Computational interaction models for automated vehicles and cyclists

Licentiatavhandling, 2023

Cyclists’ safety is crucial for a sustainable transport system. Cyclists are considered vulnerable
road users because they are not protected by a physical compartment around them. In recent
years, passenger car occupants’ share of fatalities has been decreasing, but that of cyclists has
actually increased. Most of the conflicts between cyclists and motorized vehicles occur at
crossings where they cross each other’s path. Automated vehicles (AVs) are being developed
to increase traffic safety and reduce human errors in driving tasks, including when they
encounter cyclists at intersections. AVs use behavioral models to predict other road user’s
behaviors and then plan their path accordingly. Thus, there is a need to investigate how cyclists
interact and communicate with motorized vehicles at conflicting scenarios like unsignalized
intersections. This understanding will be used to develop accurate computational models of
cyclists’ behavior when they interact with motorized vehicles in conflict scenarios.
The overall goal of this thesis is to investigate how cyclists communicate and interact with
motorized vehicles in the specific conflict scenario of an unsignalized intersection. In the first
of two studies, naturalistic data was used to model the cyclists’ decision whether to yield to a
passenger car at an unsignalized intersection. Interaction events were extracted from the
trajectory dataset, and cyclists’ behavioral cues were added from the sensory data. Both
cyclists’ kinematics and visual cues were found to be significant in predicting who crossed the
intersection first. The second study used a cycling simulator to acquire in-depth knowledge
about cyclists’ behavioral patterns as they interacted with an approaching vehicle at the
unsignalized intersection. Two independent variables were manipulated across the trials:
difference in time to arrival at the intersection (DTA) and visibility condition (field of view
distance). Results from the mixed effect logistic model showed that only DTA affected the
cyclist’s decision to cross before the vehicle. However, increasing the visibility at the
intersection reduced the severity of the cyclists’ braking profiles. Both studies contributed to
the development of computational models of cyclist behavior that may be used to support safe
automated driving.
Future work aims to find differences in cyclists’ interactions with different vehicle types, such
as passenger cars, taxis, and trucks. In addition, the interaction process may also be evaluated
from the driver’s perspective by using a driving simulator instead of a riding simulator. This
setup would allow us to investigate how drivers respond to cyclists at the same intersection.
The resulting data will contribute to the development of accurate predictive models for AVs.