Pedestrian Profiling Based on Road Crossing Decisions in the Presence of Automated Vehicles: The Sorting Hat for Pedestrian Behaviours and Psychological Facets

Abstract

Automated Vehicles (AVs) are being developed with the aim to reduce the occurrence and severity of Road Traffic Crashes (RTCs). Studies suggest AVs may improve the safety of Vulnerable Road Users (VRUs), particularly on road crossings. However, exposure to novel technology over time may lead to behavioural adaptation. Thus, understanding VRUs’ behavioural intentions towards AVs is crucial for their safe integration into traffic. We investigate four external factors pedestrians consider when crossing a road in front of an AV. An online questionnaire with 281 participants assessed crossing intentions, focusing on road gradient, weather, pedestrian–AV distance, and AV type. Personality traits and self-reported behaviour were measured. Anderson’s experimental protocol revealed all factors significantly influenced crossing decisions. Using hierarchical clustering followed by K-means clustering, the participants were classified into three different profiles: risk-averse, resolute, and indecisive pedestrians. We provide evidence of a strong link between crossing decisions, reported behaviours and psychological facets while interacting with an AV at crossings. Pedestrian profiling allows targeting preventative measures for groups based on unique characteristics, maximising efficiency thereof. Furthermore, pedestrian profiling can inform AV’s driving style to support safer road interactions. This is salient for resolute pedestrians, who take more risks, which may lead to severe RTCs.

1. Introduction

Vulnerable Road Users (VRUs) are a group of road users who use non-motorised modes for mobility and/or have reduced mobility functions [1]. Compared to other road users, this group is most prone to sustaining bodily injuries in Road Traffic Crashes (RTCs) involving a four-wheeled motorised vehicle [2]. VRUs (pedestrians, cyclists, or motorcyclists) account for over half of total RTCs worldwide [3]. The data regarding fatal RTCs in the UK and Australia for 2021 and 2022 reveal that over 50% of the victims were VRUs [4,5,6].

Numerous factors affect the occurrence and severity of RTCs, for example, environmental conditions (e.g., hostile weather), infrastructures (e.g., inadequate markings on road surfaces), vehicular characteristics (e.g., ill-serviced vehicles) or human characteristics (e.g., risk-taking behaviour or inaccurate judgement of own or others’ capacity) [7,8,9]. Traffic safety research points out human errors as one of the causes of RTCs. However, non-human parameters (for example, environmental or vehicular) also influence how individuals perceive the road environment and make decisions [8]. These factors ultimately lead the individual to take either intended or unintended actions, expressed as slips, lapses, mistakes or violations, collectively called human errors [10]. There have been various physical and non-physical interventions to counteract or reduce human errors and improve road safety, including using automation in driving. However, will automation be enough to reduce the occurrences of pedestrian errors while they interact with AVs at road crossings?

Progress in driving automation has led to the concept of Automated Vehicles (AVs) on roads. Automation refers to the accomplishment of a task by a machine which was either carried out previously by a human or not carried out by humans as finely as compared to the machines [11]. Some of the anticipated functionalities of AVs include sensing the environment, obeying traffic rules and guidelines and ensuring road user safety [12]. A scenario where a VRU (for example, a pedestrian) might interact with an AV is when they are trying to cross a road or in shared spaces where VRUs and AVs interact in less formal boundaries [13]. As AVs are a new concept and are being gradually trialled and tested on roads, some degree of uncertainty and unwillingness from pedestrians can be expected in this novel situation [14]. A questionnaire conducted for a case study revealed that most of the sample considered fully or semi-automated vehicles to be too risky for their safety and would prefer to drive the vehicles by themselves [15]. Currently, pedestrians use eye contact and gestures to communicate with human drivers. However, there is an absence of these human components in AVs, which makes it difficult to understand the AV’s intention [16]. Thus, it is important to understand how pedestrians would interact with an AV in different road environments to design and develop AVs safely.

The factors causing RTCs also influence pedestrians’ risk perception and road decision-making. Risk perception is the subjective assessment and the concern towards a possible occurrence of a dangerous situation [17]. Risk perception is dynamic and varies from person to person. Behavioural changes can be expected when there is a change in exposure towards stimuli or environmental conditions [18]. The change in exposure can lead to a change in risk perception, either by expressing less cautious behaviour in unsafe conditions or more cautious behaviour in safer conditions [18]. This is known as risk compensation or risk homeostasis [19]. To make safe crossing decisions, pedestrians need to be able to perceive the true level of risk and expose themselves to acceptable levels [20]. As decision-making is linked to risk perception, it is necessary to understand the influence of various internal and external factors on pedestrians’ road-crossing decisions.

In road-crossing contexts, risky behaviours influence decision-making [21]. An example of a scale that is used to analyse pedestrians’ road use and crossing behaviour is the Pedestrian Behaviour Questionnaire (PBQ), which has been utilised to explain the link between behaviours and decision-making [22,23,24]. Research can also benefit from gathering additional insights into pedestrians’ personality traits and how they influence their decision-making. An example of a scale that encompasses personality traits in pedestrians is the Big Five Inventory (BFI) by John and Srivastava [25,26,27,28].

In short, researchers need to understand not only what external factors affect pedestrians’ road crossing decisions while interacting with AVs but also how their behaviour and personality traits influence the decisions they make. This leads to the following section explaining the rationale of the present study.

2. The Present Study

2.1. The Rationale for the Selection of Environmental Factors

Many factors influence pedestrians’ decision-making regarding road crossings. In this research, we focus on four factors to investigate: road gradient (uphill vs. downhill), weather conditions (sunny vs. rainy), the longitudinal distance between the participant and AV (80 vs. 40 m) and the type of AV (small car vs. lorry). They are the factors that have been overlooked in existing research or have inconclusive effects on pedestrians’ road crossing decisions, or that pedestrians commonly experience, yet the effect of vehicle automation on them remains largely unexplored.

