A virtual reality experiment to study pedestrian perception of future street scenarios

The current allocation of street space is based on expected vehicular peak-hour flows. Flexible and adaptive use of this space can respond to changing needs. To evaluate the acceptability of flexible street layouts, several urban environments were designed and implemented in virtual reality. Participants explored these designs in immersive virtual reality in a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2\times 3$$\end{document}2×3 mixed factorial experiment, in which we analysed self-reported, behavioural and physiological responses from participants. Distinct communication strategies were varied between subjects. Participants’ responses reveal a preference for familiar solutions. Unconventional street layouts are less preferred, perceived as unsafe and cause a measurably greater stress response. Furthermore, information provision focusing on comparisons lead participants to focus primarily on the drawbacks, instead of the advantages of novel scenarios. When being able to freely express thoughts and opinions, participants are focused more on the impact of space design on behaviour rather than the objective physical features themselves. Especially, this last finding suggests that it is vital to develop new street scenarios in an inclusive and democratic way: the success of innovating urban spaces depends on how well the vast diversity of citizens’ needs is considered and met.

watched the new traffic scenarios sequentially in an isolated way, while the other half watched the new street paradigms in parallel on a split screen.We hypothesized that the split-screen approach supports comparison and helps to recognize the advantages and disadvantages of each scenario better than watching videos sequentially.Ultimately, the insights we gained from this study hope to eventually support the design and implementation of future transportation policies that are more inclusive and sustainable.This was our third goal.

Results
Our experimental study involved 43 participants with a mean age of 24.76 ( ± 3.73 ) years, comprising 62.8% (n = 27) female individuals.The majority of participants were students.

Number of crossing attempts
Participants were briefed to cross the road safely (i.e.without being hit by a car).Each participant experienced two different configurations (morning rush hour and evening), as illustrated in Fig. 5, in three scenarios (statusquo, inLED, and curbless), as shown in Fig. 6 (for more details see Table 2).Table 1 displays the number of Table 1.Summary of participants and recorded sessions.Rows display the number of participants per information treatment group (sequential, parallel).Columns in the right-most part of the table show the number of sessions per scenario (status-quo, inLED, and curbless).

Status-quo inLED Curbless
Information treatments (between-subjects randomized groups) participants by scenario (columns) and treatment (rows).Each participant was given two attempts per configuration in each scenario to cross the road (d in Fig. 1).Depending on the number of failed attempts, this resulted in a somewhat different total number of sessions per scenario and per treatment.Overall, with only 4.5% there was a low rate of failed attempts, but there was a significant impact of the scenario type on the failure rate (n=266, p two−sided Fisher = 1.1e − 04 ).Most of the failures occurred in scenario inLED.Apart from this, there were no significant differences in failure rates observed across treatments (sequential or parallel), indicating that this intervention had no significant impact on performing the task (n=266, p two−sided Fisher = 1.0).

Stress, safety, willingness to cross, and attractiveness
Directly after completing the crossing for each scenario, participants responded to a brief structured survey.This survey includes a Short Stress State (SSS) 35 questionnaire, and questions regarding safety, willingness to cross, and the perception of the attractiveness of a scenario (see "6.6Questionnaires").to address repeated measures of non-parametric variables, we employed an Aligned Rank Transform (ART) ANOVA test 36 (see "Materials and methods" for more details).Our analysis was structured across three levels: design scenario (Level 1), information treatment × scenario (Level 2), and information treatment (Level 3).According to our experiments, scenarios (Level 1) elicit significant variations in responses.Participants' reactions to the scenarios show the ranking status-quo > inLED > curbless (where "a > b" means that a is more attractive than b).This suggests that participants are less willing to cross unconventional street layouts, which are perceived as less attractive and less safe.This trend is visually apparent through the negative slopes of the lines in panels b-e.In panels e,f, which analyze the differences in the perceived safety for pedestrians and cars (actually, car drivers), differences are observed.Panel e shows a steep slope with significant scenario comparisons (statusquo > inLED > curbless), while panel f exhibits nearly horizontal lines with only one significant comparison (status-quo = inLED > curbless).This suggests that the aversion to unconventional scenarios is caused mainly due to the proximity of cars rather than the proximity to pedestrians.
The analysis of the comparisons of the scenarios across information treatments (Level 2) is facilitated by counting the mint and light orange brackets shown as annotations on top of panels of Fig. 2. We observe that www.nature.com/scientificreports/ the scenario status-quo receives significantly higher ratings than scenario curbless in more questions within the sequential group (4 questions) compared to the parallel group (5 questions).Similarly when comparing scenarios inLED and curbless: the sequential group perceives scenario curbless as less likely suited to cross the street at night and less attractive than scenario inLED.In addition to these observations, the parallel group is additionally less inclined to cross the street with a child and believes the street is less safe regarding interactions with cars.
Even though the small sample size may hinder the ability to draw final conclusions for the within and between-subjects comparisons due to non-significant confidence intervals even in the presence of significant p values, the information treatment (level 3) of our study seems to suggest a likely impact with respect to three additional questions: in scenario curbless, participants belonging to the parallel group exhibited a reduced willingness to cross the street with a child and appear to perceive the street as less safe regarding the interaction with cars compared to those in the sequential group.In essence, the parallel group seems to perceive scenario curbless more negatively than the sequential group.Conversely, in scenario inLED, participants in the parallel group appear to consider the street to be safer for pedestrians than those in the sequential group.In summary, Level 3 differences seem to be most frequently observed in scenario curbless, where factors such as willingness to cross the street with a child and perceived car safety are influenced by the presentation of information regarding the scenarios.However, it is needed to further explore these trends with expanded studies (see "Statistical details" and "Limitations and future outlook").

