GPU-accelerated particle filter for multi-sensor multi-target tracking

Abstract

This paper addresses particle filtering for multi-sensor multi-target tracking implemented on a graphics processing unit (GPU) architecture. The GPU-accelerated particle filter is constructed as a distributed computation particle filter in which resampling is done in a distributed manner. The simulation result gives a comparison of the tracking accuracy and computation speed of the GPU-accelerated particle filtering to those of standard sequential CPU implementations. We analyze the effect of the resampling group size, which is the most important parameter in the distributed computation particle filter. While significant speedup is observed in GPU-accelerated particle filtering, the group size provides a trade-off between tracking accuracy and computation speed.