Road gradient was selected due to limited research on its impact on pedestrian crossing decisions. While factors like road width, markings, and signals have been studied [29,30], the effect of road gradient (uphill or downhill) on pedestrian–AV interactions remains unexplored, to the authors’ knowledge. Additionally, DeLucia [31] established that visual looming (i.e., increased risk perception as an incoming object moves closer to the observer and reaches a certain threshold, at which point the observer perceives a collision is about to occur) affects the crossing decision. Thus, an exploration of the influence of visual looming of an AV (change in risk perception) with change in road gradients on pedestrians’ crossing decisions was decided to be carried out in this research.

Weather conditions were included due to mixed findings in existing studies. Movahhed et al. [32] found pedestrians took riskier crossing decisions in the rain, while Sun et al. [33] observed more caution during rainy weather with Human-Driven Vehicles (HDVs). For AVs, adverse weather increased pedestrian–AV crash instances [34]. Thus, analysing whether pedestrians perceive more risks while interacting with an AV under rainy conditions can be interesting.

Existing research has manipulated various Time-to-Arrival (TTA) values to study variations in pedestrians’ road-crossing decisions. TTA is the time available to cross before a vehicle arrives, which underpins gap acceptance (Petzoldt, 2014; as stated in Tian et al. [35]). Wang et al. [36] found that pedestrians were less willing to cross at a 2 s TTA than at 5 s. For the questionnaire, we presented gap acceptance as distances (in metres) rather than TTA (in seconds) for easy visualisation by the respondents. Based on the literature, TTA values of 3 s and 6 s were converted into longitudinal distances of 40 m and 80 m on a standard speed limit of 30 mph (48 km/h) urban road, as shown in Equation (1) below.

$$\mathrm{L}\mathrm{o}\mathrm{n}\mathrm{g}\mathrm{i}\mathrm{t}\mathrm{u}\mathrm{d}\mathrm{i}\mathrm{n}\mathrm{a}\mathrm{l} \mathrm{d}\mathrm{i}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e} (\mathrm{m}\mathrm{e}\mathrm{t}\mathrm{r}\mathrm{e}\mathrm{s}) = {\displaystyle \frac{\mathrm{S}\mathrm{p}\mathrm{e}\mathrm{e}\mathrm{d} \mathrm{L}\mathrm{i}\mathrm{m}\mathrm{i}\mathrm{t} (30 \mathrm{m}\mathrm{p}\mathrm{h})}{\mathrm{T}\mathrm{T}\mathrm{A} (3 \mathrm{a}\mathrm{n}\mathrm{d} 6 \mathrm{s})}}$$

(1)

Research on whether a vehicle’s external appearance affects pedestrian crossing behaviour was found to be inconclusive. Bayley et al. [37] found that SUVs were perceived as more threatening than smaller cars like the Volkswagen Beetle, which appeared friendlier due to their shape. However, Dey et al. [14] observed that vehicle behaviour (yielding or non-yielding) had a greater influence than its appearance or futuristic design. As noted by McGuire et al. [38], the visual looming phenomenon suggests that larger objects, such as lorries, are perceived as more threatening due to looming bias, which varies with cognitive load, psychological state, and emotions.

Gap acceptance values vary depending on inter-individual differences and situations. Changes in situations include environmental situations (e.g., weather conditions) or personal factors such as task urgency or carrying heavy luggage [39]. It was also found that gap acceptance varies with variance in time headways, the number of times attempted to cross a road by the pedestrian, and the type of vehicle [40]. Furthermore, Tian et al. [35] concluded that in a scenario with two vehicles with the same time gap, pedestrians make riskier crossing decisions when the vehicle approaches them at a higher speed. This study explored whether the same effect would exist if the vehicles were of different types and had different time gaps but the same speeds. In this way, the selected four factors are interlinked with each other. Hypotheses for these factors on pedestrians’ crossing decisions while interacting with an AV are listed in Section 2.5.

2.2. The Rationale for the Selection of the Protocol and Scales

2.2.1. Anderson’s Experimental Protocol

Many studies have analysed the influence of various factors on pedestrians’ road-crossing decisions. However, there is a lack of determination of the relationship between the factors, i.e., how they influence each other. An example of a standardised and validated experimental protocol used to overcome that limitation is Anderson’s experimental protocol, which was put forth by Anderson [41] based on Information Integration Theory (IIT). IIT is a model that reveals the cognition processes handled by humans while making daily judgements [42]. The main objective of this theory is to assess the multiple causation theory, i.e., how a decision (a single behaviour or an outcome) is made based on the function of multiple available sources [42,43].

Anderson’s experimental protocol is based on developing a complete factorial plan of the variables under investigation, and the analysis yields an explanation of how the interaction between the variables (i.e., the factors and sub-factors) influences the outcome variable (e.g., intention to cross a road). This way, researchers get insights into not just the influence of a factor on the dependent variable but also the influence of all the possible combinations of the factors [44,45,46]. A benefit of using this protocol is that it increases immersion in the scenarios presented to the participants by using a first-person perspective in the questions (for example, “Would you cross the road?”). It enables the investigation of potential interactions between the factors simultaneously, thus making it more reliable than traditional methods like surveys or interviews [47,48].

This protocol has been implemented and validated successfully in diverse research areas. Monsaingeon et al. [44] used Anderson’s experimental protocol to analyse the participants’ decision to deactivate the automated system on an AV when the automatic interface, weather, road markings and road curvature change. Delmas et al. [46] analysed how drivers’ speed preference varies with levels of automation and other environmental stimuli using Anderson’s experimental protocol. The usability of level 3 automated vehicles was evaluated using the same protocol [48]. All these studies used a series of questions to describe the scenarios. This protocol allows us to analyse the decisions made by the participants. However, we require different measures to understand the decision-making processes, personality traits and potential errors when crossing in front of AVs. These measures are further described below.

