Real-Time Pose Estimation for Autonomous Vehicles Using Probabilistic Landmark Maps and Sensor Fusion

Abstract

This study introduces a robust and accurate method for estimating autonomous vehicle position, facilitating safe navigation in urban and highway settings. The proposed technique employs a probabilistic particle filter framework, which, unlike approaches constrained by Gaussian assumptions, represents probability densities as samples, enabling more flexible position estimation. A key innovation lies in integrating a finely tuned Unscented Kalman Filter (UKF) to fuse radar and lidar data specifically for robust detection of pole-like static landmarks, whose positions and associated uncertainties are probabilistically modeled within an offline reference map. The particle filter leverages Bayesian filtering, associating UKF-derived landmark observations with this probabilistic map to refine the vehicle's pose. Broad simulation tests validate the method's effectiveness, achieving a mean localization error of approximately 11 cm in both longitudinal and lateral directions. Furthermore, the system demonstrates robustness, maintaining localization accuracy below 30 cm even with landmark position uncertainties up to 2 meters, and confirms real-time capability exceeding 100 Hz. These findings establish the approach as a reliable and precise solution for autonomous vehicle localization across various scenarios.