Travel demand modeling based on cellular probe data

A comprehensive travel demand modeling methodology using cellular probe data is presented in this thesis, which includes a static daily O-D estimation model, a mode share estimation model and a dynamic (time-dependent) O-D estimation model. At first, the cellular probe trajectories are obtained by recording all the signal-transition events of cellular probes to determine the trip origins and destinations. The ownership of cell phone was treated as a conditional probability depending on users' socio-economic factors available in the census data such as age, race, household income, etc. Thereinafter, the traveling population daily O-D demand was estimated via a robust Horvitz-Thompson estimator. The methodology was tested via a VISSIM simulation and results were compared with a conventional simple random sampling (SRS) method. The comparison outcome shows great potential of using cellular probe data as a means to estimating O-D travel demand. The mode share estimation model consists of two major parts: offline learning and online inference. The offline learning extracts the temporal and speed features from the cellular probe trajectories. The model parameters are calibrated through the offline learning process. The online inference determines the transportation mode for individual cell phone user in a real-time manner. The methodology was tested via a VISSIM based simulation and a case study designed for both the offline learning and online inference parts. The results show the great potential of using the information of cellular probe trajectories as a means to estimating the transportation mode shares. A Kalman filter based dynamic O-D estimation and prediction model is also proposed. Not like the traditional Kalman filter based O-D estimation method, in which the link traffic counts and the link traffic assignment matrix are mostly taken into consideration in the observation equation, this method utilizes the cellular probe counts crossing the cell boundaries as the observed variables and derives an assignment matrix which assigns the cellular probe counts to subsets of the links – those links covered by the cell boundaries. By conducting the simulation and field experiments, it is observed that the model is a feasible means to estimate and predict the dynamic O-D matrix.