Duration
Overall, the largest differences concerning the duration of the experiment are found in the subtasks approaching and crossing (panels b,c,e in Fig. 3).The results indicate that participants in the parallel treatment took significantly longer to approach the crosswalk compared to those in the sequential treatment ( N = 266 , p adj BY = 3.58e − 03 ), which is as well suggested by comparison with the inLED scenario.Similarly, the dura- tion of waiting (panel d in Fig. 3) was highest in the inLED scenario ( p adj BY = 7.70e − 03 ).The general trend of increased duration for the inLED scenario was also reflected in the duration of the whole experiment (panel f in Fig. 3, p adj BY = 3.45e − 03 ).The average total duration of the task increased from 43 s (± 22 s.d.) in the scenario status-quo to 51 s (± 28 s.d.) in the inLED scenario, while for curbless the average duration of the entire task was 44 s (± 17 s.d.).Overall, the average duration of the whole task for all the scenarios and information treatment is 46 s (± 23 s.d.).
The findings for crossing time (panel e) were in contrast to those for waiting and approaching.Specifically, participants took significantly longer to cross in the curbless scenario compared to both the status-quo ( p adj BY = 2.12e − 07 ) and inLED ( p adj BY = 4.68e − 06 ) scenarios.This result is consistent with the fact that the curbless scenario was the only one without a traffic light or in-pavement LEDs, allowing participants to take as much time as needed to cross.This could potentially explain the longer crossing times observed in panel e.
In comparing treatments across scenarios (level 2), the results indicate that the sequential treatment exhibited a strict ordering of status-quo < inLED < curbless.Likewise, the parallel treatment showed an ordering of status-quo ≈ inLED < curbless.

Displacement
In the following, we examine the participants' movement patterns, which encompassed both their actual physical movements in a play area of 10 m × 5 m and their ability to "teleport" within the VR environment.We first calculated the participants' speed, which was defined as meters moved per second.The standard deviation of speed ( σ speed ) served as a proxy for the uniformity in movement pace.According to panel g in Fig. 3, the standard deviation σ speed differed significantly across treatments in the curbless scenario.Specifically, participants in the sequential treatment exhibited more variation in their movement speed compared to those in the parallel treatment.Our analysis of the overall distance walked (panel h in Fig. 3) revealed the following ordering: status-quo < inLED ≈ curbless.This effect was particularly pronounced in the parallel treatment, as we observed a significant difference in distance walked between the status-quo and inLED scenarios for this treatment.

Head movement
During the VR experiment, head movement and displacement was logged to quantify the exploration of the environment [37][38][39] .The head movement for scenario curbless was not recorded due to a bug.Consequently, we only report significant differences between scenarios status-quo and inLED.Our analysis shows that pitch (i.e., looking up and down) and yaw (i.e., turning head left and right) differed significantly across the treatments in terms of both, their mean ( µ ) and standard deviation ( σ ) (panels i-l in Fig. 3).Specifically, participants looked less up and down in the inLED scenario compared to the status-quo scenario, while they turned their head more from left to right in the inLED scenario compared to the status-quo scenario..Panel a shows a significant difference between the status-quo and curbless scenarios ( N = 266 , p adj BY = 1.902e − 02 ).Moreover, panel c reveals a significant difference in the LF/ HF ratio between treatments: parallel viewers had a significantly higher ratio than sequential viewers ( N = 266 , p adj BY = 9.925e − 03 ).Specifically, panel b shows how parallel viewers had a significantly higher heart rate when approaching the crossing than sequential viewers ( N = 266 , p adj BY = 2.706e − 02).

