Non-parametric estimation of route travel time distributions from low-frequency floating car data

• Published in: Transportation research. Part C, Emerging technologies Emerging technologies, 2015-09, Vol.58, p.343-362
• Main Authors: Rahmani, Mahmood, Jenelius, Erik, Koutsopoulos, Haris N.
• Format: Article
• Language: English
• Subjects: Automatic number plate recognition; Floating car data; Kernel estimation; Non-parametric; Route travel time distribution; Sampling bias; Transport Science; Transportvetenskap
• ISSN: 0968-090X; 1879-2359; 1879-2359
• DOI: 10.1016/j.trc.2015.01.015

•A non-parametric method for the estimation of the distribution of travel times along routes from low-frequency FCD is presented.•The method is able to accurately estimate the mean and other statistics of the travel time distribution, such as the median and various percentiles.•It identifies important biases introduced when FCD are used for the estimation of path travel times and develops efficient methods for correcting them.•The study uses ANPR data for comparison purposes (which is probably one of very few papers that have done so). It also uses ANPR data to correct for sampling biases (i.e. the sample of taxis is not representative of the population of drivers).•It proposes a computationally efficient implementation that is scalable and can support real time applications with large data sets. The paper develops a non-parametric method for route travel time distribution estimation using low-frequency floating car data (FCD). While most previous work has focused on link travel time estimation, the method uses FCD observations for estimating the travel time distribution on a route. Potential biases associated with the use of sparse FCD are identified. The method involves a number of steps to reduce the impact of these biases. For evaluation purposes, a case study is used to estimate route travel times from taxi FCD in Stockholm, Sweden. Estimates are compared to observed travel times for routes equipped with Automatic Number Plate Recognition (ANPR) cameras with promising results. As vehicles collecting FCD (in this case, taxis) may not be a representative sample of the overall vehicle fleet and driver population, the ANPR data along several routes are also used to assess and correct for this bias. The method is computationally efficient, scalable, and supports real time applications with large data sets through a proposed distributed implementation.