What Can Eye Gaze Teach Us About Real-World Cycling? Insights From the Oxford RobotCycle Project
Extracted source text ยท arXiv 2606.04989
%%
%% This is file `sample-manuscript.tex',
%% generated with the docstrip utility.
%%
%% The original source files were:
%%
%% samples.dtx (with options: `all,proceedings,bibtex,manuscript')
%%
%% IMPORTANT NOTICE:
%%
%% For the copyright see the source file.
%%
%% Any modified versions of this file must be renamed
%% with new filenames distinct from sample-manuscript.tex.
%%
%% For distribution of the original source see the terms
%% for copying and modification in the file samples.dtx.
%%
%% This generated file may be distributed as long as the
%% original source files, as listed above, are part of the
%% same distribution. (The sources need not necessarily be
%% in the same archive or directory.)
%%
%%
%% Commands for TeXCount
%TC:macro \cite [option:text,text]
%TC:macro \citep [option:text,text]
%TC:macro \citet [option:text,text]
%TC:envir table 0 1
%TC:envir table* 0 1
%TC:envir tabular [ignore] word
%TC:envir displaymath 0 word
%TC:envir math 0 word
%TC:envir comment 0 0
%%
%% The first command in your LaTeX source must be the \documentclass
%% command.
%%
%% For submission and review of your manuscript please change the
%% command to \documentclass[manuscript, screen, review]{acmart}.
%%
%% When submitting camera ready or to TAPS, please change the command
%% to \documentclass[sigconf]{acmart} or whichever template is required
%% for your publication.
%%
%%
%\documentclass[sigconf, anonymous]{acmart}
\documentclass[sigconf]{acmart}
\settopmatter{printacmref=false} % Removes citation information below abstract
\renewcommand\footnotetextcopyrightpermission[1]{} % removes footnote with conference information in first column
\setcopyright{none}
%%
%% \BibTeX command to typeset BibTeX logo in the docs
\AtBeginDocument{%
\providecommand\BibTeX{{%
Bib\TeX}}}
\usepackage{graphicx}
\usepackage{subcaption}
%% Rights management information. This information is sent to you
%% when you complete the rights form. These commands have SAMPLE
%% values in them; it is your responsibility as an author to replace
%% the commands and values with those provided to you when you
%% complete the rights form.
% \setcopyright{acmlicensed}
% \copyrightyear{2026}
% \acmYear{2018}
% \acmDOI{XXXXXXX.XXXXXXX}
%% These commands are for a PROCEEDINGS abstract or paper.
%\acmConference[Conference acronym 'XX]%{Make sure to enter the correct
% conference title from your rights confirmation email}{June 03--05,
% 2018}{Woodstock, NY}
% %%
%% Uncomment \acmBooktitle if the title of the proceedings is different
%% from ``Proceedings of ...''!
%%
%%\acmBooktitle{Woodstock '18: ACM Symposium on Neural Gaze Detection,
%% June 03--05, 2018, Woodstock, NY}
% \acmISBN{978-1-4503-XXXX-X/2018/06}
% \setcopyright{none}
% \copyrightyear{2026}
%\acmBooktitle{Woodstock '18: ACM Symposium on Neural Gaze Detection,June 03--05, 2018, Woodstock, NY}
%%
%% Submission ID.
%% Use this when submitting an article to a sponsored event. You'll
%% receive a unique submission ID from the organizers
%% of the event, and this ID should be used as the parameter to this command.
%%\acmSubmissionID{123-A56-BU3}
%%
%% For managing citations, it is recommended to use bibliography
%% files in BibTeX format.
%%
%% You can then either use BibTeX with the ACM-Reference-Format style,
%% or BibLaTeX with the acmnumeric or acmauthoryear sytles, that include
%% support for advanced citation of software artefact from the
%% biblatex-software package, also separately available on CTAN.
%%
%% Look at the sample-*-biblatex.tex files for templates showcasing
%% the biblatex styles.
%%
%%
%% The majority of ACM publications use numbered citations and
%% references. The command \citestyle{authoryear} switches to the
%% "author year" style.
%%
%% If you are preparing content for an event
%% sponsored by ACM SIGGRAPH, you must use the "author year" style of
%% citations and references.
