Semantic Aware Resampling For Robust Particle Filter Localization

Abstract

Particle filters provide a principled Bayesian framework for robot localization under uncertainty, representing belief states through weighted sample sets updated via motion prediction and observation integration. Their effectiveness critically depends on the resampling strategy, which governs particle diversity maintenance while avoiding degeneracy. Traditional methods—multinomial, systematic, residual, and stratified resampling—have been extensively studied, yet the potential for semantic information to guide resampling decisions in feature-rich environments remains unexplored. We introduce Semantic-Stratified Importance Sampling (SSIS), a novel resampling strategy that partitions particles based on proximity to semantic feature categories rather than uniform weight intervals, aiming to preserve hypotheses across different semantic contexts. Through comprehensive evaluation comprising 9,500 independent trials, we conduct the first systematic comparison of SSIS against all four traditional methods, isolating the effects of particle count, semantic feature richness, motion noise, and observation confusion. Our rigorous 100-trial-per-configuration methodology distinguishes algorithmic differences from random variance, addressing a methodological gap in particle filter evaluation.