2.2.2. Pedestrian Behaviour Questionnaire

A widely used, standardised and validated version of the PBQ is the 20-item PBQ version by Granié et al. [49], encompassing five categories: positive behaviour, violations, lapses, errors, and aggression. Cronbach’s alphas (α) for violations (0.84), lapses (0.83), errors (0.79), and aggression (0.70) showed that all the dimensions except positive behaviour (0.53) had reliable internal consistency [49]. Recent studies have employed the PBQ to examine pedestrian behaviour and the occurrences of the PBQ categories by age, attitudes to road safety, and decision-making across cultures [24,50]. PBQ has also been used to understand drivers, cyclists and pedestrians’ acceptance towards AVs [51]. However, PBQ results are unique to individuals and cannot be categorised purely by scores. Since PBQ focuses on behaviour rather than personality traits, we also incorporated the Big Five Inventory by John and Srivastava [25] in our study.

2.2.3. Big Five Inventory

We used the 44-item Big Five Inventory (BFI) by John and Srivastava [25] to explore the link between pedestrians’ decision-making, road use behaviours and personality traits. The 44 questions in the BFI are grouped into the following personality traits: openness, conscientiousness, extraversion, agreeableness, and neuroticism [25]. The BFI is a widely used scale for investigating personality variability and its impact on decision-making [25,28]. Previously, ref. [52] checked for internal reliability of the BFI dimensions. All the items, openness (0.84), conscientiousness (0.83), extraversion (0.86), agreeableness (0.82), and neuroticism (0.85) had high internal reliability.

Studies have linked PBQ and BFI in various contexts [53,54,55] but not in pedestrian–AV interactions. For example, Chien et al. [26] and Aghabayk et al. [28] used BFI to examine perceptions toward automation and anger expression on roads, respectively. More recently, BFI was also used to find a relationship between users’ personality and acceptance of AVs [56].

2.3. Novelty of the Research

The research was aimed at furthering the understanding of the interaction between factors and modalities affecting pedestrians’ decision-making and, ultimately, their safety while they mentally process the stimuli as an AV approaches them. We selected the four factors with two subfactors each, as explained in Section 2.1, with road gradient being the factor that has not been studied in the context of pedestrian–AV interactions. The interactions between the four selected factors and their sub-factors are also novel, as they have not been studied together before.

We used Anderson’s experimental protocol to design the experiment and analyse the road crossing decisions for reasons mentioned in Section 2.2.1 rather than using conventional survey methods. Existing research has used Anderson’s experimental protocol to analyse the influence of various factors on, for example, difficulty in the takeover of AVs [48] or comfort in partially automated cars [45]. However, these studies fail to make the connection between road use behaviours and/or personality traits of the individuals on their takeover difficulties or perceived comfort in an AV, respectively. Thus, our study also conducted an exploratory analysis to explore if declared road-crossing behaviour and psychological traits could be linked to road-crossing decisions.

2.4. Research Objectives

The objectives of this research are:

• To investigate the effect of selected factors influencing pedestrians’ road crossing decisions while interacting with an AV.
• To quantify the effect of each factor and the likely interactions on pedestrians’ crossing decisions.
• To explore potential interactions between pedestrians’ decisions and their declared behavioural and/or psychological facets on road crossings.

2.5. Hypotheses of the Research

Based on the observations from previous research and the identified gap in the literature presented, we posit five hypotheses in the following paragraphs.

Hypothesis 1. 

Based on the looming effect, we posit that pedestrians are more likely to cross the road when an AV is approaching them from down a hill compared to when it is approaching from up a hill.

Hypothesis 2. 

Based on results analysing the frequencies and severities of pedestrian–AV crashes in different weather conditions, we posit that pedestrians are more likely to cross the road under sunny weather conditions compared to rainy weather conditions.

Hypothesis 3. 

Based on TTA values, we posit that pedestrians are more likely to cross the road when the AV is 80 m away from them compared to when the AV is 40 m away.

Hypothesis 4. 

Based on looming bias, we posit that pedestrians are more likely to decide to cross the road when a small, automated car is coming towards them compared to an incoming automated lorry.

Additionally, a fifth hypothesis is also established by combining H1–H4 for the potential interaction between the four factors:

Hypothesis 5. 

Pedestrians are more likely to decide to cross the road in the scenario that a small car approaches them from down a hill in sunny weather and 80 m away, and that the opposite effect is to be expected when a lorry is approaching from up a hill in rainy weather and 40 m away.

3. Material and Methods

3.1. Participants

Pedestrians who were English speakers, living in the UK or Australia, and above 18 years of age were eligible to participate in the study. G*Power software version 3.1.9.7 was used to find the appropriate sample size required for the statistical validity of the questionnaire [57]. Estimated effect size values of the factors being analysed are needed for the estimation of sample size. Some studies, e.g., Karageorghis et al. [58], use averaged values of effect sizes derived from previous studies that analysed similar factors to determine the appropriate sample size. However, the exact combination of four factors has not been performed before in existing studies, to the authors’ knowledge. Thus, the effect size of the factors was assumed to be moderately strong, i.e., 0.1 (between large and medium effect size values) [59]. Thus, the appropriate sample size for a ${\mathsf{\eta }}^2=0.1$ number of groups = 1 (within-group study), number of measurements = 16 (2 × 2 × 2 × 2), alpha value 0.05 and power 0.8, was 107. The correlation among repeated measures was set at 0.1 to achieve an adequate but conservative sample size estimate [60]. We also consulted existing studies that employed Anderson’s experimental protocol with a similar or greater number of scenarios. A study with 3 × 2 × 2 × 2 (24 scenarios) had collected data from 202 participants [45], and a study with 2 × 2 × 2 × 2 (16 scenarios) had collected data from 201 participants [44].

The participants were recruited via social media. A total of 365 responses were collected. Incomplete responses were filtered out, and 281 responses (243 from the UK and 38 from Australia) were included in the analysis. There were 89 women (71 from the UK and 18 from Australia) and 192 men (172 from the UK and 20 from Australia). The participants had the option to choose “other” and specify the details regarding their gender. However, this sample group was removed due to the small sample size (n = 2). The median driving experience of participants with a driving licence was seven years, and the median walking minutes per day was 30. The respondents were asked about existing issues that currently affect their mobility. Only one respondent reported that they do, thus this response was not included in the analysis to remove potential outliers.