%% Uncommenting
%% the next command will enable that style.
%%\citestyle{acmauthoryear}
%%
%% end of the preamble, start of the body of the document source.
\begin{document}
%%
%% The "title" command has an optional parameter,
%% allowing the author to define a "short title" to be used in page headers.
\title{What Can Eye Gaze Teach Us About Real-World Cycling? Insights From the Oxford RobotCycle Project}
%%
%% The "author" command and its associated commands are used to define
%% the authors and their affiliations.
%% Of note is the shared affiliation of the first two authors, and the
%% "authornote" and "authornotemark" commands
%% used to denote shared contribution to the research.
\author{Benjamin Hardin}
\orcid{0009-0000-9938-3336}
\affiliation{%
\institution{University of Oxford}
\city{Oxford}
\country{United Kingdom}
}
\email{benjamin.hardin@cs.ox.ac.uk}
\author{Efimia Panagiotaki}
\orcid{0009-0008-0189-8904}
\affiliation{%
\institution{University of Oxford}
\city{Oxford}
\country{United Kingdom}
}
\email{efimia@robots.ox.ac.uk}
\author{Daniele De Martini}
\orcid{0000-0001-6121-5839}
\affiliation{%
\institution{University of Oxford}
\city{Oxford}
\country{United Kingdom}
}
\email{daniele@robots.ox.ac.uk}
\author{Lars Kunze}
\orcid{0000-0001-5302-1938}
\affiliation{%
\institution{University of the West of England Bristol}
\city{Bristol}
\country{United Kingdom}
}
\email{lars.kunze@uwe.ac.uk}
%%
%% By default, the full list of authors will be used in the page
%% headers. Often, this list is too long, and will overlap
%% other information printed in the page headers. This command allows
%% the author to define a more concise list
%% of authors' names for this purpose.
\renewcommand{\shortauthors}{Hardin et al.}
%%
%% The abstract is a short summary of the work to be presented in the
%% article.
\begin{abstract}
Although much is known about the physical danger of cycling situations, less is understood about the perceived danger of cycling. Furthermore, perception of danger may be filtered at a subconscious level and therefore difficult for one to self-report. To this end, these subconscious perceptions can be revealed through physiological metrics such as eye gaze.
This paper explores the perceived safety of cycling in Oxford, United Kingdom and explores the ability of wearable eye tracking glasses to produce insights about the differences in perception under different environments and events.
This paper finds that eye gaze patterns change between using bike lanes, car lanes and shared bus lanes, representing different cognitive challenges of each lane type.
This paper presents that different intersections have significantly different eye gaze patterns which may have implications for cyclist stress. Finally, eye gaze patterns differ in the presence of events such as passes and pedestrians in the road compared to when cycling with no events.
This paper draws conclusions on the benefits and limitations of using wearable eye trackers to estimate stress and cyclist workload.
\end{abstract}
%%
%% The code below is generated by the tool at http://dl.acm.org/ccs.cfm.
%% Please copy and paste the code instead of the example below.
%%
\begin{CCSXML}
<ccs2012>
<concept>
<concept_id>10003120.10003121</concept_id>
<concept_desc>Human-centered computing~Human computer interaction (HCI)</concept_desc>
<concept_significance>500</concept_significance>
</concept>
<concept>
<concept_id>10003120.10003138.10011767</concept_id>
<concept_desc>Human-centered computing~Empirical studies in ubiquitous and mobile computing</concept_desc>
<concept_significance>500</concept_significance>
</concept>
<concept>
<concept_id>10003120.10003138.10003141.10010898</concept_id>
<concept_desc>Human-centered computing~Mobile devices</concept_desc>
<concept_significance>500</concept_significance>
</concept>
<concept>
<concept_id>10010405.10010481.10010485</concept_id>
<concept_desc>Applied computing~Transportation</concept_desc>
<concept_significance>500</concept_significance>
</concept>
</ccs2012>
\end{CCSXML}
\ccsdesc[500]{Human-centered computing~Human computer interaction (HCI)}
\ccsdesc[500]{Human-centered computing~Empirical studies in ubiquitous and mobile computing}
\ccsdesc[500]{Human-centered computing~Mobile devices}
\ccsdesc[500]{Applied computing~Transportation}
%%
%% Keywords. The author(s) should pick words that accurately describe
%% the work being presented. Separate the keywords with commas.
