Smartphone-Based Human-Machine-Interface for Bicycles: A Study on Behavioral Change and Learning Effects

Abstract

In future urban mobility, safe and efficient interaction between vulnerable road users and autonomous vehicles (AVs) will play a crucial role. In order to enable communication between human road users and AVs, different human-machine interfaces (HMI) are developed. Usually, these HMIs and onboard communication units are part of AVs, but some concepts exist that give cyclists communication capabilities and possibilities to interact with the human rider. This paper further investigates one of these on-bicycle HMIs that uses a smartphone mounted on the bicycle's handlebar. On the device, an application is running that augments routing apps with information about upcoming traffic scenarios and gives instructions on how to behave in certain situations. When interacting with AVs, knowing whether an HMI system influences the cyclist's behavior is crucial. Therefore, an AV can anticipate the cyclist's movement in the upcoming scenario reliably. In this paper, we focus on the research questions of whether there is a behavioral change, how it looks like, and whether learning effects with the application can be observed. We studied the behavior in a coupled Bicycle-AV-Simulator and focused on speed variations in the analysis, because of driving simulator validity. The results indicate a speed decrease after receiving app information about the upcoming scenario. However, a learning effect can be found. With an increasing number of study scenarios, the speed reduction decreases. Moreover, after receiving instructions on priority decisions, the cyclist reduces the speed if the AV takes priority and maintains or increases speed if the cyclist is prioritized.

Full text

SMARTPHONE-BASED HUMAN-MACHINE-INTERFACE FOR BICYCLES: A STUDY1
ON BEHAVIORAL CHANGE AND LEARNING EFFECTS2
3
4
5
Johannes Lindner, Corresponding Author6
Chair of Traffic Engineering and Control Technical University of Munich (TUM)7
Arcisstrasse 21, 80333 Munich, Germany8
johannes.lindner@tum.de9
ORCiD: 0000-0001-9385-245310
11
Georgios Grigoropoulos12
Chair of Traffic Engineering and Control Technical University of Munich (TUM)13
Arcisstrasse 21, 80333 Munich, Germany14
george.grigoropoulos@tum.de15
ORCiD: 0000-0002-0846-644116
17
Andreas Keler18
Chair of Traffic Engineering and Control Technical University of Munich (TUM)19
Arcisstrasse 21, 80333 Munich, Germany20
andreas.keler@tum.de21
ORCiD: 0000-0002-2326-161222
23
Klaus Bogenberger24
Chair of Traffic Engineering and Control Technical University of Munich (TUM)25
Arcisstrasse 21, 80333 Munich, Germany26
klaus.bogenberger@tum.de27
ORCiD: 0000-0003-3868-957128
29
30
Word Count: 7492 words +0 table(s) ×250 =7492 words31
32
33
34
35
36
37
Submission Date: July 31, 202338

Lindner, Grigoropoulos, Keler, and Bogenberger 2
ABSTRACT1
In future urban mobility, safe and efficient interaction between vulnerable road users and au-2
tonomous vehicles (AVs) will play a crucial role. In order to enable communication between3
human road users and AVs, different human-machine interfaces (HMI) are developed. Usually,4
these HMIs and onboard communication units are part of AVs, but some concepts exist that give5
cyclists communication capabilities and possibilities to interact with the human rider. This paper6
further investigates one of these on-bicycle HMIs that uses a smartphone mounted on the bicycle’s7
handlebar. On the device, an application is running that augments routing apps with information8
about upcoming traffic scenarios and gives instructions on how to behave in certain situations.9
When interacting with AVs, knowing whether an HMI system influences the cyclist’s behavior is10
crucial. Therefore, an AV can anticipate the cyclist’s movement in the upcoming scenario reliably.11
In this paper, we focus on the research questions of whether there is a behavioral change, how it12
looks like, and whether learning effects with the application can be observed. We studied the be-13
havior in a coupled Bicycle-AV-Simulator and focused on speed variations in the analysis, because14
of driving simulator validity. The results indicate a speed decrease after receiving app information15
about the upcoming scenario. However, a learning effect can be found. With an increasing num-16
ber of study scenarios, the speed reduction decreases. Moreover, after receiving instructions on17
priority decisions, the cyclist reduces the speed if the AV takes priority and maintains or increases18
speed if the cyclist is prioritized.19
20
Keywords: On-Bicycle Warning System; AV-VRU-Interaction; Human-Machine-Interface; Cou-21
pled Driving Simulator; Cycling Behavior; Vulnerable Road Users22

