Particle Filtering: The Need for Speed

Abstract

The particle filter (PF) has during the last decade been proposed for a wide range of localization and tracking applications. There is a general need in such embedded system to have a platform for efficient and scalable implementation of the PF. One such platform is the graphics processing unit (GPU), originally aimed to be used for fast rendering of graphics. To achieve this, GPUs are equipped with a parallel architecture which can be exploited for general-purpose computing on GPU (GPGPU) as a complement to the central processing unit (CPU). In this paper, GPGPU techniques are used to make a parallel recursive Bayesian estimation implementation using particle filters. The modifications made to obtain a parallel particle filter, especially for the resampling step, are discussed and the performance of the resulting GPU implementation is compared to the one achieved with a traditional CPU implementation. The comparison is made using a minimal sensor network with bearings-only sensors. The resulting GPU filter, which is the first complete GPU implementation of a PF published to this date, is faster than the CPU filter when many particles are used, maintaining the same accuracy. The parallelization utilizes ideas that can be applicable for other applications.

MapReduce particle filtering with exact resampling and deterministic runtime

“…27 Since the problem size remains fixed, we are actually quantifying strong scaling [38]. 28 The review also highlights challenges associated with, for example, multiple processors generating independent random number sequences, discusses the relative merits of using floating-point and fixed-point numbers and points to papers discussing architecture-specific issues (e.g., in [39][40][41][42]). 29 More mathematically, assume the ith particle has a weight (before resampling) of w i and the jth member of the new population is resampled as a copy of the ith particle with probability of = i ∈I j w i .…”

Fast, parallel implementation of particle filtering on the GPU architecture

“…We intend to generate these random sequences on the GPU device instead of the CPU to spare repetitive data transfer between the main memory of the system and the global memory of the device as recognized in [10]. The distribution of the random numbers in the resampling is critical on the quality of the estimation.…”

“…There have been some former implementations to parallel architectures [10][11][12][13][14][15][16][17][18][19]. In [11], an implementation strategy is proposed which is parallel; however, it cannot maintain the local connections of the particles.…”

“…The simple solution is creating random numbers on the CPU using any of the many implemented reliable random libraries proposed in [10]. However, these researchers are aware of the great disadvantage of this technique, mainly the huge delay raised by data transfer.…”

Parallel Implementation of Particle MCMC Methods on a GPU

“…particle filtering) [Gordon et al, 1993, Doucet andJohansen, 2011] may be implemented and combined on cheap, but high performance "Graphics Processing Unit" (GPU) cards. Related work on this topic includes the papers by Hendeby et al [2010], Lee et al [2010].…”