The participants had the option to join a lucky-draw contest for either £15 (GBP) or $30 (AUD) (three winners for each country). The online survey was advertised through social media, posters and flyers and kept open for three months, from July to September 2023. Ethical approval from Coventry (P150150) and Deakin University (SEBE-2023-19) was obtained for this online questionnaire before data collection. The respondents were also informed that the collected data was anonymised, confidential, and compliant with GDPR.

3.2. Protocol and Scales

A protocol and two self-reported scales were used in the questionnaire. Anderson’s experimental protocol [41] was implemented to analyse how pedestrians made decisions while crossing a road as an AV approached them.

Following the questions on road crossing scenarios, a 20-item Pedestrian Behaviour Questionnaire (PBQ) developed and validated by Granié et al. [49] was used to investigate self-reported pedestrian behaviour while they walk or cross a road. The PBQ scorings ranged from one (extremely infrequently or never) to six (extremely frequently or always). The instructions in Granié et al. [49] were translated from French to English. Next, the 44-item Big Five Inventory (BFI) by John & Srivastava [25] was used to uncover participants’ personality traits. Participants rated BFI statements on a five-point Likert scale, classifying from one (I disagree strongly) to five (I agree strongly).

3.2.1. Scenario Composition

Four environmental factors were selected to study pedestrians’ road crossing decisions, and Anderson’s experimental protocol enables all factors and their possible interactions to be compared consistently at the same time with the help of a questionnaire [61]. According to the methodology set by Anderson [61], different personas need to be designed to increase the immersion of the participants in the scenarios. This, in turn, is expected to increase the validity/reliability of the answers provided by the participants. This means that, instead of simply asking them if they would cross a road or not, they will be allocated protagonists based on their age and gender. Then, they were asked questions about their road-crossing decisions.

Table 1. Allocated protagonists based on the age and gender of the participant.

Protagonist Names	Age	Gender
Olivia	18–40	Female
Agatha	41 years or older
Harry	18–40	Male
David	41 years or older
Alex	18–40	Other
Andy	41 years or older

shows the featured protagonists differentiated by their age and gender.

The factors shown in

Table 2. Factors and their dimensions included in the scenarios (i.e., independent variables).

Factor	Dimensions
Road gradient	Uphill vs. downhill
Weather conditions	Rainy vs. sunny
Longitudinal distance between the pedestrian and the AV	80 vs. 40 m
Type of vehicle	Small car vs. lorry

were manipulated to construct a full factorial plan, which formulated 16 scenarios in a within-participant experimental design.

An example of a scenario for a female between 18 and 40 40-year-old is: “Olivia sees a small car about 80 m away. It is travelling down a hill towards her. It is a sunny day. If you were in the same situation as Olivia, would you cross the road now?” Additionally, we ensured the respondents would remember that the vehicle was automated by mentioning it on top of every page as they answered the questions. We also included that the AV was approaching with a constant speed but from varying distances (80 or 40 m) in the instruction paragraph.

3.2.2. Scenario Instructions

At the start of the questionnaire, the participants could view the plain language statement containing information about the possible risks, aims and objectives of the questionnaire. Next, they were asked to provide their consent by clicking on the “I am not a robot” sign and by completing a CAPTCHA verification. Subsequently, they were provided with a brief description of scenarios, each accompanied by a picture (Figure 1). The instructions provided to the participants can be found in Appendix A.

Anderson’s experimental protocol consists of fixed steps and an experimental design. The participants were able to express their crossing decision (dependent variable) on a 20-point rating scale ranging from 1 (definitely not cross) to 20 (definitely cross). The same 20-point scale was used in previous studies that implemented and validated Anderson’s experimental protocol, for example, Delmas et al. [45], Delmas et al. [46], Hurgobin et al. [47], Monsaingeon et al. [44]. The numbering on the scale was not displayed to prevent the effect of numbering bias on the responses which is in line with previous research [47,62]. The numbering was used only in the analysis of the responses.

3.2.3. Survey Procedure

The participant’s age and gender were asked to allocate a protagonist for Anderson’s experimental protocol questions. Following the series of 16 scenarios for Anderson’s experimental protocol, they were asked 20 questions from the PBQ and 44 questions from the BFI scales. At the end of the questionnaire, the participants were asked to give information on their age range, the highest level of education attained, past involvement in traffic collisions, the presence or absence of a driving licence, driving experience, and the minutes they spent walking daily. Additionally, they were asked to provide their email addresses if they wanted to be included in a lucky draw competition. The total time required to complete the questionnaire was up to 15 min. The list of questions for the sixteen scenarios in Anderson’s experimental protocol is listed in Appendix A.

4. Data Analysis

4.1. Anderson’s Experimental Protocol

The responses were imported into IBM SPSS v28.0.1.1 for repeated measures of Analysis of Variance (ANOVA). The main effects of the four factors (road gradient, weather condition, longitudinal distance between pedestrian and AV and type of vehicle) and, following that, the possible interactions between their sub-dimensions on the crossing decision were analysed. Bonferroni correction was used for post hoc tests.

Hierarchical clustering followed by K-means clustering was performed to categorise the participants into distinct clusters based on their Likert scale response to cross or not cross a road in various scenarios. The hierarchical clustering analysis generated a dendrogram, which indicated four types of possible clustering solutions. Two, three, four and five clustering options were obtained after interpreting the dendrogram. Following that, a K-means clustering analysis was performed on each option, and a decision was made that the three-clustering solution was the most distinct and significant option to categorise the collected sample. This procedure was suggested by Hofmans and Mullet [63] as it uses all data points and is less susceptible to outliers [45].

4.2. Pedestrian Behaviour Questionnaire and Big Five Inventory

The Likert-scale responses’ mean and Standard Deviation (SD) were calculated for the PBQ and BFI to find the most common PBQ category and BFI personality trait across the collected sample. These scores were also used later in the analysis to describe each cluster’s (generated from Anderson’s experimental protocol) road use behaviours and personality traits. The internal reliability values of our sample for the items in PBQ were positive behaviour (0.26), violations (0.59), lapses (0.80), errors (0.54), and aggression (0.81). The internal reliability values of our sample for the items in BFI were openness (0.52), conscientiousness (0.71), extraversion (0.70), agreeableness (0.74), and neuroticism (0.74).

