Position estimation of mobile robots considering characteristic terrain properties

Due to the varying terrain conditions in outdoor scenarios the kinematics of mobile robots is much more complex compared to indoor environments. In this paper we present an approach to predict future positions of mobile robots which considers the current terrain. Our approach uses Gaussian process regression (GPR) models to estimate future robot positions. An unscented Kalman filter (UKF) is used to project the uncertainties of the GPR estimates into the position space. The approach utilizes optimized terrain models for estimation. To decide which model to apply, a terrain classification is implemented using Gaussian process classification (GPC) models. The transitions between terrains are modeled by a 2-step Bayesian filter (BF). This allows us to assign different probabilities to distinct terrain sequences, while taking the properties of the classifier into account and coping with false classifications. Experiments showed the approach to produce better estimates than approaches considering only a single terrain model and to be competitive to other dynamic approaches.