\keywords{human-computer interaction, eye tracking, human factors, road safety, road traffic, urban infrastructure}
% \received{20 February 2007}
% \received[revised]{12 March 2009}
% \received[accepted]{5 June 2009}
%%
%% This command processes the author and affiliation and title
%% information and builds the first part of the formatted document.
\maketitle
\pagestyle{plain} % removes running headers
\section{Introduction}
Cycling continues to be a popular form of transportation that has seen significant interest in recent years due to health and environmental benefits as well as enjoyment.
In order to ensure that cycling is a safe form of transportation, it is important to understand not only situations where accidents occur, but also real time responses of cyclists to various events and environment factors. Nevertheless, it is difficult to collect real-time in-situ feedback from cyclists because they must focus full attention on the task. Many prior studies have had to rely on post-hoc or simulated environment responses from cyclists which means many situations and smaller interactions may be forgotten \cite{GADSBY2021102932, RIVERAOLSSON2023107007}. To this end, there is potential for measuring physiological metrics such as heart rate, heart rate variability, and skin response in real time and using these metrics as a proxy for stress and workload \cite{10457965, CAVIEDES2018488}. Although perhaps less explored, tracking eye gaze is another potential physiological metric that can be used to measure stress and excitement and has already shown some application to cycling studies \cite{MA202452}.
In this paper, we present findings from the RobotCycle trials in Oxford, UK conducted over several days in 2024 and 2025. We measure eye gaze fixations and horizontal gaze dispersion to understand scenarios, routes, and events that elicit the greatest reaction from the cyclist and explore the feasibility of eye gaze for measuring cyclist stress. We find that eye gaze metrics differ significantly between various intersection types, that various events that the cyclist may encounter (such as overtakes or pedestrians) produce significantly different eye gaze patterns compared to no events, and that different routes demonstrate significantly different eye gaze patterns due to varying infrastructures, traffic, and pedestrian presence. However, due to the complexity of environments and their multitude of elements which may capture visual attention, we highlight the limitations that eye gaze alone has on understanding an environment. We find that eye gaze patterns alone were not sufficient to distinguish between types of events that a cyclist faces. For these situations, object detection and segmentation is necessary. Furthermore, fixation duration and horizontal gaze dispersion alone may be insufficient for reliably estimating cyclist stress.
\section{Literature Review}
\subsection{Relationship Between Eye Gaze and Stress}
Eye gaze movements have been used extensively to evaluate stress, fatigue, and emotions \cite{yousefi2022, li2021, bend2026comparing, yang2025} across diverse domains including technical interviews, student learning, driving, and nursing \cite{behroozi2018can, Jyotsna, nemcova, ahmadi2024quantifying}. There have also been numerous methods proposed for measuring eye gaze metrics, with the most popular ones including velocity threshold (I-VT) and dispersion threshold (I-DT) methods \cite{andersson2017one, salvucci2000identifying}. Nevertheless, reliability of using eye gaze to predict psychological factors has been shown to be very dependent on the context and scenario in which the eye gaze is being measured. Intuitively, some tasks may require more scanning across a broad environment while other tasks involve deep focus on a particular point or set of points. For instance, one paper considers solving technical interview problems on a whiteboard compared to on paper and uses this as a basis of demonstration that gaze metrics vary in their meanings \cite{behroozi2018can}. Additionally, there is a question about the reliability of eye gaze in non-static environments such as walking, skipping, or jumping \cite{hooge2023robust, franchak2021adapting}. To this end, we compare eye gaze across environment factors, establishing the eye gaze baseline as the presence of no events or manoeuvres. %We analyse variability in eye gaze data to contribute to estimatesl of reliability in eye gaze metrics from a moving position.
\subsection{Risky Situations in Cycling}
%\textit{To Be completed....}
One study has found that situations that are perceived as risky generally align with the situations that are objectively risky \cite{von2020crash}. Furthermore, the study found that bike lanes did not reliably reduce crash risks and that public transport stops could increase risk. Higher traffic volume and greater intersection complexity were also related to increased crash risks, a finding that was corroborated by a study of cycling in London \cite{aldred2018cycling}.
