Robust Particle Filter for Lane-Precise Localization

Abstract

In this work, we present a method for lane-precise localization in downtown scenarios based on a geometric map and sensors present in a current production vehicle. In detail, we use low-cost GPS, odometry, lane-marking detection based on a camera, objects detected by a front radar, and events from a blind spot monitoring system. The proposed combined weight update and sampling step in a particle filter reduces the required particle density for tight measurement likelihoods and additionally improves robustness against map errors. We extend this approach to allow for mixtures of uniform and normal distributions for the intermediate belief estimate and the likelihood to make the system more robust against sensor outages. Simulations show the improved reproducibility of the new method while evaluation on logged data from downtown drives show that we can reduce the error probability from around 24% to around 0.5% compared to traditional importance weighting – while at the same time keeping system availability stable around 95%.