In this paper, we propose a novel approach based on Bayesian network to predict a moving object's future location under uncertainty. The approach includes space-partitioning schemes, popular region extraction, transformation of trajectory sequence and region sequence, frequent sequential pattern mining and the Bayesian network construction. Popular regions are used to approximate a moving object's trajectory sequences. The analyzers could determine the regions they are interested in and the system could choose the frequent region patterns including these regions to construct the Bayesian network. The popular regions will be regarded as random variables of the Bayesian network and the traversal paths of regions are used to construct the arcs between nodes of the Bayesian network. The local probability distribution at each node is obtained from the empirical data. We propose several algorithms to transform the trajectory information into the Bayesian network structure. The experiment shows that the Bayesian network allows us to perform inference and get the probabilities of all possible states of an unobserved node under the current observed data.