Predicting cycling risk at intersections with natural cycling data for speed-controlled e-bikes

Master thesis (2022)

Authors

D. Landré Mechanical Engineering

Contributors

J.K. Moore Biomechatronics & Human-Machine Control - Mechanical, Maritime and Materials Engineering (supervisor 1)
V.L. Knoop Transport and Planning - CEG (supervisor 2)
L. Marchal Crespo Human-Robot Interaction - Mechanical, Maritime and Materials Engineering (supervisor 2)
Bart Oor (supervisor 2)
Faculty
Mechanical Engineering
Copyright: © 2022 Daniël Landré
Machine learning E-bike IoT data Cycling accident analysis Intersection safety Natural cycling data
More Info
expand_more

Published Date

16-09-2022

Language

English

Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Faculty

Mechanical Engineering

Abstract

The rapid adoption of e-bikes as an alternative mode of transportation to automobiles gives rise to new methods of safety regulations for cyclists. Modern e-bikes feature Internet of Things (IoT) modules capable of collecting and sending cycling data that can be used for traffic safety analysis. This study explores the potential of using cycling data to detect dangerous intersections and then implement local low-speed areas using geofencing.

A natural cycling experiment with 10 participants is conducted, where data is collected similar to the capabilities of e-bike IoT modules (GNSS, power and IMU data) and participants are asked to cycle through Delft, The Netherlands. The data is combined with a traffic accident dataset from the Dutch government, where intersections with accidents are labelled dangerous. The cycling data is separated into separate intersection approaches, based on positional data. Metrics based on power and cadence data show the most significant statistical difference (p-values <0.02) and the largest effect size (Cohen's d >0.80) at dangerous intersections. A binary classification model is trained on the dataset, which can correctly predict whether an intersection is dangerous or safe with an accuracy of 68,2%, a specificity of 50,0%, and a sensitivity of 24,2%. This prediction is based on the metrics of a single intersection approach.

To investigate the viability of using geofencing to implement low-speed areas at designated dangerous intersections, a simulation is carried out to determine the minimum size of a geofence that effectively slows down e-bike cyclists. The minimal effective geofence size to limit a cyclist's speed at an intersection is determined by simulating cyclists approaching a geofence for several geofence perimeters, incline levels, and wind speeds. The e-bike's motor stops supporting when the geofence is entered and the cyclist is cycling above a target speed. The minimal effective geofence size ranges from 50 to 300+ m and depends on the target speed, wind speed, and road incline.

This study shows that cycling data can be used to identify dangerous intersections. Enforcing a slower speed on intersections with local geofencing is not feasible, as the geofences have to be extremely large, showing a lot of overlap on other intersections in cities due to high intersection density. Future work can be done on the intersection risk estimation method and the feasibility of speed-limiting geofences that employ active braking.