction thereby increasing overall tracking accuracy.;The interactive likelihood is constructed entirely within the random finite set framework and is integrated with a multi-Bernoulli filter. In addition, a state-of-the-art deep neural network for pedestrian detection is combined in a novel way with the multi-Bernoulli filter and interactive likelihood in order to obtain a very general and flexible random finite set-based multi-target tracking algorithm.;The performance of the algorithm is evaluated in a number of publicly available datasets (2003 PETS INMOVE, AFL, and TUD-Stadtmitte) using standard, well-known multi-target tracking metrics (OSPA and CLEAR MOT). |