Open-answer questions
The analysis of open-answer questions involved a two-stage process.First, all responses were read and labels were generated to summarise the answers in a meaningful way.Second, the labels were assigned using the LightTag software to ensure reproducibility.
The participants provided feedback on the advantages and disadvantages of the different street crossing scenarios (see "Questionnaires").Regarding the status-quo scenario, participants commonly noted advantages such as its safety ( n = 19 ), familiarity ( n = 14 ), and ease of understanding ( n = 10 ).However, the most commonly mentioned disadvantage was the wait times of pedestrians and the fact that there was only one designated spot for crossing the street ( n = 15).
The inLED scenario received mixed reviews from the participants.Among the advantages, the absence of a need to search for a crossing path for a long time was frequently noted ( n = 10 ).Moreover, participants appreci- ated the clear and visible markings of the crossing, particularly at night ( n = 1 ).Additionally, some participants found the scenario aesthetically pleasing ( n = 4 ).However, only a minority felt that the scenario was safe ( n = 5 ).The most commonly mentioned disadvantage was the perceived complexity of the scenario ( n = 14 ).Participants Each panel shows a different response measured from the participants' behaviour, which reflects statistically significant differences ( p < 0.05 ) between information treatments, scenarios, or their interaction (Information treatments × Scenario).Confidence intervals for the estimates are shown between brackets.The participant's activity was divided into four subtasks, namely familiarization with the virtual environment (a), approaching the crossing (b,c), waiting to cross (d), and crossing (e).The duration in seconds is shown on the y-axis for each subtask in panels (a-e), and for the whole task in panel (f).Panels (g,h) illustrate differences in uniformity of speed (measured as σ speed ) and the total walked distance within the VR environment.Panels (i-l) highlight differences in head movement.Specifically, we report the mean ( µ ) and standard deviation ( σ ) of change in pitch (i.e., looking up or down, see panels i,j) and yaw (i.e., turning head left or right, see panels k) as well as the cumulative total angle change in direction of the three axes per second (panel l).When shown, scenarios are grouped on the X-axis, featuring the aggregated distribution as a coloured box plot in blue, magenta, and bright orange.Violin plots show distributions for each information treatment within the same scenario: mint for sequential and orange for parallel.They are complemented by corresponding box plots displaying median values and interquartile ranges.For further details on the statistical methods see "Statistical details".also found the setting unfamiliar, which could be a potential hurdle to widespread adoption ( n = 8 ).Interest- ingly, a considerable number of participants expressed concerns about the expensive cost of implementing and maintaining the inLED scenario ( n = 6).
The curbless scenario generated mixed feedback from the participants.Many participants noted the convenience of being able to cross the street whenever they wanted ( n = 21 ).Some considered the scenario cost-efficient ( n = 4 ).However, the overwhelming majority of participants did not feel safe when crossing the street ( n = 24 ).Some participants assumed the absence of visual cues meant that car drivers had no obligation to stop for pedestrians ( n = 8 ).Furthermore, a few participants found the scenario to be complicated and tiring ( n = 6 ).Interestingly, some participants also reported that cars frequently stopped too close to them, while they were crossing the street, which was a notable disadvantage ( n = 3).
With regard to the different scenarios, participants were asked for their willingness to change their mode of transportation, and a range of responses was given.The curbless scenario received mixed feedback from the participants.Interestingly, the most common response was that participants did not want to be either a pedestrian ( n = 8 ) or a car driver ( n = 8 ) in the scenario.This might be due to the perception that scenario curbless had an increased probability that cars could hit someone.Moreover, some participants specifically stated a desire    For more details, see Table 2. ).However, some participants found being a pedestrian in this scenario to be satisfactory ( n = 5 ).Regarding the inLED scenario, a majority of participants stated that they were comfortable with being a pedestrian in the scenario ( n = 7 ).Similarly, for the status-quo scenario, a comparable number of participants reported being fine with being a pedestrian ( n = 8).
During the study, participants were asked to identify the differences they observed while crossing the street in different scenarios.The majority of participants noted that the crosswalk and traffic lights were the main differences they noticed ( n = 24 ).Notably, a significant number of participants identified safety as the primary variable that varied across scenarios ( n = 12 ).Interestingly, some participants believed that the primary differ- ence was in the level of difficulty ( n = 4 ) or familiarity ( n = 4).

Discussion
Our primary finding is that unconventional street layouts are less preferred and perceived as less safe in both the structured survey and open-answer questions.These results are mirrored by the heart rate data: we observe greater stress response in the curbless scenario, as it is unfamiliar, perceived as dangerous, and requires more cognitive resources to navigate.Hence, in our study, the status-quo street design is preferred over unconventional layouts.Nevertheless, in-street technology, such as presented in the inLED scenario, is preferred over the curbless scenario.Such technology can reserve areas for pedestrians and allows for flexible and adaptive use of space by time-of-day and day-of-week.
The inLED and curbless are two extreme examples of shared space scenarios.Although there is no clear definition of shared space, all definitions agree that shared space is characterized by the removal of a clear separation between different road users 41 and integrating vehicular traffic into social space 6 .Research on shared space has resulted in mixed results, varying from pedestrians not minding to share space 42 , to pedestrians preferring dedicated facilities 43 .From a driver's perspective, shared space results in anxiety and unease to drivers, resulting in possible lower speeds and more attention being paid to other shared space users 44 .
The open-answer questions shed more light on the perceived differences between the different scenarios.A frequently mentioned concern was safety, which aligns with the results of the structured question.The frequent mentioning of familiarity by participants suggests the significant cognitive effort required to adapt to new traffic paradigms and a strong preference for the status-quo scenario, which could potentially outweigh the benefits of adopting new transportation innovations due to the status quo bias 45 .While the safety distance to stop did not vary between the scenarios from the participants' perspective, the stopping distance appeared closer for the participants.This insight is valuable for planners, who need to take into account the effects of removing visual cues in street designs.
The inLED scenario took the longest to complete, which is consistent with the open-answer questions, where participants mentioned confusion due to the overwhelming information displayed.Interestingly, the curbless scenario without a timing mechanism took the longest to cross.This finding suggests that the convenience of crossing the road can be increased by giving more time to pedestrians.
One indirect measure of presence in VR is head movement.Increased head movement indicates that individuals are likely to behave and judge as in reality 46 .In two scenarios, head movement was recorded.Our study found that the scenario influenced participants' visual exploration patterns, with the inLED scenario requiring more focus on the ground plane and participants turning their heads from the left to right (head yaw), and looking more up-and-down (pitch) in the status-quo scenario.This is intuitive, as the novel aspect of this scenario was the hexagonal road pavers that needed to be explored, directing more attention to the left and right rather than up and down.Exploratory research furthermore indicated that an increased head yaw, or scanning, might be due to higher anxiety 47 .Lower self-reported safety for the inLED scenario also points to higher anxiety in this scenario.
The content analysis of the answers given to the open-answer questions delivered further intriguing insights.Participants commented on the diverging maintenance costs for the different scenarios, suggesting that they should be considered not only as road users but also as citizens and taxpayers.This indicates the importance of not only highlighting the immediate benefits and drawbacks for road users but also of considering organizational issues such as costs.
When asked about their willingness to change their mode of transportation in different scenarios, participants provided surprisingly mixed responses.In the curbless scenario, many participants were afraid of hitting pedestrians as car drivers and therefore explicitly stated not wanting to drive a car.On the contrary, other participants expressed fears of being hit by a car as pedestrians and, therefore, preferred to be car drivers.These findings highlight the importance of planners to consider the unintended consequences of their designs on the behaviour of road users if the safety and equity of road use are ultimately improved.
When asked to state the differences, many participants identified safety, familiarity, and difficulty as key variables that varied across scenarios.This suggests that participants were more focused on the affordance of the space 48 and the behavioural effects of the new installations than on physical features.This highlights the importance of considering how citizens use and feel about a space when designing transportation infrastructure.
In our experiments, parallel viewers tended to favour conventional scenarios over unconventional ones more frequently than sequential viewers.This finding indicates that having the ability to compare three street scenarios enabled participants to focus primarily on the drawbacks (instead of the advantages) of a novel scenario compared to a familiar one.It is likely not coincidental that the presentation of information made the greatest difference in questions relating to the most vulnerable traffic participants, namely children, and the transport mode with the greatest potential to cause harm: a car.Furthermore, our study found that parallel viewers needed a longer approach time to cross, implying greater caution than sequential viewers.Furthermore, the information treatment appears to influence the comprehension of the scenarios.Specifically, parallel viewers showed a more constant movement pace, implying they may have comprehended the single scenario better than sequential

