Can priors be trusted? Learning to anticipate roadworks

This paper addresses the question of how much a previously obtained map of a road environment should be trusted for vehicle localisation during autonomous driving by assessing the probability that roadworks are being traversed. We compare two formulations of a roadwork prior: one based on Gaussian Process (GP) classification and the other on a more conventional Hidden Markov Model (HMM) in order to model correlations between nearby parts of a vehicle trajectory. Importantly, our formulation allows this prior to be updated efficiently and repeatedly to gain an ever more accurate model of the environment over time. In the absence of, or in addition to, any in-situ observations, information from dedicated web resources can readily be incorporated into the framework. We evaluate our model using real data from an autonomous car and show that although the GP and HMM are roughly commensurate in terms of mapping roadworks, the GP provides a more powerful representation and lower prediction error.