MCMC data association for multitarget tracking

In this thesis, we describe a new Monte Carlo methods based framework for multitarget tracking problems and three MCMC algorithms for data association problems. The framework uses MCMC samplers to estimate data association solutions. System state is maintained and updated with particle filters.;In solving the data association problems, we developed three MCMC sampling algorithms. In the first algorithm, a restricted data association field is constructed under the same assumptions of JPDAF. All data association samples generated and accepted comply with a gating validation matrix and the assumptions. In the second algorithm, the sampling processes is done in a relaxed data association field where samples comply with the gating validation matrix, but not the strict assumptions. The data association probabilities are redefined in accommdate to the relaxed assumptions. In the third algorithm, samples are generated from a data association field that are not restricted to either the assumptions and the gating validation matrix. A new proposal function to generate the Markov Chain is developed.;Because the framework only uses Monte Carlo methods, it does not require strict assumptions for the system models. The system is also highly configurable. By choosing the right parameters, the framework can be easily adapted to applications with different assumptions and system models. The algorithms are completely interruptable so are very feasible for real-time applications. Track initiation and elimination processes are included in the data association algorithms which makes the framework a complete one. Experimental results show that the performance is acceptable. For off-line multitarget tracking problems, the system can achieve close-to-optimal solutions.