4.3. Synthesis of the Protocol and the Scales

The data collected for PBQ and BFI were found to be non-normal; thus, non-parametric tests were performed. First, those responses were subjected to the Kruskal–Wallis test (non-parametric ANOVA) to examine if a difference exists between PBQ and BFI characteristics of the participants under different clusters. Then, Dunn’s test (a non-parametric post hoc test) was used to perform pairwise comparisons to investigate further and observe the level of difference existing between mean values of PBQ and BFI scores between a cluster and two other clusters.

5. Results

5.1. Demographics

Table 3. Demographics of the collected sample across the UK and Australia.

Age Group	United Kingdom (71 Females, 172 Males)		Australia (18 Females, 20 Males)
	Female	Male	Female	Male
18–40 (Total: 234)	53	152	15	14
Above 41 (Total: 47)	18	20	3	6

shows a breakdown of participants based on their age, gender, and location.

The number of responses for the two countries was imbalanced; thus, a cultural comparison was not performed. However, given that both are English-speaking countries with similar traffic rules and regulations, the combined data was included in further analysis. We also did not find significant differences in crossing intentions between genders.

5.2. Analysis Conducted on the Crossing Decisions

First, analysis was conducted on the whole sample to find out the influence of all the selected factors on the Likert scale responses for road crossing decisions. We conducted a 2 (road gradient: uphill vs. downhill) X 2 (weather: sunny vs. rainy) X 2 (distance between AV and pedestrian: 80 vs. 40 m) X 2 (type of AV: small car vs. lorry) ANOVA. It revealed that participants were willing to cross the road the most in the scenario where a small, automated car approaching from 80 m away was travelling up a hill towards them on a sunny day, on a scale ranging from 1 to 20 (Mean = 13.5, SD = 5.6).

The scenario with the exact opposite modalities, i.e., a lorry travelling down a hill towards the participant from 40 m away from them on a rainy day, was the scenario where most participants were less willing to cross (Mean = 6.0, SD = 5.7).

Second, the effects of all the factors on crossing decisions were found to be significant at p < 0.001 with varying effect sizes. A ${\mathsf{ŋ}}_{\mathbf{p}}^2$ value equal to or more than 0.14 signifies a large size effect, equal to or more than 0.06 signifies a medium size effect, and less than that signifies a small size effect [59].

Table 4. Main effects of factors on pedestrians’ road crossing decisions.

Factor	F-Value (p < 0.001)	${\mathsf{ŋ}}_{\mathbf{p}}^2$ Values	Effect Size
Road gradient F (1, 280)	98.051	0.259	Large
Weather F (1, 280)	51.544	0.155	Large
Distance F (1, 280)	148.068	0.346	Large
Vehicle type F (1, 280)	19.454	0.065	Medium

shows that road gradient, weather and distance had a large effect size, whereas the type of vehicle had a medium effect size on the crossing decision.

Thirdly, pairwise comparisons were conducted to analyse the relations between the sub-dimensions as per Anderson’s experimental protocol. It revealed three significant interactions between the sub-dimensions and their combinations as listed below.

We only list the significant interactions for the sake of brevity.

• Between road gradient and longitudinal distance, F (1, 280) = 11.172, p < 0.001, ${\mathsf{ŋ}}_{\mathit{p}}^2$ = 0.038. Pairwise comparisons showed that when the AV was travelling up the hill and the distance was 80 m, participants were more willing to cross the road. See

  Table 5. Interaction between road gradient and longitudinal distance between the AV and pedestrian on mean crossing decisions and standard errors.

  Road Gradient	Distance (m)	Mean	S.E.
  AV travelling up the hill	80	12.519	0.257
  	40	8.069	0.274
  AV travelling down the hill	80	10.346	0.251
  	40	6.795	0.264

  .
• Between the weather and longitudinal distance, F (1, 280) = 4.919, p = 0.027, ${\mathsf{ŋ}}_{\mathit{p}}^2$ = 0.017. Pairwise comparisons showed that when the weather was sunny and the longitudinal distance between the AV and the pedestrian was 80 m, participants were more willing to cross the road. See

  Table 6. Interaction between weather and longitudinal distance between the AV and the pedestrian on mean crossing decisions and standard errors.

  Weather	Distance (m)	Mean	S.E.
  Sunny	80	12.105	0.253
  	40	7.817	0.269
  Rainy	80	10.760	0.239
  	40	7.048	0.266

  .
• Between all four factors (1, 280) = 5.008, p = 0.026, ${\mathsf{ŋ}}_{\mathit{p}}^2$ = 0.018. It was found that participants preferred when the AV was a small car, travelling uphill on a sunny day and 80 m away from them while crossing a road. See

  Table 7. Interaction between road gradient, weather conditions and longitudinal distance between the AV and the pedestrian and the type of vehicle on the mean crossing decision and standard errors.

  Road Gradient	Weather	Distance	Type of Vehicle	Mean	S.E.
  AV coming up the hill	Sunny	80	Small car	13.473	0.333
  			Lorry	12.722	0.348
  		40	Small car	8.203	0.345
  			Lorry	8.577	0.352
  	Rainy	80	Small car	12.103	0.317
  			Lorry	11.776	0.332
  		40	Small car	8.342	0.353
  			Lorry	7.157	0.338
  AV coming down the hill	Sunny	80	Small car	11.498	0.331
  			Lorry	10.726	0.337
  		40	Small car	7.630	0.332
  			Lorry	6.858	0.337
  	Rainy	80	Small car	9.993	0.343
  			Lorry	9.167	0.359
  		40	Small car	6.665	0.341
  			Lorry	6.028	0.342

  .

5.3. Clustering Based on Crossing Decisions

Following the analysis of the whole sample, a clustering analysis was performed on the Likert scale responses for road crossing decisions.