Lindner, Grigoropoulos, Keler, and Bogenberger 3
INTRODUCTION1
In urban mobility planning, efforts are made to increase traffic safety and more sustainable. Be-2
sides others, one can recognize these two trends: autonomous driving and the promotion of cycling3
in urban areas. In today’s urban traffic, the interaction between motor vehicles and cyclists is as-4
sociated with statistics about high fatality rates of cyclists in interaction with motorized traffic5
(1, 2). In the future, this interaction could be improved by excluding human errors through au-6
tomation and communication technology. Automated vehicles are equipped with onboard units7
to enable communication with other road users and infrastructure, but how is the situation with8
cyclists? Traditional cycling is considered inexpensive, space-effective, and simple, and bicycle9
designs are not expected to change significantly in the future. However, few safety add-ons will10
likely be included in bicycle designs when market penetration of autonomous vehicles and V2X-11
communication (Vehicle-to-Everything communication, including infrastructure and other vehi-12
cles) coverage is high. Simple onboard units for better perception of cyclists and human-machine13
interfaces (HMI) to communicate to the human bicycle rider other vehicles’ intentions or warning14
of safety-critical situations are discussed in the literature, besides other concepts (3–5). Even today,15
cyclists often use smartphone-based map services for routing while fixing the smartphone to the16
handlebar. This motivated Lindner et al. (3) to study whether extending those routing applications17
with additional safety-related information and behavior instructions in specific scenarios is possi-18
ble. So far, whether an on-bicycle HMI influences cycling behavior has yet to be analyzed. For19
interaction with AVs and training the involved prediction algorithms, a behavioral change is highly20
relevant. This paper further analyzes the coupled driving simulator study in Lindner et al. (3) of21
an automated vehicle (AV) and a bicycle while the cyclist uses a smartphone as a communication22
device on the operational level of the bicycle ride. The research questions addressed in this paper23
are: (1) does cycling behavior differ with and without an on-bicycle HMI system in use, (2) if yes,24
how does the behavioral change look like, and (3) can learning effects be observed when using the25
HMI system? We first provide an overview of the state of the art of on-bicycle HMIs and bicycle26
simulators for road user interaction studies, followed by a description of the experimental setup27
and the simulator study. We then present the results of the bicycle behavior analysis, including28
the influence of HMI messages on cycling behavior and learning effects. Finally, we discuss the29
results regarding driving simulator validity, summarize our findings, and provide an outlook for30
future work.31
STATE OF THE ART32
On-Bicycle Human-Machine-Interfaces33
Traditionally human-machine interface (HMI) research comes from the automotive domain to en-34
sure and emend appropriate interaction between conventional vehicles and their drivers (6). With35
the development of advanced driver assistant systems (ADAS) toward a higher level of automa-36
tion, new possibilities to communicate with the driver and other human road users have emerged.37
In recent years, a particular type of HMI was discussed intensively in research, the external HMI38
(eHMI) (6–13). Examples of HMIs are text messages displayed on the outside of the vehicle, light39
strips mounted on the windshield changing color depending on the vehicle’s state, or concepts that40
imitate human behavior, such as eyes installed at the vehicle front looking at other road users. For41
external HMIs, many studies exist on which type of eHMI might be the best for communicating the42
AV’s driving intention. So far, there is no clear recommendation for HMI design (3, 7). Also, there43
is criticism about the concepts because many are only investigated in 1-to-1 interaction, which is44

Lindner, Grigoropoulos, Keler, and Bogenberger 4
not necessarily the case in reality. It could especially lead to issues addressing information to one1
specific road user, ensuring that no other person misinterprets the eHMI’s message. Moreover,2
there is criticism in the literature that the interaction for some road users, mainly AV to pedestrian,3
is much more investigated than for others. For example, AV-to-bicycle interaction is rarely investi-4
gated, although this mode of transport plays a vital role in urban traffic (3, 7). Moreover, Dietrich5
et al. (14) states that in today’s traffic, such encounters are mostly resolved implicitly in urban6
traffic, i.e., through the kinematic motion of an approaching vehicle (14). For example, the driver7
decelerates to communicate the driver’s intention to let a pedestrian cross. For safe interaction be-8
tween vulnerable road users (VRUs) and AVs, both communication methods, implicit and explicit,9
should be integrated into the communication framework of an automated vehicle. For bicycles,10
there is little research in HMI development or investigation of AV-to-bicycle interaction (15, 16).11
Nevertheless, some visions and experimental designs of on-bicycle HMIs exist (3, 5, 17, 18). These12
concepts range from haptic interfaces (vibration motors at the handlebar) over auditory interfaces,13
like helmet audio, to visual interfaces using laser projections, a head-up display (HUD), or smart-14
phones for cyclists. In contrast to HMI research for AVs, so far, these concepts should only be15
considered as safety add-ons. All these concepts are currently in the conceptual phase. The studies16
found focus on user acceptance, comparison of reaction times, and identification of the best com-17
munication modality of such a novel HMI system for bicycles (3, 5, 17, 18). The impact on cycling18
behavior has not been researched. Especially when interacting with AVs, it is essential to study the19
driving behavior of cyclists using such an HMI system. The AV’s prediction algorithms need to be20
trained to determine whether there is a change in driving behavior and how it influences cyclists’21
behavior to correctly interpret cyclists’ movements. In this paper, we address this research gap to22
identify the influence of one specific on-bicycle HMI on the driving behavior of a cyclist.23
Bicycle Simulators for Road User Interaction Studies24
In order to study novel interaction concepts, driving simulators can be used to research in a safe and25
reproducible environment. The most experience in the driving simulation domain has been gained26
with car driving simulators. Car driving simulators can differ greatly in their hardware setups, from27
simple keyboard-screen setups to highly elaborate systems using moving platforms, LED-Screens,28
and real vehicle mock-ups (15, 19–21). Also, bicycle simulators are increasingly used to study road29
user interaction and new infrastructure designs (15, 16, 22–25). As visualization methods, usually30
screens or virtual reality headsets are used. There are major differences in the sensory equipment31
and the force feedback loops of the simulators. Very elaborate setups exist, including steering force32
feedback, a motion platform (roll and pitch angels), brake force sensors, and a headwind simulator,33
only to name a few properties (25). On the other hand, simpler hardware setups measure only the34
main driving parameters, steering angle (without force feedback), and speed using a bicycle in a35
fixed frame (15). Despite the differences in the hardware setup of bicycle simulators, it still needs36
to be conclusively clarified which components lead to higher absolute validity of the study results.37
The requirements for bicycle simulators are also very different compared to car driving simulators.38
The physics of a bicycle ride is complex to reproduce in a simulator, for example, due to miss-39
ing centrifugal forces in fixed-base simulators. So far, no moving-base bicycle simulator exists.40
The reasons for this are manifold, but two major ones are financial resources and the possibility41
of falling off the bike in the simulator and harming the user if vehicle dynamics a not modeled42
correctly. Also, the communication patterns of cyclists must be detectable in a bicycle simulator43
for proper examination of road-user interaction (26). The communication patterns include explicit44

