Evaluation of Optimal Covariance Models for EKF-based Wheeled Mobile Robot Localization

Accurate and reliable localization is a key requirement in autonomous mobile robotics. The performance of EKF-based localization is highly dependent on the proper configuration of its process and measurement noise covariance matrices. This paper investigates the optimization of an Extended Kalman Filter (EKF)-based localization framework through three covariance tuning strategies, using Particle Swarm Optimization (PSO). The study focuses on the fusion of odometry and Absolute Positioning System (APS)measurements for successful localization of a wheeled mobile robot performing trajectory tracking.The first strategy optimizes the diagonal elements of the EKF's process and measurement noise covariance matrices (R and Q), while the second method tunes motion model coefficients ( and ) that govern uncertainty propagation from encoder data.The third approach introduces an adaptive model where angular velocity -measured by a gyroscope -is used to dynamically adjust these coefficients.Each method is evaluated on five real-world trajectories across seven fitness modes using RMSE, MAE, and MAXE metrics.Results show that all three strategies outperform the baseline EKF.Among them, the method based on tuning the process and measurement noise covariance matrices delivers the most consistent improvements, while the gyro-adaptive method achieves peak performance in highly dynamic scenarios.However, the latter displays higher variance, motivating future use of nonlinear models for improved robustness.The findings highlight the benefits of optimization-driven and context-aware tuning in enhancing EKF localization performance.