Conclusions
Value of integrating people's feedback Overall, our findings highlight the importance of seeking early feedback in any urban planning action and aligning with current practices of incremental and tactical urban planning.Participants' assumptions that the absence of visual cues influences car driver behaviour emphasize the importance of street design elements (e.g.markings, signs, delineated areas), to ensure driver awareness of their obligations to stop for pedestrians.

Comparing between options improves feedback
We also found that the ability to compare multiple scenarios allowed participants to more easily identify the drawbacks of a novel scenario compared to a familiar one.Physiological data supported our qualitative findings, suggesting increased perceived and measurable stress in unfamiliar scenarios.

Communication is important
By comparing two communication strategies, this study points out how this can be as relevant as the actual content of the proposed changes in cities.This highlights the importance of informing people effectively, not only focusing on the technical design of proposals for successful policy-making.
Tangible, designed visual cues to support street use are important Furthermore, we highlight the importance of tangible design elements, such as signs and demarcated areas, which increase perceived safety.We also recommend retaining simple markings as they positively change the perception of space.

Consequences and behavioural changes centre people's attention
Finally, our study shows that participants were more focused on the consequences of changes, the affordance of space, and the behavioural effects of new layouts rather than objective differences.This underscores the need for a user-centred approach when designing streets to meet a community's preferences and needs.

Importance of people-centred urban planning innovation
In conclusion, our study has provided valuable insights into the impact of unconventional street layouts on road users' behaviours and perceptions.Our findings suggest that such layouts are generally perceived as less attractive, less safe, and less preferred by participants.This highlights the importance of designing streetscapes that meet community needs and preferences and providing opportunities to familiarize oneself with new scenarios.We recommend that planners and policy-makers consider multiple dimensions of street design innovations, including organizational issues, costs, and other benefits and drawbacks for citizens.Additionally, we urge planners to take into account citizen perception to prevent unintended consequences such as the switch to less sustainable solutions.

