|M.Sc Student||Matusevich Mark|
|Subject||Off-Line Improvement of Generic Group Trackers|
|Department||Department of Computer Science||Supervisor||Professor Dan Geiger|
Trackers are information fusion systems which estimate the trajectories of objects located in some area of interest. Trackers are online algorithms which makes them suboptimal algorithms. The sub-optimality which results from the trackers being online algorithms cannot be overcome when the main objective is the estimation of the current situation in the area of interest. However, there are applications where the analysis of the history of events in the area of interest is required.
In this thesis, we aim to fix tracker errors that result from the online nature of trackers via new algorithms. We first formulate the problem of off-line improvement of a group tracker output. We then define a probabilistic model for the group tracker output improvement problem, and propose the Maximum A posteriori Probability (MAP) explanation for the model as a solution approach for the problem. We show that the problem of finding MAP for the model can be reduced to a Network Design Problem (NDP). Then, we present an algorithm for a subclass of NDP, including all instances of NDP resulting from this reduction. The algorithm is not polynomial in the worst case, but is efficient enough to allow an evaluation of our technique on large scale tracking problems. Next, we present an algorithm generating instances of group tracker output improvement problems. Finally, we evaluate the proposed solution approach using the above tools. As the result of this evaluation, we find our solution approach to be very successful: correctly restoring 84.43% connections, while introducing only 0.14% wrong ones.