ARTERIAL PROBABILISTIC TRAFFIC MODELING AND REAL-TIME TRAVEL TIME PREDICTION WITH VEHICLE PROBE DATA USING MACHINE LEARNING

Loading...
Thumbnail Image

Files

Publication or External Link

Date

2018

Citation

Abstract

This study proposes a probabilistic modeling framework for the estimation and prediction of link-based arterial travel time distribution using GPS data. The spatiotemporal correlations of the network are modeled using a directional acyclic graphical model, and several external variables in the prediction model are included to yield a better prediction in a variety of situations. This study also aims to investigate the effects of each factor on the travel time and the uncertainty associated with it.

In the proposed model, factors such as weather conditions, seasons, time of day, and day of the week are added as external variables in the graphical model. After determining the structure of the model, Streaming Variational Bayes (SVB) is used for training and parameter inference; this offers a valuable option when constant streaming data is utilized. SVB adaptively changes its parameters gradually with a lower computational cost, which makes the process less time-consuming and more efficient. The analysis shows that incorporating external variables can improve the model performance.

The data used in this study is INRIX vehicle trajectory raw data from four months - February, June, July, and October of 2015 - which makes it possible to take into account the effects of seasons and weather conditions on travel time and its uncertainty. One of the products of this study is a framework for vehicle trajectory data cleaning process including trip identification, removing outliers, and cleaning the trips data. Once the data are cleaned and ready to use, they should be mapped to the roads. The Hidden Markov Model (HMM) map matching algorithm is used to map the GPS latitude/longitude data to the Open Street Map (OSM) base map and find the traversed links between each pair of GPS points of vehicle trajectories.

Finally, a novel procedure to compare any travel time prediction model with any available commercial routing API is proposed and tested to compare the proposed model with Google API.

Notes

Rights