Lindner, Grigoropoulos, Keler, and Bogenberger 5
cues, like hand gestures, and implicit cues, like head movement, body posture, and pedaling pace1
(15, 27). For simulator studies that research road user interaction, the usage of Coupled Driving2
Simulators can be beneficial (15, 28–32). In this simulation approach, multiple humans can inter-3
act in the same virtual environment using separate simulators. In the study discussed in this paper,4
a cyclist can interact with an AV passenger (15). Coupled driving simulator studies come with the5
drawbacks of higher implementation effort for study and software but promise higher validity con-6
cerning the interaction of road users. Driving simulator studies must always be critically examined7
regarding validity because elaborate setups or study methods do not necessarily increase validity8
(33–35). For behavioral validity in driving simulator studies, two terms must be distinguished:9
absolute and relative validity (35). Absolute validity compares the exact driving and interaction10
parameters (e.g., absolute speed) of a simulator experiment to the ones of a real-world experiment.11
Relative validity compares the tendency towards a driving action (e.g., braking or accelerating) to12
the real world. Most driving simulators, also the coupled Bicycle-AV simulator used for the study13
discussed in this paper, fulfill the criteria for relative validity but insufficiently for absolute validity14
(36, 37).15
METHODOLOGY16
Experimental Setup17
FIGURE 1:Bicycle Simulator with Smartphone mounted on the Bicycle’s Handlebar. Ap-
plication showing the Pre-Info Screen in a Occlusion Scenario and both Variants of Screens
after Priority Decision
The study analyzed in this paper was conducted in a coupled driving simulator consisting18
of a separate bicycle and automated vehicle (AV) simulator. The coupling of the simulators in a19
multi-user simulator enables investigation of the interaction between those two road users. A more20
detailed description of the simulator setup and the software solution is described in Lindner et al.21
(15). The bicycle simulator consists of a training stand to fix the bicycle’s position. The speed22
can be detected using an infrared sensor measuring the number of rotations of a metal cylinder23
driven by the bicycle’s rear wheel. The steering angle is extracted by a magnetic rotary encoder24
measuring the rotation of a metal plate connected to the front wheel. Besides the speed and steering25
angle parameter, it could be measured whether cyclists give hand signals. For measuring the hand26
signals, a depth camera was used to obtain the skeleton points of the study participant. A pre-27
trained convolutional neural network can then detect whether or not a hand signal was given. More28

Lindner, Grigoropoulos, Keler, and Bogenberger 6
information about the used method to detect hand signals can be found in Malcolm et al. (38). The1
visualization and sound system for the bicycle simulator was covered by one 56 inch monitor with2
integrated speakers. The natural field of view of the monitor is 50°, but was increased by the virtual3
camera to 90°. What is unique about this study is a smartphone mounted on the handlebar that4
could show different screens connected to audio signals (information signals and warning signals),5
which are linked to the simulation (see, Figure 1). The cyclist receives routing information and6
notifications about upcoming scenarios and priority decisions via this mobile application.7
The AV simulator is a low-fidelity driving simulator with a desktop screen setup, consisting8
of three 24 inch monitors and a separate speaker. Unique in this study was a tablet as HMI device9
that can, just like for the cyclist, display different screens and give audio signals. In specific scenar-10
ios, the AV passenger could control the behavior of the AV via priority decisions in the upcoming11
scenario, meaning prioritizing the cyclist or their vehicle. In future traffic, including AVs, a change12
of prioritization rules may have several advantages. It can improve traffic performance, enhance13
AVs’ acceptability, or reduce interaction scenario complexity. Especially the last point is highly14
relevant for the widespread adoption of AVs. In complex interaction scenarios the AV cannot re-15
solve, vehicle control must be handed over to a human driver or a remote control center operator16
(3). Reducing the probability of the emergence of this complex scenario type and providing tools17
to resolve these situations can facilitate AV operation greatly.18
The software solution to conduct the coupled simulator study has two main components:19
the simulation engine and the mobile application. As a simulation engine, the game engine20
Unity3D was used. A benefit of using game engines is that a solution for multiplayer games21
usually exists, significantly reducing the implementation effort. Besides the networking and ren-22
dering tasks, the vehicle controls are the most important task of the simulation engine. In order to23
control the bicycle, the inputs must be processed and applied to the bicycle model. The parameters24
speed and steering angle are fed to a physics model (using Unity’s physics engine) to control the25
bicycle. In other bicycle simulators, the Single-Track Model is used to control the bicycle in the26
simulation. For this simulation framework, we chose the approach of physics models because a27
broader range of 3D maps, including slopes and different road surfaces, can be simulated more eas-28
ily. In contrast, the Single-Track Model represents vehicle movements in two-dimensional space29
(39). The AV control in the simulation depends on the cyclist’s movement. Both road users follow30
a fixed route but should enter study scenarios at the same time. Thus, the AV follows a predefined31
path while the speed adopts the cyclist’s speed. Also integrated into the simulation framework is32
the link to the web-based mobile application. The mobile application screens can be updated via33
an API (application programming interface) from the simulation engine. It conveys information34
about the current phase in the scenario and enables the AV passenger to decide about the priority35
in an upcoming scenario. The application usually receives state information from the simulation36
engine via the API. In the case of a priority decision, the mobile application provides the current37
scenario state to the simulation engine.38
Simulator Study39
Participants followed a predefined route during the driving simulation, including 5 different study40
scenarios out of the three scenario types Static Bottleneck,Left Turn and Occlusion (Figure 2). The41
mobile application routes the cyclist through the virtual city between the scenarios with dedecated42
navigation screens. The AV was controlled in a way that it adopts it’s speed to the cyclist’s posi-43
tion on the route in order to arrive at the next scenario at the same time. The scenarios differ in44