It has been found that intersections can induce higher levels of stress for cyclists \cite{CAVIEDES2018488}. The authors also found that protected bike lanes were associated with reduced stress levels. One study also found that discontinuities in paths, such as cycle paths ending, could increase stress as well as situations when there is not enough space for bicycles \cite{su10072379}. High speed differentials between cyclists and vehicles have also been found to be stressful \cite{GADSBY2021102932}.
\section{Eye Gaze and Cycling}
The systematic review by Ma et al., 2024 presents an overview of eye gaze tracking in cycling studies \cite{MA202452}, finding that there is a lack of standardised practice for variables such as glances and fixations and highlights that fixations may not consider all of the data that cyclists use from their peripheral vision. Furthermore, the paper considers how types of workload may vary extensively by environment but this has yet to be studied.
One study considered eye gaze patterns of cyclists at intersections \cite{rupi2019}, finding that intersections which force cyclists to share a lane with vehicles induce more eye gaze patterns associated with stress. Another study found that the presence of pedestrians significantly disrupts a cyclist's fixation on the path before them as well as when there are changes to the cycling path \cite{MANTUANO2017408}. The paper proposes the need for greater visual separation between pedestrians and cyclists when they share existing infrastructure. One real-world study of cyclists found that subjective risk was related to the spatial complexity of the environment, and this complexity contributed to more hectic gaze behaviour, shorter fixation durations, and more fixations away from the direction of travel \cite{von2020gaze}. Most importantly, the study found that how far ahead cyclists looked and how far their gaze deviated from the direction of travel are more informative measures of perceived danger than fixation durations. Finally, one paper has found that right turns for cyclists (in a right-hand traffic context) can induce greater mental workload than left turns \cite{rawson2025eyes}.
\def\rqone{How do eye gaze metrics compare when the cyclist is utilising a bike, bus, or car lane?}
\def\rqtwo{How does eye gaze compare for roundabouts, signalised crossings, unsignalised turns, and zebra crossings?}
\def\rqthree{How do eye gaze metrics compare across events (as defined in Table~\ref{tab:annotation_labels})?}
\def\rqfour{How do cyclists respond to passes when a bike lane is present versus when no bike lane is present?}
\def\rqfive{How does cyclist eye gaze compare across routes?}
\section{Methodology}
\subsection{Research Questions}
The research is motivated by the following questions:
\begin{enumerate}
\item[\textbf{RQ1.}] \rqone
\item[\textbf{RQ2.}] \rqtwo
\item[\textbf{RQ3.}] \rqthree
%\item[\textbf{RQ4.}] \rqfour
\item[\textbf{RQ4.}] \rqfive
% \item[\textbf{RQ6.}] \[Do something with the segmentation.... what objects do cyclists fixate on most.\]
\end{enumerate}
\subsection{Data Collection}
Data was collected using the first-generation Meta Aria research glasses \cite{aria} which include two eye tracking cameras, two external-facing mono scene cameras, and one POV RGB camera. A cyclist wore the glasses while cycling around 3 different routes in Oxford over the course of several days in 2024 and 2025. An overview of the routes is presented in Figure~\ref{fig:route_plan}. %and in-depth details of the routes are outlined in the Oxford RobotCycle project paper \cite{robotcycle}.
Routes were cycled in both directions and routes demonstrate low-speed environments of 20 and 30 mph roads through city centre and urban neighbourhood areas. Data was collected from 12 runs (5 city centre, 4 north, 3 south). Each route generally took 20-30 minutes to complete, allowing ample time for diverse situations to be encountered including emergency vehicles, close vehicle passes, and near misses with jaywalking pedestrians. The cyclist also wore the RobotCycle backpack \cite{robotcycle} which contains a number of sensors for monitoring the environment. While data from this backpack is not included in this paper, it is worth noting as this backpack may influence cyclist agility, speed, and manoeuvring.
%IMU data collected?
%Heart Rate data collected? Heart rate data was also collected from the cyclist on their [GARMIN WHAT MODEL] smartwatch.
