Estimating traffic intensity of cyclists using flow interpolation

Abstract

Literature shows that there is still a lack of objective, quantitative information about cycling traffic for urban researchers and planners. At the moment, gravity models are the most standard models for traffic prediction. However, these models have some difficulties fitting to local measurements. Machine learning algorithms can fit to local measurements very well and have become viable in recent years due to the increase in computing power of modern hardware. Induction loops and manual counting by hand have traditionally been often used methods for (cycling) traffic counting and traffic data collection. Due to the increase in smartphone usage, GPS data has become a viable alternative to these traditional methods in recent years. Therefore, there lies a lot of potential in combining new and upcoming data sources for traffic information such as GPS with more conventional data sources such as traffic counts using machine learning. This research aims to investigate if and how cyclist traffic intensity can be estimated using machine learning algorithms to combine GPS tracks and local traffic counts. The municipality of Tilburg in the Netherlands is selected as the scope for this research, because of the availability of data for this area.

Possible input features for the machine learning algorithms were based on literature. GPS cyclist intensities, spatial distance, road surface type, the width of roads and attractivity were found to be suitable input features. The Fietsersbond network was chosen over the OpenStreetMap network, since the former contained most of the possible input features. GPS tracks from the B-Riders and Fietstelweek were mapmatched to the network to provide traffic intensities to combine with the traffic counts, which are based on the ’Fietstelprogramma Gemeente Tilburg’. Seven different machine learning regression algorithms are tested, and their outcomes were assessed using K-fold and Leave-one-out cross-validation.

There was found to be very little correlation between the GPS cyclist intensities and the traffic counts. The outcomes of every single machine learning algorithm show that the B-Riders and Fietstelweek GPS data are unsuitable for estimating the traffic intensity of cyclists using flow interpolation since all of them have r2 scores that are zero or negative. Future research, using more extensive and less biased GPS data samples, could provide further insight into the possibilities of combining GPS tracks and local traffic counts to estimate cyclists traffic intensity on a network.