Using GPS Data to Understand Variations in Path Choice

A comprehensive set of GPS vehicle location data from Lexington households was analyzed to determine if such data can be helpful in improving path choice assumptions in traffic assignment models. Primarily, the portion of the data used consisted of a reconstruction of the street network and the lists of street segments in each path. Analysis was based on “matches” of trips, e.g., pairs of trips with similar origins and destinations. Matches were obtained for trips within households and for trips across households. Statistics used to compare trips in matches were a path deviation index and the percentage of identical links. It was found that the path chosen on a trip was quite sensitive to the location of the origin and destination and that the chosen path most often differed considerably from the shortest time path across the network. Paths for trips made by the same driver were very consistent over time; paths by different drivers showed more deviations even when the trip ends were the same or very similar. Recommendations are made as to how GPS data on path choice can be better collected in the future and for improvement of traffic assignment models. INTRODUCTION A path is a sequence of links (e.g., road segments) and nodes (e.g., intersections) that comprise a trip from an origin to a destination. Notions of path choice by travelers are fundamental to the traffic assignment step in travel forecasting models and to many traffic simulation models. The current methods used by planners for modeling path choice in traffic assignment have been developed largely in the absence of objective empirical evidence of actual path choices. Theories of user-optimal equilibrium assignment and stochastic multipath traffic assignment have proven quite useful to planners, but those algorithms’ underlying assumptions related to path choice have not received an adequate level of validation. Furthermore, these algorithms are most often applied to a network overlaid on a coarse zone system, and the implications of different levels of zonal aggregation on the validity of path choice assumptions are nearly unknown. There has been a recent interest in “microsimulation” for travel forecasting, which drastically reduces the level of spatial aggregation but greatly increases the amount of computation. The objective of this study is to explore the use of objective path choice data to begin to understand the differences between actual behavior and traffic assignment theory and practice. A recent data collection effort in Lexington, Kentucky (1) employed the Global Positioning System (GPS) to track vehicles over an extended period of time. The data set is unique in its comprehensiveness, involving all trips for a single vehicle from 100 households and 216 drivers over a one-week period of time. More than 3000 trips are represented in this sample. These data allows analysis of actual path choices by drivers, analysis of the stability of path choice for the same driver taking the same trip at different times of day and different days of week, and comparisons of different drivers taking essentially the same trip. The data also allows comparisons of paths of trips with similar, but not identical, trip ends. Of particular importance to this study is a set of derived data from the raw GPS data from Lexington that identifies the sequence of street segments for each trip. In network terms, each segment is a link. The Lexington network consists of about 13,000 separate street segments (or links) representing virtually every road in the metropolitan area. The raw GPS coordinates (longitude-latitude) had been matched to street segments so all the links in a path can be identified. The Lexington data can be used to compare exact paths for sets of trips that have identical or similar ends. These sets of trips are referred to here as “matches”. Given the large number of trips in the Lexington data, Jan, Horowitz and Peng 2 there are many valid matches. Matches can then be classified as to the type of path deviations seen, compared to the theoretical shortest path, and analyzed quantitatively for the degree of deviation.