Data was collected and handled in accordance with the University of Oxford central privacy and ethics approval.
\begin{figure}[h]
\centering
\begin{subfigure}{0.20\textwidth}
\includegraphics[trim=490 75 450 150, clip,width=\textwidth]{NL4.jpg}
\caption{North Loop: 6.05 km}
\end{subfigure}
\begin{subfigure}{0.20\textwidth}
\includegraphics[trim=490 75 450 150, clip,width=\textwidth]{CL4.jpg}
\caption{Central Loop: 5.38 km}
\end{subfigure}
\begin{subfigure}{0.20\textwidth}
\includegraphics[trim=490 75 450 150, clip,width=\textwidth]{SL4.jpg}
\caption{South Loop: 5.99 km}
\end{subfigure}
\caption{Data Collection Routes in Oxford.}
\label{fig:route_plan}
\end{figure}
\subsection{Analysis}
Eye gaze projection points onto the scene were calculated using the Meta Aria Research Toolkit \cite{aria}.
To estimate eye movement patterns, a variation of the popular I-VT algorithm \cite{salvucci2000identifying} was utilised using the pymovements python package version 0.26.2 \cite{pymovements}.
Segments where the bike was stationary at traffic lights, before they started the route, or after they reached the finish were excluded as the cyclist is not focused on the riding during these moments.
To analyse fixation durations and horizontal gaze dispersions across the environment variables in our research questions, we first log-transformed our data to address a strong-right skew present in the distributions. We then performed Welch's ANOVA on the log-transformed data as an omnibus test followed by Games-Howell post-hoc comparisons for more detailed analysis between variables. In our results, the \textit{p} value represents the value from the Games-Howell post-hoc comparisons, the \textit{g} value represents Hedges' g, and the \textit{ratio} value represents the geometric mean ratio of the variables.
%Code used for analysis can be found at the following GitHub link:
% [Not actually doing this for this paper, will only add if time permits] Object detection and segmentation was performed on the videos using yolo26x-seg \cite{} with the default confidence threshold of 0.25 with all classes enabled. In post analysis, classes were filtered to include only....
\subsection{Data Annotation}
A summary of annotations labels can be found in Table~\ref{tab:annotation_labels}. Labels that should be noted include \textit{ego\_passing} which involves the cyclist overtaking a vehicle that was parked in a position that is not a normal parking spot and required the cyclist to adjust their trajectory for it. A \textit{pass} occurs whenever a vehicle passes the cyclist in the same direction as the cyclist is travelling and a \textit{close\_pass} was labelled when the vehicle passed closer than would be expected given the available road space. There were few close passes in the final data. The \textit{oncoming\_pass} label represents a vehicle travelling in the opposite direction of the cyclist and was only an annotated event when the road was narrow or the lane was shared for both directions. Thus, an oncoming pass from a vehicle in an opposite lane separated by a lane line was not annotated. Passes and oncoming pass events only apply to vehicles, motorcycles, and scooters and do not include pedestrians or other cyclists. Passes were annotated starting at the moment they appear in the video, however, the cyclist is likely to have heard the vehicle passing from behind. Thus, we perform analysis on a few seconds preceding the pass. A \textit{vehicle\_pull\_out\_front} annotation was applied whenever a vehicle crossed the cyclist's path in the near distance in a way that caused the cyclist to reduce speed or change trajectory. Data was annotated using ELAN 7.1 \cite{elan}. Annotation was performed by one researcher to ensure maximum consistency in annotation style across samples.
\begin{table}[]
\centering
\caption{Route Annotation Labels}
\begin{tabular}{l l}
\hline
\textit{infrastructure\_type} & bike\_lane\\
& car\_lane \\
& shared\_bus\_lane \\
& bike\_path \\
& shared\_path \\
& off\_road \\
\hline
\textit{event} & ego\_passing \\
& pass \\
& close\_pass \\
& oncoming\_pass \\
& pedestrian\_in\_lane \\
& vehicle\_pull\_out\_front \\
\hline
\textit{intersection\_context} & roundabout \\
& passing\_signalised\_crossing \\
& left\_turn \\
& right\_turn \\
& zebra\_crossing \\
& crossing\_road \\
& lane\_change \\
\hline
\textit{other\_context} & bike\_stopped \\
& construction\_zone \\
& object\_in\_road \\
& emergency\_vehicle
\end{tabular}
\label{tab:annotation_labels}
\end{table}
For \textit{infrastructure\_type}, the data is labelled based on where the bike is in the scene, not what infrastructure is available. For instance, a given segment may have a bike lane available but the cyclist was in a turn lane instead. For this segment, the data would be labelled as \textit{car\_lane}. A \textit{bike\_path} represents a lane only available to bikes and with a median between other lanes such as pedestrian paths or car lanes. A \textit{shared\_path} represents a path that is used by both pedestrians and cyclists.
