Improving Vehicle Positioning and Visual Feature Estimates through Mutual Constraint

An argument is presented that a two-way relationship between position and visual information can be utilized to improve estimates of vehicle position as well as the interpretation of the surrounding environment. Simulation results for a formulation based on dynamic Bayesian networks illustrate that the belief in vehicle position is influenced by the measured visual features, while the belief in the value of said visual features is in turn influenced by GPS measurements of vehicle position. The weight of these influences, as well as the uncertainty of the resulting state estimates, depends directly upon the rate-of-change of the observed features. This information sharing approach represents a successful integration of these two technologies that leads to more plausible vehicle positioning and visual feature estimation that will be useful for purposes of automated driving and navigation systems.