Title

Parallel and distributed algorithms for data association and application to multitarget tracking

Date of Completion

January 1997

Keywords

Engineering, Electronics and Electrical

Degree

Ph.D.

Abstract

In the first part of this dissertation, we developed several novel and highly-efficient parallelizations of an existing serial multitarget tracking algorithm based on an Interacting Multiple Model (IMM) state estimator embedded into the 2 D assignment framework. The parallelizations developed were for both a distributed-memory high-performance computer (HPC) and a general-purpose shared-memory MIMD multiprocessor. For a sparse Air Traffic Surveillance (ATS) problem, the results of our work show that: (i) our coarse-grained shared-memory parallelization across the numerous tracks found in a multitarget tracking problem is robust, scalable, and can realize superlinear speedups, unlike previously proposed fine-grained parallelizations, and (ii) a SPMD distributed-memory parallelization using relatively simple task allocation algorithms has excellent performance and can realize near linear speedups.^ In the second part of this dissertation, we developed a novel dynamically adaptable m-best 2 D assignment algorithm and multi-level parallelization for a shared-memory multiprocessor. The m-best 2 D assignment algorithm is more efficient than the "best" m-best 2 D assignment algorithm currently in the literature, especially in dynamic multitarget tracking environments. For both simulated data and the same ATS problem, the results of our work show that: (i) a non-intrusive 2 D assignment algorithm switching mechanism enables the numerous 2 D assignment problems generated in the m-best assignment framework to be efficiently solved, and (ii) a multi-level parallelization of the partitioning task and the data association interface task enables many independent and highly parallelizable tasks to be executed in parallel.^ In the third part of this dissertation, we developed a novel m-best S D assignment algorithm and its shared-memory parallelization. To date, there has been no precedent to our m-best S D assignment algorithm. Some of the more novel aspects of this work include: (i) an efficient localization solution via a coarse-grained parallelization of the numerous static SD assignment problems generated in the m-best assignment framework, (ii) the formulation of higher-level composite (super) measurements and their corresponding joint event probabilities, and (iii) an efficient tracking solution via a series of dynamic 2 D assignment problems based on the composite measurement lists. We demonstrate m-best S D on a passive sensor multitarget tracking problem consisting of an unknown number of targets (emitters) based on multiple time samples of measurements originating from multiple high frequency direction finders. ^