Lindner, Grigoropoulos, Keler, and Bogenberger 7
FIGURE 2:Study Scenarios
infrastructure and turning relations of the study participants. The scenario types are described in1
the following:2
•Static Bottleneck3
There are two scenarios with a static bottleneck (truck and construction site) on a urban4
road with one line in each driving direction. In one of the scenarios, the bottleneck blocks5
the cyclist’s lane (obstacle: truck). On the other one, it blocks the AV’s lane (obstacle:6
construction site). No bicycle path exists, so the cyclist rides on the road. No other road7
users are present in the scenario.8
•Left Turn9
The study includes two Left Turn scenarios. The scenarios are embed in to a road network10
with one lane in each driving direction. Both scenarios are the same from the infrastruc-11
ture point of view and turning relations. The cyclist wants to turn left, while AV goes12
straight on the opposite lane. There was only a variation when deciding about prioritiz-13
ing the AV passenger received additional screen information on whether the passenger’s14
decision would improve traffic flow. For the cyclist, the displayed information does not15
change. The only difference for the cyclist is receiving the priority decision message16
later with additional traffic information because the AV passenger’s decision duration is17
significantly longer than without additional information, as Lindner et al. (3) found. No18
other road users are present in the scenario.19
•Occlusion20
There are two scenarios where buildings act as visual obstructions impeding the cyclist’s21
line of sight. The cyclist goes straight, either on a bicycle path or the road(one lane per22
driving direction), while the AV crosses from the right. Without communication devices,23
the AV could be seen only briefly before the conflict point to increase the criticality of24
the situation. No other road users are present in this scenario.25
Each scenario could have several variations regarding communication; see Figure3. Scenar-26
ios could be carried out as Baseline Scenario, which are referred to as Default. This means there27
is no additional communication, and the right-of-way is regulated like in normal conditions in to-28
day’s road traffic. In scenarios where communication is investigated (Communication Scenario),29
there are two further variations of the priority decision instance. This can be the AV, simulating a30
case where the traffic control center communicates the AV an optimized traffic control strategy, or31
the AV passenger, who is free in the decision. In this simulation, there were only the options to32
prioritize the own vehicle or the cyclist, because no other road users were present in the simula-33

Lindner, Grigoropoulos, Keler, and Bogenberger 8
FIGURE 3:Variation of Communication and Priority Decisions in Study Scenarios
tion. Each communication scenario starts with the Pre-Info Phase and screen information about1
the upcoming scenario for AV passenger and cyclist (and an audio signal). This app screen depicts2
the infrastructure, the involved road users and their driving intention (see Figure 1). Also there3
is the distance to the conflict point is shown on the screen and updated in real-time. This phase4
is followed by the Decision Phase. The decision can be manually executed by the passenger or5
automatically by the AV. There is one special case for the scenario decision, as already mentioned6
in the description of the Left Turn scenario. At certain scenarios the AV passengers receive addi-7
tional information on the planned decision. A feedback is displayed on whether the decision has a8
positive or negative impact on traffic flow. From the bicycles point of view this information is not9
relevant. After the Priority Decision phase, both study participants (AV passenger and cyclist) get10
notified about the priority decision and execute the instruction (Execution Phase). Not all of these11
phases are recognizable for the cyclist, and there are some differences in handling the app informa-12
tion compared to the AV. While the AV can instantly react to the decision or other instructions and13
adopt it’s driving behavior, the cyclist requires a Reaction Phase. Every time the cyclist receives14
a message, some reaction time is involved. Also, the priority decision is not present for the cyclist15
and, therefore, the Decision Phase. The cyclist only receives the Pre-Info and the priority decision16
message, which is depicted at the bottom of Figure 3. The study included 16 simulation runs and17
two participants (13 female, 19 male - Age group 18-24: 12, 25-39: 19, 40-59: 1) each. If the18
simulation run is completed, every participant goes through 18 scenarios, including 3 Baseline and19
15 Communication Scenarios. More details about the study, the participants, the procedure, and20
the app screens can be found in Lindner et al. (3).21
RESULTS22
This section analysis the bicycle trajectories of the coupled driving simulator study. The segments23
in which the cyclist received a notification via the mobile application are particularly interesting.24
The notification is visual information in the form of a new screen and an audio warning signal.25

Lindner, Grigoropoulos, Keler, and Bogenberger 9
We separately analyze the Pre-Info message and the Priority Decision message. The focus of1
the behavior analysis is the speed variation. We are aware that there might be other interesting2
parameters like lateral position to study, but with the given experimental setup, the speed variation3
(compared to baseline scenario) can be analyzed with the highest validity. The Discussion and4
Limitations section will address the topic of driving simulator validity in more detail. Note also5
that in the following analysis, the scenarios are no more separated by decision instance. The reason6
is that from the cyclist’s point of view, the notifications and screens do not differ for the decision7
instance. The cyclist only receives the Pre-Info and Priority Decision Messages in Communication8
Scenarios.9
Cyclist Behavior after Pre-Info Message10
FIGURE 4:Speed Reduction after Pre-Info Message: Comparison of Baseline with different
Experience Levels of the Cyclists with the Mobile Application
In the following, we analyze the cyclist’s speed after receiving the Pre-Info notification.11
The screen on the smartphone gives an overview of the upcoming scenario, showing a map, all12
involved road users, and their driving intention. For this analysis, all five scenarios are used. The13
scenarios can be compared because all study scenarios’ boundary conditions during the Pre-Info14
Phase are very similar. It is on a straight road (in one scenario on a bicycle path), including15
no other road users approaching a four-way intersection or a bottleneck. Also, the analysis only16
considers only relative speeds. For the level of detail in this analysis, the assumption of scenario17
comparability is thus sufficient. In Figure 4 we can see the speed variation in percent (100 percent18
is the highest speed in that interval) as a function of the distance in meters after the Point of19
Message Receiving (PMR). We could identify the reaction phase in the interval of 0 - 4m. This20
corresponds to approximately 1 second, considering the cyclist’s average speed. We identified the21

