Neural Terrain Classification and Fuzzy Adaptive Covariance Tuning for Improved Outdoor Mobile Robot Localization

Accurate and reliable localization is essential for autonomous outdoor mobile robots operating with low-cost sensors under varying terrain conditions. In such environments, the performance of an Extended Kalman Filter (EKF) strongly depends on the appropriate selection of process- and measurement-noise covariance matrices, while fixed covariance parameters often fail to represent terrain-dependent uncertainty. The terrain-specific covariance parameters obtained offline using PSO are incorporated into a Mamdani-type fuzzy inference system, which adapts the EKF covariance matrices online according to the estimated terrain type and its associated class-dependent reliability coefficient. The framework was experimentally validated on a wheeled mobile robot using commanded velocity inputs, low-cost Global Positioning System (GPS) measurements, and Real-Time Kinematic Global Navigation Satellite System (RTK-GNSS) reference trajectories collected over multiple outdoor terrain types. Compared with a baseline EKF using fixed covariance matrices, the proposed adaptive approach reduced the average root mean square error by 9.7% and the mean absolute error by 10.2%, with the largest improvements observed on deformable terrains. The results demonstrate that terrain-aware covariance adaptation provides an effective strategy for improving outdoor localization performance using low-cost sensing systems.