A three-cluster solution was adopted to classify the participants into distinct clusters based on their varying decisions to either cross or not cross a road in the sixteen scenarios.

Table 8. Demographics of participants in each cluster.

Demographic	Clusters
	Cluster 1 (n = 52)	Cluster 2 (n = 27)	Cluster 3 (n = 202)
Male	30	21	141
Female	22	6	61

shows the demographics in each cluster.

For all the clusters and factors, uphill, sunny, 80 m and small car modalities have the highest mean values, indicating that these were the preferred factors. However, for participants in cluster 2, the lorry was found to have a minor preference over small cars. The table can be found in Appendix B.

We did not explore socio-demographic differences within the clusters for two reasons. First, we did not find significant effects of socio-demographic variables on crossing intentions in the overall sample. Second, splitting the clusters further would have resulted in highly uneven group sizes, especially due to the gender imbalance in clusters 2 and 3.

Figure 2 is a line drawing to visualise the mean values of crossing decisions on the Y-axis (on a scale of 1–20) of scenarios with the most and least ratings. Only the significant interactions among the factors are marked on the X-axis. The interactions marked on the X-axis under the coloured lines representing three different clusters show the significant interactions for each cluster, respectively.

The “positive crossing decisions” on the positive Y-axis refer to the Likert-scale rating for scenarios with the significantly highest likelihood of crossing. The “negative crossing decisions” on the negative Y-axis refer to the Likert-scale rating for scenarios with the significantly lowest likelihood of crossing.

For example, risk-averse pedestrians (yellow lines) were the most willing to cross the road in scenario-US8C with a mean response rating of 10.04 on a 20-point scale, whereas they were the least willing to cross the road in scenario-DR4L with a mean response rating of only 2.04 on the same scale.

The full configurations of the scenarios are provided in the figure caption. Additionally, for the same cluster of participants, only the four interactions: R*D, R*W*AV, D*AV, and W*D (the interactions listed in the X-axis bounded by the yellow line) significantly affected their road crossing decision. The other two clusters, represented by red and green lines, can be interpreted similarly.

On the X-axis, R means road gradient, W means weather, D means longitudinal distance between the AV and the protagonist, and AV means the type of vehicle. On the Y-axis, UR8C stands for Uphill, Rainy, 80 m and Small Car, US8C stands for Uphill, Sunny, 80 m and Car, DR4L stands for Downhill, Rainy, 40 m and Lorry, and DR4C stands for Downhill, Rainy, 40 m and Small Car.

The Standard Deviation (SD) values for crossing decisions in each cluster can be found in Appendix C.

5.3.1. Cluster 1: Risk-Averse Pedestrians

The first cluster (n = 52) was composed of participants who were, in general, the most reluctant to cross the road under any conditions.

Thus, this group of people were called risk-averse pedestrians. ANOVA revealed significant interactions (individual and between factors) in this cluster, as shown in

Table 9. Significant factors and interactions in Cluster 1.

Factor	F-Value (1, 51)	${\mathsf{ŋ}}_{\mathit{p}}^2$ Values	Effect Size
Road	36.048	0.414	Large
Weather	22.467	0.306
Distance	47.040	0.480
Vehicle type	20.974	0.291
Road * Distance	10.953	0.177
Weather * Distance	6.772	0.117	Medium
Distance * Vehicle type	5.131	0.091
Road * Weather * Vehicle type	10.016	0.164

The asterisk symbol represents the interaction between the factors.

.

5.3.2. Cluster 2: Resolute Pedestrians

Cluster two (n = 27) was found to be the most distinct among the three clusters because participants in this cluster expressed comparatively positive crossing decisions.

This cluster was thus named resolute pedestrians. There were no significant effects for any of the factors in this cluster at the individual level. However, there were two significant interactions between the factors, as shown in

Table 10. Significant interactions in Cluster 2.

Factor	F-Value (1, 26)	${\mathsf{ŋ}}_{\mathbf{p}}^2$ Values	Effect Size
Road * Distance	5.847	0.184	Large
Road * Weather * Vehicle type	6.311	0.195

The asterisk symbol represents the interaction between the factors.

.

5.3.3. Cluster 3: Indecisive Pedestrians

Cluster three (n = 202) was composed of participants with characteristics similar to those of risk-averse pedestrians, but the effect size of all four factors was lower for them. In contrast, the highest and lowest means of crossing decisions were higher compared to those of the first cluster. They were named indecisive pedestrians. This was the only cluster of participants for whom the interaction effect of all four factors was significant. ANOVA revealed significant interactions (individual and between factors) in this cluster, as shown in

Table 11. Significant factors and interactions in Cluster 3.

Factor	F-Value (1, 201)	${\mathsf{ŋ}}_{\mathit{p}}^2$	Effect Size
Road	32.213	0.232	Large
Weather	27.502	0.162
Distance	140.378	0.348
Vehicle type	23.062	0.053	Small
Road * Distance	18.416	0.013
Weather * Vehicle type	17.561	0.024
Road * Weather * Distance * Vehicle type	18.105	0.002

The asterisk symbol represents the interaction between the factors.

.

Figure 3 illustrates the effect sizes of only the significant interactions of each cluster.

The interaction between road gradient and distance (R*D) was significant in all three clusters. Cluster 1 (risk-averse pedestrians) was found to have larger effect sizes in general, and the opposite could be observed for cluster 3 (indecisive pedestrians). Cluster 3 was the only cluster whose crossing intentions were significantly influenced by the interaction between all four factors, i.e., road x weather conditions x distance between AV and pedestrian x type of AV. The clusters are defined in light of the interpretation of their self-reported measures rather than a moderation of situational effects in the following Section 6.

5.4. Analysis of PBQ and BFI Within the Clusters

The PBQ was measured on a scale of 1 (extremely infrequently or never) to 6 (extremely frequently or always).