\section{Results}
%\subsection{Answering Research Questions}
\subsubsection{RQ1. \rqone}
Throughout Oxford and particularly on the chosen routes, bike lanes are mostly very narrow (the width of a single bike), unprotected, and separated from car lanes by broken lines rather than solid lines. Because of this, we expect that the presence of a bike lane does not have a meaningful impact on the perceived safety compared to a car lane. Bus lanes, however, are quite broad and provide a wide space between the cyclist and other vehicles, meaning we should expect a measurable difference in eye gaze metrics.
The strongest effect of fixation duration occurs between car and bus lanes (p<0.001, g=-0.21, ratio=0.80), where the fixation durations are about 20\% shorter in the car lane. Another significant effect occurs between bike and car lanes (p=0.001, g=0.07, ratio=1.08), where fixations are shorter during bike lane riding. However, the difference is only 8\% and the sample size is large, which may mean the significant effect is not as strong as it appears.
When comparing bike lanes to bus lanes, there is a significant effect (p=0.023, g=-0.14, ratio=0.86), representing about 14\% shorter fixation durations in bike lanes than bus lanes. Density of fixation durations by infrastructure type can be found in Figure~\ref{fig:fixation_duration_infrastructure_type}.
Concerning the strongest effect of infrastructure on dispersion, dispersion was approximately 8.5\% larger during bike lane riding compared to car lane riding (p<0.001, g=0.09, ratio=1.085), a similar mean difference to what was found with fixations for these two lane types. No other significant effects were found between infrastructure types on dispersion. Density of dispersions by infrastructure type can be found in Figure~\ref{fig:dispersion_infrastructure_type}.
These findings are not surprising given the narrow nature of most Oxford bike lanes, leaving the cyclist almost as exposed to cars as if they were in the car lane. Thus, the similarities between bike and car lanes that we see in fixation and dispersion comparison align with expectations. With the large width and general solid painted stripe of the bus lane, along with bus drivers in Oxford who are very used to navigating around cyclists, this leaves more time for the cyclist to fixate on the environment and requires less scanning behaviour as is seen in the greater fixation durations of the bus lane.
We also calculated fixation frequency, measured in fixation events per minute, and found that bike lanes had the highest frequency at 142.2 fixations per minute, followed by car lanes at 80.5 fixations per minute, and finally bus lanes at 62.8 fixations per minute. These results combined with the dispersion results indicate that when in bike lanes, the cyclist is able to more rapidly move their fixation on points in the environment, possibly representing increased cognitive load as they rapidly switch between fixation points. Since the bus lane demonstrated the longest fixation duration times, it is not surprising that it also demonstrates the lowest fixation frequency.
\begin{figure}[h]
\centering
\caption{Fixation Duration by Infrastructure Type}
\includegraphics[width=0.8\linewidth]{output.png}
\label{fig:fixation_duration_infrastructure_type}
\end{figure}
\begin{figure}[h]
\centering
\caption{Dispersion by Infrastructure Type}
\includegraphics[width=0.8\linewidth]{output2.png}
\label{fig:dispersion_infrastructure_type}
\end{figure}
\subsubsection{RQ2. \rqtwo}
Although labelled in the dataset, we exclude analysis during scenarios where the cyclists crossed another road via a yield because this intersection type occurred very infrequently, only once per run in the south loop.
