Thursday 28 Mar 2013: Variational Bayesian and belief propagation based multi-target tracking
Ata Ur-Rehman - Loughborough University
Harrison 106 14:00-15:00
Multi-target tracking is a challenging problem which has occupied various researchers. It more generally has many applications, such as surveillance, intelligent transportation,
submarine tracking, animal tracking for behavioural analysis and human computer interfacing. A successful multi-target tracking system requires a reliable identification of targets, which can be
achieved by adopting an appropriate data association technique.
We have recently developed a novel data association technique which assigns multiple measurements to a single target in a two stage process. In the first stage, measurements originating from all the targets are
grouped by using variational Bayesian (VB) clustering and then at the second stage these clusters are assigned to targets by using a belief propagation (BP) method. The advantage of using VB is that it
automatically determines the number of clusters which can fit the measurements. This is very helpful in multi-target tracking: a case where the number of targets is unknown and remains changing. By using
BP, our technique provides a solution to assign multiple clusters to a target.
The aim of the talk is to present our recent work on robust data association technique for multi-target tracking. The focus of the presentation will be: a) How the variational Bayesian approach is used to
group the measurements originating from different targets? b) How the new belief propagation technique is developed to assign groped measurement to targets? c) How the new data association technique is
used within the multi-target tracking framework?