Content area
Full Text
A local weight mean comparison scheme for high-speed particle filters along with its corresponding architecture is presented. The proposed scheme arranges a particle index by a comparison between the relevant weight and the current mean weight so that as many high-weighted particles as possible can be firstly accessed in the next sampling step. It breaks the sequential nature of resampling and basically maintains the same accuracy as that of the systematic resampling (SR) scheme. The proposed architecture pipelines the current resampling step and the next sampling step, and improves the speed by times compared with generic architectures based on the SR scheme and the residual systematic resampling scheme.
Introduction: Particle filters (PFs) have shown great promise in hand- ling complex nonlinear and/or non-Gaussian problems requiring dynamic state estimations such as signal processing, computer vision, communication, navigation, etc. Common PFs are composed of three steps including sampling, weight calculation and resampling. Owing to their computational complexity, it is necessary, in real-time appli- cations, to turn to their hardware implementations for faster execution. Two generic hardware architectures for PFs have been developed in [1] based on the systematic resampling (SR) scheme and the residual systematic resampling (RSR) scheme, respectively. Despite a significant reduction in the memory requirement, the two architectures still suffer a slow speed from the time-consuming resampling step. And so are the cases with the architecture employing the compact threshold-based resampling scheme [2] and the architecture based on the residual resam- pling scheme [3]. Generally, the speed of current existing architectures for PFs is crucially affected by the resampling operation because of its sequential nature.
In this Letter, we propose a local weight mean comparison scheme for high-speed PFs. Once a particle weight is generated, the proposed scheme arranges the relevant particle index according to the comparison result between the weight and the current mean weight in order that as many high-weighted particles as possible can be accessed first in the next sampling step. It breaks the sequential nature of resampling and basically has no sacrifice...