Lindner, Grigoropoulos, Keler, and Bogenberger 10
reaction phase but do not analyze the reaction time in detail in this paper. After the reaction phase,1
the action of the cyclist starts. It is a temporary speed reduction with its maximum between 8 and2
10 meters after the PMR. Moreover, the plot shows that a learning effect of using the application3
is recognizable. The data is split into the cyclist’s experience level with the application compared4
to the baseline, corresponding to the Default scenarios without app intervention. In one simulation5
run, the cyclist drives through 15 scenarios, including app notifications. From this study design,6
we derive three experience levels (0 - 4, 5 - 9, and 10- 14 scenarios experience). After receiving7
the app notification, the speed reduction increases with a lower experience level. With a higher8
experience level, the speed reduction decreases. At the highest experience level, the speed curve9
is close to but not yet equal to the baseline case, and still, a speed reduction can be recognized in10
comparison to the baseline.11
Cyclist Behavior after Priority Decision Message12
FIGURE 5:Speed profile of the Cyclist after Priority Decision Message compared to the
Baseline
This section analyses the influence of the Priority Decision message, which corresponds to13
the Execution-Phase. Figure 5 shows the speed after the instruction to give way to the AV (Priority:14
AV, red) or to get priority (Priority: Bicycle, blue) compared to a baseline scenario (grey). The15
baseline scenarios in this plot only include scenarios in which the AV has priority according to16
today’s traffic rules. For reasons of clarity in the plot, the opposite case, baseline scenarios with17
priority for cyclist, is not visualized, but the cyclist’s speed choice looks very similar and can18
be considered equal for this comparison. The speed is visualized with the mean value and the19
25 (bottom) and 75 (top) quartiles. The app message is received at the PMR (Distance = 0m).20
For this analysis, all study scenarios are considered, because we focus on the action, which is21

Lindner, Grigoropoulos, Keler, and Bogenberger 11
an increase/decrease in speed, not the absolute speed value. For different research questions, the1
scenarios should be evaluated separately. The reaction phase is less clearly visible in this plot2
compared to the previous one. Based on the previous Pre-Info plot, we assume a similar reaction3
time of approximately 1 second. One can also argue that reaction time is reduced, because after4
the Pre-Info message the cyclist also awaits the next screen. However, this is a research question5
for another analysis. After the reaction phase, we could observe the following.6
• Priority Bicycle: In these scenarios, the cyclist received the message to have priority.7
Compared to the initial average speed, the speed after receiving the message leads to8
maintaining or increasing speed of the cyclist’s velocity, starting about 4m after the mes-9
sage was received.10
• Priority AV: The cyclist has to give way to the AV at an intersection or a bottleneck. The11
analysis shows that the cyclist reduces speed to give way to the AV. The minimum speed12
is reached 6 - 10m after the PMR.13
• Baseline: The baseline corresponds to the data of Default scenarios. Since the initial14
app notification’s position varies, the Default scenarios’ trajectories start at the PMR of15
the non-default scenarios. The baseline plot shows that the speed in default scenarios16
remains relatively constant.17
When comparing the absolute speeds in the plot, one can see that the average speed in the Default18
scenario (Baseline) is higher than in both Communication Scenarios (Priority: AV and Priority:19
Bike). Also the difference between the quartiles of the speed distribution remain in a similar range20
compared to the baseline scenario.21
Comparison of Scenarios separated by Priority Decision Message22
FIGURE 6:Analysis of a Bottleneck Scenario including lateral Position Distribution, Speed
and App Notifications (Priority Bicycle, top and Priority AV, bottom)
This section analyzes the trajectories of the cyclist of all study participants and simulation23

Lindner, Grigoropoulos, Keler, and Bogenberger 12
runs. It visualizes some of the results from the sections above, additionally with positional infor-1
mation by the example of one Bottleneck Scenario. The top subplot in Figure 6 shows a bottleneck2
scenario with the obstacle on the cyclist’s lane and priority given to the cyclist. The bottom sub-3
plot visualizes the same scenario with priority for the AV. Again, the analysis will be performed4
based on actions because only relative validity could be assumed for the simulator results (36, 37).5
This means the maneuver intention can be interpreted as valid (e.g., braking intention, lane change6
intention) but not the absolute trajectory (e.g., exact position, speed). All figures include an aggre-7
gated visualization of the trajectories of all simulation runs and participants. The trajectories are8
depicted as area, while the color indicates the speed. The top and bottom border of the area is the9
25 (bottom) and 75 (top) quartiles of the y-position of a 6 m discrete element in the x-direction.10
The black middle line represents an x-discrete element’s median. The top two box plots in each11
subplot indicate the position of the Pre-Info and Execution phases. In this plot, not the PMR but12
the whole time the cyclist received the respective screen information. Thus, the app information’s13
position can be directly compared with the bicyclist’s speed and position. The variation of the14
notification position has several explanations. In scenarios with the AV passenger as the priority15
decision instance, the timing depends on the decision duration of the passenger. Other influencing16
factors are the speed of the cyclist and latency effects. Moreover, it was possible in the simulator17
to give and detect hand signals. A general observation of that simulator study is that participants18
rarely use hand gestures, too little to include them in this plot meaningfully. In the top subplot, the19
cyclist approaches the scenario from the right side of the figure. After the Pre-Info, the app informs20
the cyclist to have priority against today’s existing traffic rules. The plot clearly shows that after21
receiving the Pre-Info, the cyclist slows down. The cyclist increases or maintains the speed after22
receiving the second message about the priority decision. The bottom subplot shows a scenario in23
the same environment. Only the priority decision has changed, and now the AV has priority. Like24
in the first case, a deceleration could be observed after the Pre-Info screen. The cyclist gives way25
to the AV and decelerates as a consequence of the priority decision.26
DISCUSSION27
This study’s main objective is to identify the cyclist’s behavioral changes and learning effects. On-28
bicycle HMIs are developed as an add-on for cycling safety using V2X communication. As we29
could identify in the literature review, these HMI concepts should also improve the encounter with30
automated vehicles. Prediction of other road users is an essential aspect of AV research. Especially31
critical is the prediction of VRU behavior, but also complex at the same time. An additional system32
like on-bicycle HMIs could further increase the complexity of the prediction task. If the bicycle33
HMI affects behavior, the behavior should be analyzed and used as input for prediction algorithms.34
In this paper, we focus on speed variations because the relative validity of the bicycle simulator35
can be assumed. Other parameters like lateral position or eye movement to study attentiveness36
may also have an influence, but could not be studied validly given the available simulator setup.37
Considering this background, we can answer the first and second research questions of whether38
and how the behavioral change of cyclists changes. After the Pre-Info notification, we can observe39
a speed reduction after the PMR. Given the average cyclist’s speed, the deceleration process starts40
after approximately one second of reaction time, corresponding to 4m. The results indicate a41
behavioral change in speed adaption. One aspect must be highlighted here, which also answers the42
third research question. The study participants have a learning effect when using the application,43
as shown in Figure 4. The speed reduction decreases with an increasing number of Communication44