Limitations and future outlook
Real-world experiments are, of course, expected to provide the highest ecological validity, but they are unfeasible in the context of the future scenarios studied by us 49 .It is known, however, that the use of VR technologies to test responses of subjects to changes in built environments can provide high ecological and internal validity 50 through realism 29 and experimental control 34 .On one hand, ecological validity reflects the correspondence between experimental settings and real-world situations.On the other hand, internal validity allows to isolate causal relationships.Nevertheless, more extensive or longitudinal VR studies would be favourable, as subject familiarity with the scenarios may change over time, thereby increasing the likelihood that novel scenarios would eventually become acceptable.Hence, our VR study might underestimate medium-term public acceptance of new street layouts.Among others, the real-world observation of others navigating safely might enhance perceived safety.However, (perceived) safety issues during a transition phase would probably remain.
Additionally, our focus on futuristic scenarios (reliant on technology that is not yet fully established) may also limit current societal acceptance.In general, head movement and other users' physiological and behavioural response patterns might be underestimated compared to real-world scenarios, as the VR settings did not feature noise, weather conditions and other sources of environmental variability.Furthermore, a higher diversity of types www.nature.com/scientificreports/ of street users (including bikes, eScooters, and other vulnerable people) could help to increase the realism of the virtual environment.Accordingly, multi-user experiments might benefit from even higher levels of realism and complexity, in particular regarding enhanced interactions between vehicles, environment and people.Generally, well-controlled VR experiments contrast with real-life unpredictability, which might have a larger impact on physiological responses than our information treatment.
In this context, we want to point out some issues with heart rate (HR) and physiological sensing.While some measurements, for example, those of electrodermal activity (EDA) or heart rates have become an experimental standard to assess stress levels, some would also want to consider respiration rates, pose estimation through computer vision techniques, and eye tracking for gaze and pupil change analysis.Furthermore, the use for example of electrocardiography (ECG), which can improve HR collection data, electroencephalography (EEG), or even functional magnetic resonance imaging (fMRI) scans could provide additional insights.In this context, one needs to consider additional ethical issues, including the question of whether more invasive or pervasive monitoring would be justified by the benefits created.Many of these measurements would not be feasible or desirable in real-world applications or may raise significant issues with regard to the protection of data or subjects.One should also consider the principle limitations of data-driven approaches, particularly when inferred variables are involved, i.e. variables other than those measured or measurable directly.In fact, one should be mindful that many data science applications involve a considerable amount of challenges related to matters of context, interpretation, and causality 51 .
Further note, while the sample size of 43 participants used in this study with repeated measures appears to be limited, it is still comparable to other recent experiments using VR technologies studying pedestrian behaviour in built environments 29,30,34,[52][53][54][55] .Moreover, our research has still precision and statistical power to provide valuable insights into most of the differences within and between subjects explored.Nevertheless, future research might benefit from increasing the number and diversity of participants.Additionally, the study of effective communication strategies in the context of informing about public policy and urban design would benefit from a more comprehensive and detailed design of alternatives.
Another challenge in VR experiments replicating large-scale environments is space availability to effectively and realistically represent movement through the virtual world.Apart from "teleporting" and controller-based displacement already implemented in this study, the use of treadmills may provide a more consistent collection of data regarding displacement and pose, independently of environment size.Overall, the study of the spatial data could unravel insights regarding user occupancy of space and their relationship with other elements located in the environment.

Experimental design
The experiment was conducted in collaboration with the Decision Science Laboratory at ETH Zurich.The laboratory was responsible for participant recruitment and payment, ensuring that experimental data and personal information were kept separate and safe by design.The study was preregistered 56 and received IRB approval from the Ethics Commission of ETH Zurich (IRB Approval Date: 2021-01-28, IRB Approval Number: EK 2020-N-183).All experiments were performed in accordance with applicable guidelines and regulations.All participants provided written informed consent before they participated in the experiment and were compensated for their time according to the laboratory's payment scheme.
The participant was briefed to cross the street in three different scenarios.Through this, a participant would experience three different levels and strategies of implementation of new technologies applied to street design with the interplay of autonomous vehicles.To that end, we designed three different street environments (Fig. 5) for our VR experiments: the status-quo scenario served as a baseline, with the street characterized by regular sidewalks and a traffic light controlling the pedestrian crossing.The inLED scenario allowed for a dynamic allocation of uses, assuming a smart pavement consisting of hexagonal modular pavers 10,11 incorporating LEDs, which are capable of changing colour based on the intended use of the street.In contrast, the curbless scenario relied on smart vehicle sensing and the adaptive behaviour of people, enabling a dynamic use of the space.Consequently, these diverse scenarios put the participants in the position to answer a series of related questions.
Apart from the differences in designs, all three scenarios were designed similarly (Fig. 6).Each scenario featured a park adjacent to a straight long road with curves at both ends of the road.The curves blocked the line of sight and allowed for the generation of pedestrians and cars out of sight of the participant.The participant could freely move in this area but was discouraged from moving too far in one direction along one side of the street, namely by limiting teleportation to certain areas (i.e.alongside the street) and potentially by the instructors reminding them that their task was to cross the street.
Each of the three street scenarios was experienced twice in two different configurations by all the participants (i.e., within-subject factor), once during rush hour and once during the evening.The scenarios were designed such that street usage varied throughout the day.During rush hour, the street was characterized by a steady stream of commuters, while in the evening, it was used by families as a space for play and dining.Only the order of the scenarios was randomized across participants.
Prior to starting the VR task, participants filled out an SSS for establishing a stress level baseline for aftertreatment differences assessment (see a in Fig. 1).Then, participants watched a 3-min video to become familiar with the traffic environment.The presentation of the videos varied between subjects: participants in the sequential group watched videos from a bird's-eye view for every scenario separately in sequential order matching the randomized order experienced afterwards for performing the task in the three VR scenarios.Instead, participants in the parallel group watched the same videos, but on a split-screen, simultaneously in a side-by-side layout, allowing for a direct comparison across scenarios.This gave them additional qualitative information about each street design.In both cases, the window for the video clip of each scenario was depicted in the same size and aspect ratio: i.e., each scenario video window occupied 1/3 of the width of the screen, resulting in sequential screen layout having one single video playing with most of the screen empty, and parallel visualization having all the three videos filling the entire width of the screen (see b in Fig. 1).In Table 2, a summary of the treatments for the experiment design is shown with a motivation of the interventions, both for the between-subjects information treatments and for the within-subjects scenarios.
The task that participants had to perform was identical in every design scenario and configuration (see d in Fig. 1).Moreover, they had two attempts to perform the task per configuration in case of being hit by a car and failing.To complete the task, participants had three locomotion options, which could be freely switched between (1) physical walking and turning, (2) virtual walking and turning using a controller, and (3) teleportation.Teleportation allowed participants to move automatically to a desired location by pointing the controller and was only possible on the sidewalk and restricted in range to 8 m to prevent participants from simply teleporting across the street (Fig. 6).The virtual walking and teleportation options were introduced due to the large space requirements of the VR setting, although participants also had ample physical space to move around in a 10 m × 5 m play area.After ensuring a comfortable fit of the head-mounted display, participants were given two controllers and found themselves in a virtual tutorial environment.An instructor guided them through a fixed tutorial sequence for getting accustomed to VR and learning the different locomotion techniques (see f in Fig. 1).
After completing the task, in both configurations for each design scenario, participants were asked to fill in a survey assessing their emotional state and perception of the scenario (see e in Fig. 1 and "Questionnaires").To complete the survey, they removed the VR headset and switched to a desktop computer.The survey included questions related to stress, perceived safety, and willingness to cross and share the street.At the very end of the experiment, after performing the task in the three scenarios, participants were asked to reply to a set of openanswer questions regarding general impressions and behaviour change based on the scenarios, as well as general socio-demographic data.All these questions that were answered are listed in the paragraph titled ""6.6Questionnaires"".All the questionnaires were designed, hosted and served from the web service provided by Qualtrics and the resulting data stored in their servers.

