Markov Random Fields for Basic Scene Analysis of Occupancy Grids

We develop an approach to updating the representation of the environment in near proximity to a platform based upon limited-quality onboard sensor readings. Our method builds upon occupancy grid mapping techniques, and uses a Markov random field as the probabilistic modeling paradigm for resolving discrepancies between neighboring cells in the occupancy grid. This approach is readily implemented with intermittent data of obstacle reports, which is a behavior typically encountered when using forward look sonars operating on unmanned underwater vehicles. We demonstrate the approach through simulation studies of sonar performance in environments with multiple obstacles. Using receiver operator characteristic analysis of the resulting filtering scheme, we show how the method can be tuned to improve performance based upon the expected environmental complexity and sonar performance characteristics. Numerical results from the simulations are presented to show how the method compares to conventional Bayesian filtering.