Arterial travel time distribution estimation and applications.

Abstract

Travel time estimation on signalized urban arterials has been one of the biggest challenges in transportation engineering. This thesis focuses on the characterization of arterial travel times by estimating the travel time distributions and collecting GPS data from probe vehicles to predict travel times. The main factors that affect travel time patterns on an arterial link roughly fall into four categories: geometric structure, driving behavior, signal control and traffic demand. Four states of travel time for through-through vehicles are defined. State 1 (non-stopped) and State 3 (stopped) can be approximated using mixture normal densities while State 2 (nonstopped with delay) and State 4 (stopped with delay) can be approximated with uniform distributions. When prior travel time data is available, travel time distributions could be estimated by EM algorithm. Otherwise, they can be estimated based on signal control and geometric structure of the arterial. Link travel times are then extended to route travel times. A method based on Markov Chain is proposed to estimate mean route travel time. Results suggest that the proposed method can capture the relationship of link travel times well and provide an accurate estimation of mean route travel time. Combined with travel time data collected from GPS probe vehicles, a real-time traffic condition identification approach based on Bayes theorem is proposed. Numerical examples show a single GPS probe is able to identify real-time traffic condition successfully in most cases. In addition, GPS travel times can also be used to refine the existing travel time distributions using Bayesian update. Finally, a comprehensive case study based on the NGSIM Peachtree Street Dataset is demonstrated. Travel time distributions estimated from signal timing and the geometry iii are considered as prior distributions. Traffic condition identification process is performed and the probability one travel time sequence belongs to each traffic condition is calculated. Data are then classified according to the posterior probabilities. Finally, a Bayesian update is run to calculate posterior distributions under each traffic condition combining with classified data. This update process can be repeated iteratively when new GPS data are available. The results obtained from Bayesian update are also compared to those estimated from EM algorithm. Overall the EM algorithm fits the data better than Bayesian update. However, sometime Bayesian approach could reflect the real world situation when some data is missing while EM does not.