Non-parametric anomaly detection in trajectorial data

摘要: The main part of this licentiate thesis concerns parallelization recursive estimation methods, both linear and nonlinear. Recursive deals with the problem extracting information about parameters or states a dynamical system, given noisy measurements system output plays central role in many applications signal processing, identification, automatic control. Solving Bayesian is known to be computationally expensive, which often makes methods infeasible real-time for problems large dimension. As computational power hardware today increased by adding more processors on single chip rather than increasing clock frequency shrinking logic circuits, most powerful way improving execution time an algorithm. It has been found that several optimal filtering are suitable parallel implementation, certain ranges sizes. concluded from experiments substantial improvements can achieved performing "tailor"-made parallelization, compared straightforward implementations based multi-threaded libraries. For suggested parallelizations, speedup number cores have provided up 8 times double quad-core computer. evolution computer architectures unfolding rapidly, same will become available. developed do not, course, scale infinitely, but definitely exploit harness some next generation platforms, allowing state applications.