The highest scored PBQ items in each category were positive behaviour: “I thank a driver who stops to let me cross” (Mean = 5.7), error: “I cross between vehicles stopped on the roadway in traffic jams” (Mean = 4.4), violation: “I cross diagonally to save time” (Mean = 4.5), aggressive behaviour: “I have gotten angry with a driver and hit their vehicle” (Mean = 5.9), lapse: “I forget to look before crossing because I am thinking about something else” (Mean = 5.4).

The participants were asked to choose one of the five options, ranging from 1 (disagree strongly) to 5 (agree strongly) while answering the BFI questions. It was found that agreeableness was the most common psychological trait (35.2%), whereas extraversion was the least common (9.3%) in the collected sample.

5.5. Synthesis of Road Crossing Decisions, PBQ and BFI

Kruskal–Wallis test revealed there were significant differences between some of the PBQ and BFI items’ scores of participants in each cluster.

The following PBQ items: violation, aggression, and lapses were found to be significantly different across risk-averse (C1), resolute (C2) and indecisive (C3) pedestrians.

The following BFI item: Agreeableness was found to be significantly different across risk-averse (C1) and resolute (C2) pedestrians.

Following that, Dunn’s test was performed to calculate the differences in mean scores for the PBQ and BFI items, as shown in

Table 12. Dunn’s test for the difference in mean rankings for PBQ and BFI items between the three clusters.

Categories	Clusters	Difference in Average Rank Scores Between Two Clusters	Std. Error	p-Value
Pedestrian Behaviour Questionnaire
Aggression	Risk-averse-indecisive	−41.600	12.543	<0.001
	Risk-averse-resolute	−56.029	19.133	0.003
Error	Risk-averse-indecisive	−47.433	12.596	<0.001
	Risk-averse-resolute	−42.312	19.215	0.028
Violation	Risk-averse-indecisive	−52.148	12.596	<0.001
	Risk-averse-resolute	−63.578	19.214	<0.001
Lapses	Risk-averse-indecisive	−34.307	12.489	0.006
	Risk-averse-resolute	−42.444	19.051	0.026
Big-Five Inventory
Agreeableness	Risk-averse-resolute	−41.410	19.244	0.031

.

Figure 4 shows radial diagrams to visualise the PBQ and BFI mean rank values obtained from Dunn’s test. It can be observed that risk-averse pedestrians (C1) ranked the highest on positive behaviour and agreeableness, and the least on violation and openness. Resolute pedestrians (C2) ranked the highest on aggression and openness, and the least on positive behaviour and neuroticism. Indecisive pedestrians (C3) ranked the highest on errors and extraversion, and the least on positive behaviour and agreeableness.

6. Discussion

In our study, Anderson’s experimental protocol was used to collect and analyse responses on pedestrians’ road-crossing decisions across 16 different scenarios. The decisions were then related to their declared behaviours while crossing a road, as well as their psychological facets. The scenarios were used to manipulate the factors: road gradient, weather, longitudinal distance between pedestrians and the AV and the type of vehicle.

The first objective was to investigate the potential effect of selected factors and their interactions on pedestrians’ road-crossing decisions while interacting with an AV. ANOVA conducted on the whole sample (N = 281) showed all four factors had either a large (road gradient, weather, and longitudinal distance between pedestrians) or medium (type of vehicle) effect size on pedestrians’ crossing decisions.

The large effect size of road gradient on crossing decisions was a novel and interesting find, bridging a gap in the road safety literature, as there were no existing studies investigating this topic to the best of our knowledge. In this way, hypothesis 1: “Based on the looming effect, we posit that pedestrians are more likely to cross the road when an AV is approaching them from down a hill compared to when it is approaching from up a hill” is consistent with our findings.

Weather conditions (rainy vs. fair, or sunny) were found to create inconclusive effects on crossing decisions that involved HDVs and AVs in the consulted literature. Some studies have concluded that pedestrians make riskier road crossing decisions in rainy conditions Movahhed et al. [32], whereas some studies show that pedestrians are more cautious and practice caution in the presence of unfair weather conditions [33]. According to our results, rain decreased pedestrians’ intentions to cross the road, whilst sunny weather encouraged road-crossing behaviour in the presence of AVs. Therefore, it can be assumed that pedestrians consider their safety to be more important than comfort when it comes to risks while crossing the road in the rain. This is in line with the risk homeostasis theory by Wilde [19], where individuals adapt their behaviour (and preferred risk levels) depending on a benefit versus risk trade-off. Thus, our findings support hypothesis 2: “Based on results analysing the frequencies and severities of pedestrian–AV crashes in different weather conditions, we posit that pedestrians are more likely to cross the road under sunny weather conditions compared to rainy weather conditions”.

Next, it was found that pedestrians preferred crossing the road when the AV was 80 m away from them compared to 40 m. This is an addition to numerous existing studies, which point out that pedestrians tend to prefer larger TTA values (which were expressed in terms of distance in this questionnaire) while crossing a road [36,40]. Therefore, our results support hypothesis 3: “Based on TTA values, we posit that pedestrians are more likely to cross the road when the AV is 80 m away from them compared to when the AV is 40 m away”.

Next, McGuire et al. [38] concluded that visual looming affects the perceived threat of the incoming vehicle. Visual looming was also found to influence collision detection, TTA and road crossing responses [31]. The findings from this study supported this, as pedestrians preferred to cross the road in front of a small, automated car compared to a large, automated lorry. Therefore, we have evidence to support hypothesis 4: “Based on looming bias, we posit that pedestrians are more likely to decide to cross the road when a small, automated car is coming towards them compared to an incoming automated lorry”.

These two findings relating to TTA and the type of vehicle are also congruent with the risk homeostasis theory by Wilde [19], as the results point out that the participants adapted their acceptable levels of risk by making a trade-off between levels of safety and perceived risk.

The second objective was to quantify the effect of each factor and the potential interactions between them on pedestrians’ crossing decisions. Uphill movement of a small, automated car on a sunny day, 80 m away, was preferred most by the participants (supporting the fifth hypothesis). Interaction analysis revealed three significant interactions:

• Road gradient X longitudinal distance between the participant and the AV.
• Weather Xlongitudinal distance between the participant and the AV
• and road gradient X weather X longitudinal distance between the participant and the AV X type of vehicle.

