In this paper we discuss a system that can learn personal maps customized for each user and infer his daily activities and movements from raw GPS data. The system uses discriminative and generative models for different parts of this task. A discriminative relational Markov network is used to extract significant places and label them; a generative dynamic Bayesian network is used to learn transportation routines, and infer goals and potential user errors at real time. In this paper we focus on the basic structures of the models and briefly discuss the inference and learning techniques. Experiments show that our system is able to accurately extract and label places, predict the goals of a person, and recognize situations in which the user makes mistakes (e.g., taking a wrong bus). (permanent copy, local copy)
Published in Modeling Others from Observations (MOO 2005)
WR-07