Multi-prediction particle filter for effcient memory utilization

The sampling importance resampling particle filter (SIR PF) is a common tool for nonlinear/non-Gaussian state estimation. The SIR PF is a memory-hungry algorithm, and the estimation accuracy is better with more particles (memory). However, the SIR PF does not utilize the memory effectively. In this paper, we propose a multi-prediction (MP) PF with two-stage resampling to use memory effectively. At similar accuracy, proposed MP-PF gives 74% and 49.5% memory reduction with 4.1% and 1.1% performance loss respectively in our experiments. With equal memory requirement, proposed MP-PF can improve the estimation accuracy.