Passing a signalised crossing shows significance compared to roundabouts (p=0.004, ratio=1.27, g=0.22), representing that fixation durations are about 27\% longer during signalised crossings compared to roundabouts. At near significance (p=0.065), fixations at signalised crossings are 22\% longer than during right turns as well (ratio=1.22, g=0.19). Although not significant, signalised crossings produce longer fixations overall than every other intersection type. The density distribution of fixation durations by intersection type is presented in Figure~\ref{fig:fixation_duration_intersection}.
This may indicate that when passing signalised crossings, fixations occur on specific elements (such as the traffic signal itself or pedestrians waiting to cross), while other intersections require rapid scanning of the environment for vehicles entering and exiting as well as the desired exit point, leading to shorter individual fixations as gaze shifts more frequently. This is likely particular to Oxford where many of the signalised crossings are only for pedestrians and do not contain joining or crossing roads at the signal. In those scenarios, there is less of a burden for the cyclist to scan for traffic that may join from other directions.
Concerning gaze dispersion, we find a significant effect for left turns compared to roundabouts (p=0.031, ratio=1.29, g=0.29), where gaze dispersion during left turns is 29\% greater than during roundabouts. Signalised crossings compared to roundabouts represent a near significant effect (p=0.057, ratio=1.19, g=0.19), where gaze dispersion during signalised crossings is 19\% greater than during roundabouts. No other pairwise comparisons reached or were near significance. This may indicate that roundabouts allow for more fixed gaze as there is more constrained flow of traffic compared to left turns or places where pedestrians are likely to cross. Overall, roundabouts elicit a noticeably different gaze pattern compared to other intersection types, as roundabouts represent the shortest fixations and the most focused gaze. The density distribution by intersection type is presented in Figure~\ref{fig:dispersion_intersection}.
\begin{figure}[h]
\centering
\caption{Fixation Duration by Intersection Type}
\includegraphics[width=0.8\linewidth]{output3.png}
\label{fig:fixation_duration_intersection}
\end{figure}
\begin{figure}[h]
\centering
\caption{Dispersion by Intersection Type}
\includegraphics[width=0.8\linewidth]{output4.png}
\label{fig:dispersion_intersection}
\end{figure}
\subsubsection{RQ3. \rqthree}
In general, there was no statistically significant effect in fixation duration between event types (Welch's ANOVA p=0.586). The Games-Howell post-hoc comparison confirmed this finding, as no pairwise comparison achieved significance (p<0.05). Geometric mean fixation durations across the event types are very consistent: roughly 354 ms (\textit{pedestrian\_in\_lane}), 346 ms (\textit{vehicle\_pull\_out\_front}), 343 ms (\textit{oncoming\_pass}), 357 ms (\textit{ego\_passing}), 379 ms (\textit{pass}). Across thousands of fixations and five event types, the spread is only about 35 ms.
We performed the analysis for gaze dispersion metrics across events and found a similar null result (Welch's ANOVA p=0.207). The Games-Howell post-hoc comparison confirmed this finding, as no pairwise comparison achieved significance (p<0.05). Nevertheless, there is a 23\% difference in dispersion across events. Events involving lateral spatial relationships (\textit{close\_pass, pass, pedestrian\_in\_lane, ego\_passing}) tend toward more dispersed gaze, while events ahead in the road (\textit{oncoming\_pass, vehicle\_pull\_out\_front}) tend toward more focused gaze. Lateral events require scanning across the cyclist's flanks (checking adjacent vehicles, monitoring overtaking, tracking pedestrians on cross-paths), which intuitively would result in broader gaze dispersions. Events which require attention to a single point ahead, such as the oncoming vehicle or the vehicle pulling out, should produce more focused gaze. Nevertheless, this pattern does not reach significance. Densities of fixation durations by event and dispersions by event are presented in Figure~\ref{fig:fixation_event} and Figure~\ref{fig:dispersion_event}, respectively.
When we compared fixations during event periods (n = 2979) to non-event periods, we found that a Welch's ANOVA on log-transformed durations revealed a statistically significant difference (p<0.001); however, the effect was extremely small in magnitude (g=0.08).
Geometric mean fixation duration was approximately 393 ms during non-event periods compared to 362 ms during events, a difference of roughly 31 ms, or 8.7\% (95\% CI: 4.2\% to 13.4\%). This finding indicates that gaze patterns do indeed change in the presence of events, but high variation in eye gaze patterns means we cannot distinguish between types of events from gaze metrics alone.