Lindner, Grigoropoulos, Keler, and Bogenberger 13
Scenarios, including app notifications. One can state the hypothesis that a new task increases the1
cyclist’s cognitive workload and directs the visual attention from the road to the mobile phone,2
resulting in a speed reduction, as found in previous simulator studies (40). This result indicates3
that this on-bicycle HMI influences speed adaption. However, not only a behavioral change but4
also a learning effect could be observed. Whether the speed reduction can be neglected after a5
certain training period has yet to be answered. Besides the speed reduction after the Pre-Info6
Message, the behavior after receiving the Priority Decision message was analyzed. This behavior7
is even more interesting for interaction with AVs. After the priority decision in favor of the AV, one8
can expect a speed reduction of the cyclist to let the AV pass. The other way around, the cyclist9
maintains speed or slightly accelerates. Figure 5 also contains the baseline scenario. Compared to10
the communication scenarios, the average speed is higher, excluding the effect of speed reduction11
after the AV gets priority. An explanation could be that due to the Pre-Info message before, the12
study participants are waiting for new messages and therefore reduce their speed to react quickly13
to the instructions. Since we used all scenarios in this analysis and since absolute speed values14
might not be represented validly in the simulator, this finding needs to be reviewed using another15
study method before making a definitive statement. For the behavior after the Priority Decision16
message, no learning effect could be observed, except for a trend that the initial speed was higher17
for scenarios with the highest experience level. When conducting further studies on the learning18
effect, the absolute speed in the scenarios separated by priority decision type should be included.19
What could also influence the study results, in general, is the fact that the study was conducted on20
a bicycle simulator. In this simulator setup, the bicycle is held in a fixed frame that does not allow21
any movement. The participant does not have the usual cycling task, including balancing out the22
own and the bicycle’s weight. Moreover, the benefit of driving simulators is that scenarios can be23
investigated safely, which negatively impacts behavioral simulator validity at the same time (41).24
Study participants will not behave like in an actual safety-critical situation because the presence in25
the simulation may be insufficient and, therefore, subjective safety might not be represented validly26
in a driving simulator. The scenario design could also impact the results because, except for one27
AV, no other road users were involved. More road users could lead to a more cautious driving28
style. For future investigations, we recommend using controlled field test experiments to include29
not only the parameter speed but also get insights into the variation of the lateral position. Also30
included should be eye tracker measurements to study attentiveness in the scenarios. The field test31
experiment also has the benefits that the vehicle dynamics can be included, and the possibility of32
falling or crashing because of inattentiveness is present. The learning effects must also be studied33
in long-term experiments or experiments with many more repetitions to evaluate whether the speed34
variation effect can be neglected after a familiarization phase.35
Limitations36
One major drawback of the study results is the limited sample size, as discussed in Lindner et al.37
(3). More trials must be carried out in order to obtain more reliable results. Also, the composition38
of the sample collective is not representative because mainly students within younger age groups39
took part. Another limitation is that only two road users were involved in the studied scenarios,40
the AV and bicycle. The concept’s applicability must be studied further for scenarios involving41
multiple interacting road users. For comparability of the scenarios and the different study partici-42
pants, the study design with a 1-to-1 interaction was chosen. Already often described in this paper43
is the topic of driving simulator validity. Studies with a driving or bicycle simulator always have44

Lindner, Grigoropoulos, Keler, and Bogenberger 14
the limitation that, to a certain extent, reality cannot be reproduced entirely. In this study, the main1
restriction from the simulator was the missing vehicle dynamics and limited field of for the cyclist.2
We thus do not expect a high value of absolute validity for distance and speed perception. When3
receiving an app notification, there are likely other effects than speed variation. Possible effects4
can be higher variance in lateral position because of a more unstable bicycle ride due to the shifted5
attention to the smartphone screen. Moreover, the absolute speed value may vary during study6
scenarios compared to a regular bicycle ride because the cyclists keep attention and wait for new7
instructions. These two effects cannot be studied validly with this simulator setup.8
CONCLUSION9
This paper investigates the variation of a cyclist’s behavior in scenarios with an on-bicycle HMI.10
We focus on relative speed variation for reasons of driving simulator validity. We could observe11
a speed reduction of the cyclist after receiving a Pre-Info message. More interestingly, study12
participants have a learning effect when using the application. With an increasing number of13
communication scenarios, the speed reduction decreases. After notifications of priority decisions,14
the cycling speed differs depending on priority. If the cyclist has the right-of-way, the average15
speed remains constant or increases. If the cyclists must give way to the AV, the speed temporarily16
decreases. These results show that the on-bicycle HMI in use influences the cyclist’s behavior,17
specifically on the speed variation, as long as the cyclist is in the early learning stage of the system.18
Future studies must investigate whether a behavioral change can still be observed after a more19
extended training period. The results also show that when designing an on-bicycle HMI of this20
or a different type, evaluating the behavioral changes induced by this bicycle safety add-on is21
necessary. Especially in the interaction with autonomous vehicles, predicting road user behavior22
is an essential part that must be addressed.23
ACKNOWLEDGMENTS24
This work is part of the research project @CITY – Automated Cars and Intelligent Traffic in the25
City. The project is supported by the German Federal Ministry for Economic Affairs and Energy26
(BMWi), based on a decision taken by the German Bundestag, grant number 19A17015B. The27
authors are solely responsible for the content of this publication.28
AUTHOR CONTRIBUTIONS29
The authors confirm contribution to the paper as follows. Study conception: JL, GG, AK, KB;30
literature review: JL; methodology: JL, GG, AK; data collection, data preparation, and implemen-31
tation: JL; analysis and interpretation of results: JL, GG, AK, KB; draft manuscript preparation:32
JL, GG, AK. All authors reviewed the results and approved the final version of the manuscript.33
References34
1. European Commission, Facts and Figures: Roads Inside Urban Areas, 2022.35
2. European Commission, Facts and Figures: Cyclists, 2023.36
3. Lindner, J., G. Grigoropoulos, A. Keler, P. Malcolm, F. Denk, P. Brunner, and K. Bogen-37
berger, A mobile application for resolving bicyclist and automated vehicle interactions at38
intersections. In 2022 IEEE Intelligent Vehicles Symposium (IV), 2022, pp. 785–791.39
4. Kapousizis, G., M. B. Ulak, K. Geurs, and P. J. Havinga, A review of state-of-the-art40

