Unified mixture-model based terrain estimation with Markov Random Fields

Abstract

This paper proposes a Markov Random Field (MRF) representation for sensor and terrain information fusion in a 2.5D map. Unlike in the previous works, the proposed MRF formally models the sensor pose and measurement uncertainties, thus allowing the measurements to be appropriately fused with terrain information. Additionally, the MRF's graphical modelbased representation allows for an easy modification to the probabilistic dependencies among variables, permitting a more flexible and general model including terrain spatial correlations to be studied. The use of an MRF representation also makes it easier to perform factorization and inference on any variable subset of interests. Results show that the addition of a terrain MRF model not only helps reduce the estimation error, but also serves as a basis for terrain property characterization, which is useful for future terrain analyses such as traversability assessments in ground robot navigation.