Technical details
To gain a deeper understanding of how street design influences the walking experience of pedestrians 57,58 , we captured participants' physiological responses.Specifically, we measured participants' heart rate and heart rate variability, as well as their activity levels, using a breast strap (Garmin HRM-Pro) connected through a smartwatch (Garmin Forerunner 55) and a smartphone (HTC Desire 19+) via Bluetooth to the App Garmin Connect to synchronize and store the anonymous data.
For conducting the VR part of the experiment, our setup consisted of a Meta Quest 2 head-mounted display and hand controllers connected to a desktop computer (Intel®Core TM i7-11700 11th generation [8 Cores, 16 MB Cache, 2.5 GHz up to 4.9 GHz], NVIDIA®GeForce®RTX 3070 TM 8 GB GDDR6; 32 GB [16 GB x 2, DDR4, 2.933 MHz]; M.2-PCIe-SSD with 1 TB + SATA with 1 TB, 7200 1/min) via Air Link.The VR application was developed in Unity 2020.3.32f1 using the C# programming language.The virtual environment depicted in Fig. 5 was designed with high-quality assets from the Unity Asset Store ("Module Based City Pack" and diverse cars).Realistic, rigged 3D models of pedestrians were taken from Mixamo 59 .

Data collection and pre-processing
Forty-three participants ( n participants = 43 ) were randomly assigned to the sequential or parallel information treat- ment group (between-subjects factor) and completed the task of crossing the virtual street in a randomized order across the three scenarios (within-subjects factor): status-quo, inLED, and curbless.Each scenario was repeated at least twice, following the fixed order for the two configurations: morning rush hour (RH) and evening (EV) (see panel d in Fig. 1).In cases of task failure, additional attempts were made, providing participants with second opportunities to cross the street again in the formerly failed scenario and configuration (see Tables 1, 2).No participant reported experimental fatigue either verbally during the experiment or in the free-answer questionnaire.
Physiological measurements were taken continuously from the beginning of the experiment until the end resulting in a continuous time series (more details in paragraph "Heart rate data").Behavioural data were taken separately at every VR session, defined as every run to perform the task in a given scenario and configuration until the participant accomplished crossing the street or failed because of being hit by a car ( n sessions = 266 ).Behavioural data were computed from the VR environment logs, which recorded the position (x, y, z) and orientation of the VR headset in three angles (pitch, yaw, roll).From this raw log data, distance walked, speed, and angle change for pitch, yaw, roll, and fortotal angle (as the sum of pitch and yaw) were computed.The structured self-reported survey (comprising SSS, acceptance, safety perception and assessment of willingness to cross) was answered after completing an entire scenario in both configurations ( n structured surveys = 129 ).Physiological and behavioural time series were synchronized and matched.Later physiological continuous data was cropped according to the starting and ending time of each session to have a combined time series per VR session including physiological and behavioural data.
We recorded the first-person perspective in VR of participants for the entire duration of the experimental session.These recordings were later manually timestamped, which allowed mapping all data points to four different subtasks.The first subtask, which involved participants exploring the virtual environment, was labelled as familiarization (1).Following this, participants typically walked towards the curb (or perceived border of the area suitable for cars) in preparation for the road-crossing task, which we referred to as approaching (2).Subsequently, each participant had to wait in front of a (virtual) crosswalk, a subtask we termed waiting (3) until the red light turned green or he felt it was safe to cross.Finally, participants completed the task by crossing the street, a subtask www.nature.com/scientificreports/we called crossing (4).The subdivision of the task into these subtasks facilitated comparisons of completion times across the different stages as well as highlighting particular differences within the task.Consequently, time series are cropped as well to match these subtasks.In the case of the lack of one of the subtasks (e.g., because a participant did not wait before starting to walk or before crossing), missing values are imputed with the closest 10 s of data from the contiguous subtasks.Time series for each session are summarised by computing their means ( ν ), medians ( η ), and standard deviations ( σ ).These values were later used for statistical analysis.The answers to the structured self-reported survey were normalized for each participant by computing the differences between each value reported after experiencing a scenario and the initial baseline value registered by each participant at the beginning of the experiment.These differences are later used in the statistical tests.