%\subsubsection{RQ4. \rqfour}
\begin{figure}[h]
\centering
\caption{Fixation Duration by Event}
\includegraphics[width=0.8\linewidth]{output5.png}
\label{fig:fixation_event}
\end{figure}
\begin{figure}[h]
\centering
\caption{Dispersion by Event}
\includegraphics[width=0.8\linewidth]{output6.png}
\label{fig:dispersion_event}
\end{figure}
\subsubsection{RQ4. \rqfive}
% We can compare within each direction of the route too.
The south route had the highest average fixation duration at 754 ms, followed by north at 710 ms, and the centre route at 692 ms. These results intuitively make sense as the southern route contains much longer straight stretches of road. Figure~\ref{fig:fixation_by_route} and Figure~\ref{fig:dispersion_by_route} detail the density distribution of fixations and dispersions, respectively, by routes.
There is a statistically significant effect between fixation durations for the centre and north routes (p=0.019, g=-0.05, ratio=0.95) and between the centre and south routes (p=0.005, g=-0.06, ratio=0.93). However, there is no significant effect between the north and south routes (p=0.767, g=-0.015, ratio=0.98).
Concerning gaze dispersion, there is a significant effect between the centre and north route (p<0.001, g=0.11, ratio=1.11) and between the north and south route (p<0.001, g=-0.14, ratio=0.88). There is no significant effect between the centre and south route (p=0.357, g=-0.03, ratio=0.97). Overall, the north route elicits more focused gaze than both the centre and south routes.
% TODO: Tie these back in with earlier results on infrastructure types relative to fixation and dispersions.
% Really, I need to train a mixed effects model
\begin{figure}[h]
\centering
\caption{Fixation Duration by Route}
\includegraphics[width=0.8\linewidth]{output7.png}
\label{fig:fixation_by_route}
\end{figure}
\begin{figure}[h]
\centering
\caption{Gaze Dispersion by Route}
\includegraphics[width=0.8\linewidth]{output8.png}
\label{fig:dispersion_by_route}
\end{figure}
\section{Discussion and Conclusion}
This study measured the eye gaze of a cyclist across diverse routes in Oxford, UK under different traffic and weather conditions across 2024 and 2025. The routes represent a broad mix of bike lanes, shared bike lanes, bike paths, bus lanes, bus traffic, car traffic, and pedestrian presence.
%Consider 5 seconds or so before these events.
Our results indicate that raw fixation durations alone proved not to be a useful discriminator across event types and thus fixation duration alone is not enough to predict an event type. These results highlight the importance of understanding which objects the cyclist fixates on at which points in time. This will be an important area of future research to improve the usefulness of the dataset. Nevertheless, fixation durations and gaze dispersions provide useful insights about cyclist responses to infrastructure, notably that roundabouts generally allow for more focused attention due to the constrained flow of traffic.
There are a few limitations to the present study. Future work may consider the inclusion of inertial measurement unit (IMU) data in the analysis to improve the understanding of fixations in relation to head movements. Furthermore, future work should use the gaze projection points on the environment to understand not just what is in the scene but what the cyclist is directly looking at during fixations. Understanding not only when the cyclist fixates, but what the cyclist fixates on could further enhance the insights about what triggers a fixation in diverse conditions and environments.
Our study reveals the importance of the metrics of fixation duration and horizontal gaze dispersion for understanding how cyclists perceive and interact with various traffic environments. However, our results highlight the limited ability of these two metrics alone to predict stress or workload of a rider. % this is because there's no clear idea of complex/difficult environment = shorter fixations or something like that.
\begin{acks}
The authors would like to acknowledge the immense work of the RobotCycle and Oxford Robotics Institute team, particularly Pratik Somaiya and Daniel Marques, for organising, collecting, and anonymising the data that was used for this project.
\end{acks}
%%
%% The next two lines define the bibliography style to be used, and
%% the bibliography file.
\bibliographystyle{ACM-Reference-Format}
\bibliography{references}
% %%
% %% If your work has an appendix, this is the place to put it.
% \appendix
\end{document}
\endinput
%%
%% End of file `sample-manuscript.tex'.