Lindner, Grigoropoulos, Keler, and Bogenberger 15
bicycle technologies affecting cycling safety: level of smartness and technology readiness.1
Transport Reviews, Vol. 43, No. 3, 2023, pp. 430–452.2
5. Hou, M., K. Mahadevan, S. Somanath, E. Sharlin, and L. Oehlberg, Autonomous Vehicle-3
Cyclist Interaction: Peril and Promise, Association for Computing Machinery, New York,4
NY, USA, p. 1–12, 2020.5
6. Bengler, K., M. Rettenmaier, N. Fritz, and A. Feierle, From HMI to HMIs: Towards an6
HMI Framework for Automated Driving. Information, Vol. 11, No. 2, 2020.7
7. Dey, D., A. Habibovic, A. Löcken, P. Wintersberger, B. Pfleging, A. Riener, M. Martens,8
and J. Terken, Taming the eHMI jungle: A classification taxonomy to guide, compare, and9
assess the design principles of automated vehicles’ external human-machine interfaces.10
Transportation Research Interdisciplinary Perspectives, Vol. 7, 2020, p. 100174.11
8. Rouchitsas, A. and H. Alm, External Human–Machine Interfaces for Autonomous12
Vehicle-to-Pedestrian Communication: A Review of Empirical Work. Frontiers in Psy-13
chology, Vol. 10, 2019.14
9. Merat, N., T. Louw, R. Madigan, M. Wilbrink, and A. Schieben, What externally presented15
information do VRUs require when interacting with fully Automated Road Transport Sys-16
tems in shared space? Accident Analysis & Prevention, Vol. 118, 2018, pp. 244–252.17
10. Chang, C.-M., K. Toda, T. Igarashi, M. Miyata, and Y. Kobayashi, A Video-Based Study18
Comparing Communication Modalities between an Autonomous Car and a Pedestrian. In19
Adjunct Proceedings of the 10th International Conference on Automotive User Interfaces20
and Interactive Vehicular Applications, Association for Computing Machinery, New York,21
NY, USA, 2018, AutomotiveUI ’18, p. 104–109.22
11. Forster, Y., F. Naujoks, and A. Neukum, Your Turn or My Turn?: Design of a Human-23
Machine Interface for Conditional Automation. Proceedings of the 8th International Con-24
ference on Automotive User Interfaces and Interactive Vehicular Applications, 2016.25
12. Bazilinskyy, P., D. Dodou, and J. de Winter, Survey on eHMI concepts: The effect of text,26
color, and perspective. Transportation Research Part F: Traffic Psychology and Behaviour,27
Vol. 67, 2019, pp. 175–194.28
13. Li, Y., H. Cheng, Z. Zeng, H. Liu, and M. Sester, Autonomous Vehicles Drive into Shared29
Spaces: eHMI Design Concept Focusing on Vulnerable Road Users. In 2021 IEEE Inter-30
national Intelligent Transportation Systems Conference (ITSC), 2021, pp. 1729–1736.31
14. Dietrich, A., P. Maruhn, L. Schwarze, and K. Bengler, Implicit Communication of Auto-32
mated Vehicles in Urban Scenarios: Effects of Pitch and Deceleration on Pedestrian Cross-33
ing Behavior. In Human Systems Engineering and Design II (T. Ahram, W. Karwowski,34
S. Pickl, and R. Taiar, eds.), Springer International Publishing, Cham, 2020, pp. 176–181.35
15. Lindner, J., A. Keler, G. Grigoropoulos, P. Malcolm, F. Denk, P. Brunner, and K. Bogen-36
berger, A coupled driving simulator to investigate the interaction between bicycles and37
automated vehicles. In 2022 IEEE 25th International Conference on Intelligent Trans-38
portation Systems (ITSC), 2022, pp. 1335–1341.39
16. Kaß, C., S. Schoch, F. Naujoks, S. Hergeth, A. Keinath, and A. Neukum, A Methodolog-40
ical Approach to Determine the Benefits of External HMI During Interactions Between41
Cyclists and Automated Vehicles: A Bicycle Simulator Study. In HCI in Mobility, Trans-42
port, and Automotive Systems. Driving Behavior, Urban and Smart Mobility (H. Krömker,43
ed.), Springer International Publishing, Cham, 2020, pp. 211–227.44