Heart rate data
Participants in VR environments are exposed to a variety of stimuli that can elicit a variety of physiological and psychological responses 60 .A variety of sensors, including heart rate sensors, head movement sensors, and displacement sensors, can be used to measure these responses.Heart rate variability (HRV) is a measurement of the time difference between successive heartbeats.It is a useful indicator of the body's physiological response to stress and other stimuli.Heart rate sensors can be used in VR experiments to track participants' HRV as they navigate virtual environments.This data can reveal how participants react to various aspects of the environment, such as visual cues or sound effects.
We collected heart rate data continuously, at a rate of 1 Hz from the start to the end of the VR experiment.Outliers in the dataset were removed to guarantee data quality.In such an experimental setting, it is possible to have outliers in the collected heart rate data.Outliers can be caused by a variety of factors, including measurement errors, physiological abnormalities, and participant noncompliance.While pre-processing the data, we looked at the outliers and removed them in a way that did not skew the findings.We resampled and adjusted the data to one sample per second [61][62][63][64] in order to provide a good accuracy to the HRV analysis and sufficient sampling points 65 .Timestamps between different sensing sources, including the VR environment logs and sensing devices, were matched to accurately align data across sources.The whole heart rate time series for each participant was split into individual sessions corresponding to each VR scenario.
In VR experiments, analyzing the power spectral density (PSD) of physiological signals such as heart rate can provide insights into the autonomic nervous system response, which can help in understanding the participants' emotional or cognitive states 66,67 .The PSD represents the power of various frequencies in a signal mathematically.The PSD gives information about the distribution of power at different frequencies in the heart rate signal in the case of heart rate variability.The PSD's HF component is linked to parasympathetic (vagal) activity, whereas the PSD's LF component is linked to sympathetic and parasympathetic activity 68 .
We analyzed the PSD of heart rate data from various participants to investigate the effects of diverse scenarios in a VR experiment on the autonomic nervous system.We used a wearable device (Garmin HRM-Pro) to collect heart rate data, and each participant went through a series of VR scenarios, each intended to elicit a particular emotional response.We used the Welch method to analyze PSD because it reduces noise and gives a more accurate estimate of the PSD 69 .To obtain a more accurate estimate of the PSD, the Welch technique divides the signal into overlapping segments and then averages the periodograms of these segments.The technique reduces noise caused by the signal's non-stationarity by using overlapping segments, and it improves the estimate of the PSD by averaging the periodograms.The PSD was then integrated into the LF (0.04-0.15 Hz) and HF (0.15-0.4 Hz) bands to determine the total power in each band.The LF/HF ratio was determined as the ratio of LF power to HF power.The LF/HF ratio is a commonly used metric for autonomic nervous system activity.A higher LF/HF ratio is commonly read as a sign of sympathetic dominance or a shift in the balance of sympathetic and parasympathetic activity toward sympathetic dominance.This could be due to stress, nervousness, or other conditions that cause sympathetic activation.A lower LF/HF ratio, on the other hand, is frequently linked with parasympathetic dominance or a shift towards parasympathetic activity, which may be associated with relaxation and recovery 40 .

Statistical details
To characterise the independent and dependent variables of the experiment, we employed the Shapiro-Wilk test 70 to check normality and the Levene test 71 for equal variance.Different statistical methods were applied depending on the type and distribution of the data for assessing the association between variables and the informational treatments and scenarios.However, as the majority of the experiment variables did not follow a normal distribution, we utilized non-parametric tests to evaluate differences across treatments and scenarios.
For categorical variables, both independent (i.e., gender, experience with VR experiments, ownership of a car, driving license, and walking habits,) and dependent (i.e., counts of failed attempts for crossing the street, favourite scenario reported at the end of each experiment) a two-tailed Fisher's exact test 72 was used on the contingency table of occurrences for each treatment group due to the small sample size of some of them (i.e., below 5).
We statistically tested the independence of the treatments.The only variables from the socio-demographic data that were not completely randomized and showed unbalance with regards to the treatment groups were ownership of a car ( N = 43 p two-sided Fisher = 1e − 03 ), driving license ownership ( N = 43 p two-sided Fisher = 2.499e − 02 ), and reported previous experience in VR experiments ( N = 43 , p two-sided Fisher = 5e − 04 ).Hence, we controlled for the effect of these covariates in combination with the main interventions (information treatment, and scenario) in all the statistical analyses and they didn't show any significant effect on the responses ( p > 0.05).
For the analysis of the effects of the interventions, we performed an Aligned Rank Transform (ART) ANOVA test.This allowed us to perform factorial non-parametric analyses when violations in the assumptions of normality and variance were present and it could handle repeated measures 36 as in the case of the mixed experiment designed.
• What is your preferred mode of transportation for your daily chores (work, groceries, etc.)? (walking, bike, public transportation, motorbike, car) • Do you walk (> 500 m) daily?(Y/N) Short Stress State (SSS) Questionnaire All of the items of the short stress state questionnaire 35 are answered on a 10-point scale, once before and subsequently after each scenario.
• I feel dissatisfied.
• I feel alert.
• I feel depressed.
• I feel sad.
• I feel active.
• I feel impatient.
• I feel annoyed.
• I feel angry.
• I feel irritated.

