Marie-Curie fellowship from 2014/1/1 until 2016/5/31
Fernando José Iglesias García
Pranab Mandal & Anton Stoorvogel
Traditional tracking solutions are not suitable in complex non-linear systems. Due to the lack of a general closed-form solution suitable for generic non-linear models, numerical approximations must be used. Sequential Monte Carlo (SMC) methods or particle filters have become during the last two and half decades increasingly popular due to their flexibility to cope with non-linearities and provable strong theoretical guarantees.
The cost of the flexibility of SMC comes in their high computational demand, as their principle of operation consists of generating samples resulting from numerical simulations. In order to circumvent the inefficiencies of SMC dealing with high dimensions, the use of SMC in conjunction with Markov chain Monte Carlo (MCMC) methods has been proposed in the literature during the last decade.
In large harbour areas we need to track the movement of hundreds of ships in real-time. To this end, diverse sources of information including multiple radar systems and GPS are available. The current technology of SMC ain conjunction with MCMC is simply not able to tackle this problem. Specifically, multiple objects require a large number of particles and therefore a huge increase in computational effort. Furthermore, the large number of sensors make the measurement update of the particles numerically challenging.
The goal of this research is to develop efficient algorithms able to address multiple object tracking problems like the one described. This shall be done by means of techniques such as genetic algorithms to improve the MCMC-based particle filter and ensure fast convergence even when large amounts of sensor data must be processed. Moreover, by using modern processing tools such as random finite sets and Langevin MCMC we will show how to address multiple object tracking much more efficiently.