Lindner, Grigoropoulos, Keler, and Bogenberger 16
17. Oczko, M.-C. H., L. Stratmann, M. Franke, J. Heinovski, D. S. Buse, F. Klingler, and1
F. Dressler, Integrating Haptic Signals with V2X-based Safety Systems for Vulnerable2
Road Users. In 2020 International Conference on Computing, Networking and Communi-3
cations (ICNC), 2020, pp. 692–697.4
18. Berge, S. H., M. Hagenzieker, H. Farah, and J. de Winter, Do cyclists need HMIs in future5
automated traffic? An interview study. Transportation Research Part F: Traffic Psychology6
and Behaviour, Vol. 84, 2022, pp. 33–52.7
19. Kummetha, V. C., A. Kondyli, and S. D. Schrock, Analysis of the effects of adaptive cruise8
control on driver behavior and awareness using a driving simulator. Journal of Transporta-9
tion Safety & Security, Vol. 12, No. 5, 2020, pp. 587–610.10
20. Himmels, C., J. Venrooij, M. Gmünder, and A. Riener, The influence of simulator and11
driving scenario on simulator sickness. In Proceedings of the Driving Simulation Con-12
ference 2022 Europe VR (A. Kemeny, J.-R. Chardonnet, and F. Colombet, eds.), Driving13
Simulation Association, Strasbourg, France, 2022, pp. 29–36.14
21. Himmels, C., V. Andreev, A. A. Syed, J. Lindner, F. Denk, and A. Riener, Are Head-15
mounted Displays Really Not Suitable for Driving Simulation? A Comparison with a16
Screen-Based Simulator. In 2023 IEEE Intelligent Vehicles Symposium (IV), 2023, pp. 1–17
6.18
22. Keler, A., J. Kaths, F. Chucholowski, M. Chucholowski, G. Grigoropoulos, M. Spangler,19
H. Kaths, and F. Busch, A bicycle simulator for experiencing microscopic traffic flow20
simulation in urban environments. In 2018 21st International Conference on Intelligent21
Transportation Systems (ITSC), 2018, pp. 3020–3023.22
23. Bogacz, M., S. Hess, C. Calastri, C. F. Choudhury, A. Erath, M. A. B. van Egger-23
mond, F. Mushtaq, M. Nazemi, and M. Awais, Comparison of Cycling Behavior between24
Keyboard-Controlled and Instrumented Bicycle Experiments in Virtual Reality. Trans-25
portation Research Record, Vol. 2674, No. 7, 2020, pp. 244–257.26
24. Brown, H., C. Sun, and Z. Qing, Investigation of Alternative Bicycle Pavement Markings27
with the Use of a Bicycle Simulator. Transportation Research Record, Vol. 2662, No. 1,28
2017, pp. 143–151.29
25. Fischer, M., G. Temme, K. Gröne, D. M. Garcia, G. Grolms, and J. Rehm, A VRU-30
simulator for the evaluation of pedestrian-and cyclist-vehicle interaction –Design criteria31
and implementation. In Proceedings of the Driving Simulation Conference 2022 Europe32
VR (A. Kemeny, J.-R. Chardonnet, and F. Colombet, eds.), Driving Simulation Associa-33
tion, Strasbourg, France, 2022, pp. 153–160.34
26. Färber, B., Communication and Communication Problems Between Autonomous Vehicles35
and Human Drivers", bookTitle="Autonomous Driving: Technical, Legal and Social As-36
pects, Springer Berlin Heidelberg, Berlin, Heidelberg, pp. 125–144, 2016.37
27. Grigoropoulos, G., P. Malcolm, A. Keler, F. Busch, and K. Bogenberger, Predicting Bicy-38
clist Maneuvers using Explicit and Implicit Communication, 2022.39
28. Bazilinskyy, P., L. Kooijman, D. Dodou, and J. de Winter, Coupled simulator for research40
on the interaction between pedestrians and (automated) vehicles, 2020.41
29. Noyce, D., J. Kearney, S. Chakraborty, Y. Jiang, and K. Santiago-Chaparro, Multi-Modal42
Distributed Simulation Combining Cars, Bicyclists, and Pedestrians, 2018.43
30. Song, Y. E., C. Lehsing, T. Fuest, and K. Bengler, External HMIs and Their Effect on the44
Interaction Between Pedestrians and Automated Vehicles. In Intelligent Human Systems45

Lindner, Grigoropoulos, Keler, and Bogenberger 17
Integration (W. Karwowski and T. Ahram, eds.), Springer International Publishing, Cham,1
2018, pp. 13–18.2
31. Lehsing, C., A. Kracke, and K. Bengler, Urban perception - A cross-correlation approach3
to quantify the social interaction in a multiple simulator setting. In 2015 IEEE 18th Inter-4
national Conference on Intelligent Transportation Systems, 2015, pp. 1014–1021.5
32. Sawyer, B. D. and P. Hancock, Development of a Linked Simulation Network to Evaluate6
Intelligent Transportation System Vehicle to Vehicle Solutions. Proceedings of the Human7
Factors and Ergonomics Society Annual Meeting, Vol. 56, No. 1, 2012, pp. 2316–2320.8
33. Blaauw, G. J., Driving Experience and Task Demands in Simulator and Instrumented Car:9
A Validation Study. Human Factors, Vol. 24, No. 4, 1982, pp. 473–486.10
34. Milleville-Pennel, I. and C. Charron, Driving for Real or on a Fixed-Base Simulator: Is It11
so Different? An Explorative Study. Presence, Vol. 24, No. 1, 2015, pp. 74–91.12
35. Zöller, I. M., Analyse des Einflusses ausgewählter Gestaltungsparameter einer Fahrsimu-13
lation auf die Fahrerverhaltensvalidität. Ph.D. thesis, Technische Universität, Darmstadt,14
2015.15
36. Törnros, J., Driving behaviour in a real and a simulated road tunnel—a validation study.16
Accident Analysis & Prevention, Vol. 30, No. 4, 1998, pp. 497–503.17
37. Godley, S. T., T. J. Triggs, and B. N. Fildes, Driving simulator validation for speed re-18
search. Accident Analysis & Prevention, Vol. 34, No. 5, 2002, pp. 589–600.19
38. Malcolm, P., G. Grigoropoulos, A. Keler, H. Kaths, and K. Bogenberger, Analysis of Bicy-20
clist Communication in a Simulator Environment, 2021.21
39. Schramm, D., M. Hiller, and R. Bardini, Single Track Models, Springer Berlin Heidelberg,22
Berlin, Heidelberg, pp. 225–257, 2018.23
40. Engström, J., E. Johansson, and J. Östlund, Effects of visual and cognitive load in real and24
simulated motorway driving. Transportation Research Part F: Traffic Psychology and Be-25
haviour, Vol. 8, No. 2, 2005, pp. 97–120, the relationship between distraction and driving26
performance: towards a test regime for in-vehicle information systems.27
41. de Winter, J. and R. Happee, Advantages and Disadvantages of Driving Simulators: A28
Discussion, 2012.29