Safety, willingness, attractiveness
In addition to the short stress state questionnaire, the following questions were asked after participants encountered each street scenario: • How safe do you feel due to the proximity and number of pedestrians/cars?(1-7) • How willing are you to cross the street with a kid/at night?(1-7) • Please rate how attractive you found the scenario as a pedestrian (0-5)

Figure 1 .
Figure 1.Design, phases, devices and data collection during the experimental setting.

Figure 2 .
Figure 2. Self-reported responses in the structured survey.Each panel presents a different response to the structured survey, which shows statistically significant differences ( p < 0.05 ) between information treatments, scenarios, and their interaction (Information treatment × Scenario).Confidence intervals for the estimates are shown between brackets.Panel (a) displays the distributions for the self-reported "active" feeling on the Short Stress State (SSS) questionnaire.Panels (b-f) show responses with statistically significant differences for scenarios and interactions between interventions.When shown, scenarios are grouped on the X-axis, featuring the aggregated distribution as a coloured box plot in blue, magenta, or bright orange.Violin plots show distributions for each information treatment within the same scenario: mint for sequential and light orange for parallel.They are complemented by corresponding box plots displaying median values and interquartile ranges.For further details on the statistical methods see "Statistical details".

Figure 4 ,
Figure 4, displays the low frequency (LF)/ high frequency (HF) ratio (see "Heart rate data").It is interpreted as an index of autonomic balance, where higher values indicate greater sympathetic dominance and lower values indicate greater parasympathetic dominance40 .Panel a shows a significant difference between the status-quo and curbless scenarios ( N = 266 , p adj BY = 1.902e − 02 ).Moreover, panel c reveals a significant difference in the LF/ HF ratio between treatments: parallel viewers had a significantly higher ratio than sequential viewers ( N = 266 ,

Figure 3 .
Figure 3. Behavioural responses.Each panel shows a different response measured from the participants' behaviour, which reflects statistically significant differences ( p < 0.05 ) between information treatments, scenarios, or their interaction (Information treatments × Scenario).Confidence intervals for the estimates are shown between brackets.The participant's activity was divided into four subtasks, namely familiarization with the virtual environment (a), approaching the crossing (b,c), waiting to cross (d), and crossing (e).The duration in seconds is shown on the y-axis for each subtask in panels (a-e), and for the whole task in panel (f).Panels (g,h) illustrate differences in uniformity of speed (measured as σ speed ) and the total walked distance within the VR environment.Panels (i-l) highlight differences in head movement.Specifically, we report the mean ( µ ) and standard deviation ( σ ) of change in pitch (i.e., looking up or down, see panels i,j) and yaw (i.e., turning head left or right, see panels k) as well as the cumulative total angle change in direction of the three axes per second (panel l).When shown, scenarios are grouped on the X-axis, featuring the aggregated distribution as a coloured box plot in blue, magenta, and bright orange.Violin plots show distributions for each information treatment within the same scenario: mint for sequential and orange for parallel.They are complemented by corresponding box plots displaying median values and interquartile ranges.For further details on the statistical methods see "Statistical details".

Figure 4 .
Figure 4. Physiological responses.Each panel presents statistically significant differences ( p < 0.05 ) of low- frequency/high-frequency (LF/HF) ratio metric based on the heart-rate data for different subtasks within the VR session task: subtask approaching in panels (a,b) and whole task in panel (c).Confidence intervals for the estimates are shown between brackets.When shown, scenarios are grouped on the X-axis, featuring the aggregated distribution as a coloured box plot in blue, magenta and bright orange.Violin plots show distributions for each information treatment within the same scenario: mint for sequential and light orange for parallel.They are complemented by corresponding box plots displaying median values and interquartile ranges.For further details on the statistical methods see "Statistical details".

Figure 5 .
Figure 5. Participants crossed a street wearing VR headsets in three different design scenarios: conventional traffic light controlled pedestrian crossing (status-quo, left), dynamic pavement for right-of-way allocation (inLED, centre), and flat street (curbless, right).Additionally, each traffic scenario was presented for two different times of the day with distinctive lighting, use, and traffic conditions: morning rush hour (top row) and evening traffic (bottom row).For more details, see Table2.
Figure 5. Participants crossed a street wearing VR headsets in three different design scenarios: conventional traffic light controlled pedestrian crossing (status-quo, left), dynamic pavement for right-of-way allocation (inLED, centre), and flat street (curbless, right).Additionally, each traffic scenario was presented for two different times of the day with distinctive lighting, use, and traffic conditions: morning rush hour (top row) and evening traffic (bottom row).For more details, see Table2.

Table 2 .
Characterization of experimental variations: information treatment, scenarios and configurations with its motivations.
SequentialEach video only shows one scenario at a time Each scenario is displayed for 1 min Scenarios are displayed consecutively (one after another) Each scenario is displayed once Size of video window is 1/3 of the treatment parallel Total visualization time of all videos is 3 min Baseline for explanation of scenarios Vol.:(0123456789) Scientific Reports | (2024) 14:4571 | https://doi.org/10.1038/s41598-024-55073-xwww.nature.com/scientificreports/ to switch from being a pedestrian to being in a car (driver) ( n = 6