Implementation and performance of a parallel multitarget tracking particle filter

Presented are results on a design, implementation, and performance evaluation of parallel multitarget tracking particle filters using a cluster of computers as a parallel computing environment. The parallel algorithms developed are based on the Joint Multitarget Probability Density (JMPD) algorithm for multiple target tracking. Since the resampling step is crucial in any particle filter parallelization, three techniques are implemented and experimented on a cluster for resampling - centralized, distributed with proportional allocation, and distributed with nonproportional allocation. Independent partitioning, coupled partitioning, and adaptive partitioning for importance sampling within the JMPD filter are also implemented in parallel. Results from a comprehensive experimental study with simulated large-scale multitarget tracking scenarios (up to 50 targets) for performance evaluation are presented. A comparative analysis is provided based on the obtained experimental data.