20210225_ロボティクス勉強会_パーティクルフィルタのMAP推定の高速手法「FAST-MAP」を作ってみたMori Ken
Particle filters (PFs) are used for the discrete approximation of dynamic and non-Gaussian probability distributions using numerous particles. Maximum a posteriori (MAP) estimation, which is a point estimation method to extract a unique state value from the probability distributions formed by a PF, functions appropriately against multimodal distributions. However, MAP entails an enormous calculation cost. Therefore, we propose a method to perform MAP estimation with a low calculation cost by compressing the information configured by PF, using adaptive vector quantization. For MAP estimation with 900 particles, the proposed method reduced the computational cost by approximately 96% compared to the conventional method and maintained the same estimation accuracy during the simulation.