Positional Forecasting From Logged Training Data Using Probabilistic Neural Networks

摘要

This paper presents methodology focused on learning movement tendencies using GPS data and time stamps to forecast future movement locations. It is argued that people move throughout regions of time in established, but variable patterns, and that a person's normal movement can be learned by machines. Location extraction from raw GPS data in combination with a probabilistic neural network (PNN) is proposed for learning human movement patterns. Using time as an input to a PNN over a distribution of data, normal tendencies of movement can be forecasted by analyzing the probabilities of a target being at projected locations within a set of frequented locations. In addition to time, locations a target has visited during a specific, discrete time region can be used as input to a PNN. This gives a PNN a context with which to forecast the target's movement given its current localized movement patterns. Ultimately, the results produced are probabilities that a target will be at a certain location as well as the time window in which it will be there.