Bayesian network movement model
Andreas E Züfle · SIGSPATIAL Special · 2016
To unify and integrate various types, resolutions, and levels of uncertainty of user location data, we employ a Bayesian learning approach, which allows to iteratively add new knowledge to refine a model describing the motion of an object in space and time. By explicitly modelling uncertainty, we maintain a notion of data reliability, such that the confidence of any query and mining results can be assessed. Based on this motion model for individual users, we present our algorithms for estimating the continuous location of users based on sparse observations. Our approach uses a global traffic model using a Markov-chain as an apriori-model, which is learned from all available historic trajectory data. Starting from this apriori-model, we use observations of individual users to adapt this model, using a forward backward approach to add new knowledge to the model. This yields as user-specific aposteriori-model, which captures information of their observation, and uses the apriori-knowledge to model the error and uncertainty in-between these observations. As verified by our empirical study on real trajectory data, this model allows to predict the location of objects in-between discrete observation much more accurate than competing approaches, thus significantly reducing the uncertainty in spatio-temporal data.