Variations in interactions between the factors were found to cause a variation in the risk perceived by the pedestrians, thus influencing their crossing decisions; thus, our findings also support the fifth hypothesis.

Next, clustering analysis on the road crossing decisions revealed three distinct clusters. The third objective was to investigate potential interactions between pedestrians’ road crossing decisions, their self-reported behaviours and psychological facets at road crossings. This investigation was carried out per cluster to compare the three distinct profiles. Although the three clusters varied in size, this did not compromise the statistical validity of the results. However, one limitation of the smaller cluster sizes is that they are more likely to lead to type II errors. Pedestrians’ declared behaviour was found to be an effective predictor of pedestrians’ psychological traits and vice versa. The three clusters are discussed further in the following sections.

6.1. Risk-Averse Pedestrians

In the first cluster (n = 52), i.e., risk-averse pedestrians, participants were reluctant to cross the road in most of the scenarios. They ranked the highest on positive behaviours and agreeableness, and the lowest on violation and openness.

Their mean crossing responses were the lowest among the clusters, but they had the most significant interactions between factors, suggesting they processed more information and took fewer risks when crossing. Their PBQ and BFI scores aligned with Zheng et al. [53] and Qu et al. [54], who found that positive behaviour predicts agreeableness. These individuals, being risk-averse, were less violent and exhibited safer behaviours, consistent with our findings. Openness was also a significant predictor of positive behaviour in Zheng et al. [53] and Qu et al. [54], and our results further support it. Openness was also a positive predictor of aggression [54] due to its strong link to sensation seeking, which can manifest as positive behaviour or aggression. In the second cluster, higher openness may correlate with more aggressive behaviours.

6.2. Resolute Pedestrians

In the second cluster (n = 27), i.e., resolute pedestrians, participants were characterised by high ranks for aggression and openness and low ranks for positive behaviour and neuroticism.

This group had the highest mean crossing responses across all sixteen scenarios, showing the highest willingness to cross. They also had the fewest significant factor interactions, indicating minimal consideration of surrounding stimuli before crossing. This aligns with Joshanloo [55], who links openness to risk-taking. However, our findings contrast with research suggesting that aggression correlates with neuroticism [64] and that openness predicts positive behaviour [53]. These inconsistencies may stem from contextual differences, as Zheng et al. [53] studied urban residents’ living and travelling patterns, and Dam et al. [64] examined personality traits and aggression, while our study focused on road-crossing decisions in the presence of AVs and stimuli.

6.3. Indecisive Pedestrians

In the third cluster (n = 202), i.e., indecisive pedestrians, participants were characterised by high ranks for error and extraversion and lower ranks for positive behaviour and agreeableness. Their mean crossing decision values were in-between the two other clusters but were uniquely influenced by the combination of all four factors, suggesting they processed stimuli but were indecisive. While [53] found that positive behaviour and agreeableness reduce errors, this cluster showed the opposite. Though extroverts are typically confident [25], our findings indicate indecision, likely due to contextual differences. Indecisive pedestrians might be hesitant to make decisions while interacting with novel AV technology, thus committing more errors. Therefore, future research should focus on bolstering the knowledge of pedestrians regarding AVs’ capabilities to mitigate the occurrence of human errors at road crossings. This is paramount as pedestrians may reproduce these errors they already make while interacting with non-automated vehicles.

6.4. Synthesis

The study distinguished participants’ behavioural intention, self-reported road-use behaviour, and psychological traits as predictors across the three clusters. Using Anderson’s experimental protocol and two scales, it explored pedestrian–AV interactions and factor dynamics. This is where we highlight the novelty of our study by using Anderson’s experimental protocol on factors that have inconclusive effects on pedestrian road crossing decisions and further establishing the links with road use/crossing behaviour and personality traits. Our findings highlight risky environmental and road features crucial for AV design and safety guidelines.

Evaluating pedestrian characteristics helps predict behaviours during AV interactions at crossings and informs future road safety campaigns. Pedestrian profiling enables targeted safety messaging and allows road authorities to tailor strategies to different profiles [65]. As suggested by the same research, focusing on risk profiles can improve safety for high-risk groups like resolute pedestrians or groups prone to making human errors, like indecisive pedestrians. Future research could explore additional factors in scenarios to assess the consistency of PBQ and BFI predictor findings with this study.

7. Limitations

A potential limitation is the sample size and representativeness, as convenience sampling led to uneven distribution across age groups and genders. However, in our sample, we did not find significant effects between socio-demographic variables and cluster properties based on Anderson’s protocol.

A literature review drove the rationale behind choosing the investigated factors. We did not include popularly used factors such as the speed of approaching AV because it has been researched to vast depths, and we also knew that it would be difficult for the participants to visualise a dynamic factor such as speed via a questionnaire.

We used PBQ and BFI, which are standardised and validated scales. The internal consistencies (α) in some dimensions of the scales were lower than the values reported in their original literature. This may be explained by differences between validation samples and our participants (e.g., demographic or cultural). For example, PBQ was validated using data exclusively from the French population [49], and BFI was checked for internal reliability using the American population [52]. The scales’ reliability has been established in prior work, and our findings align conceptually with existing literature, suggesting that the lower α reflects sample-specific variation but not a psychometric weakness. Another reason could be that the PBQ was designed for non-AV-pedestrian interactions and may not always be consistent for AV-pedestrian interactions.

This study serves as a foundation for experimental designs, such as pedestrian simulation studies that capture actual behaviour rather than intentions. Using pictures to illustrate scenarios was less immersive than videos, but practical for smartphone-based responses. The use of images might have under-rendered the looming effect and omitted the acoustic cues and presence of other road users. However, the use of Anderson’s protocol has been proven to be a valid, immersive and efficient tool in a wide range of studies, including automated driving [44,45,46,48], thus it is a solid method to understand pedestrian intentions.

8. Recommendations for Future Work

Future research should aim for a more balanced sample in terms of incorporating pedestrians with mobility issues (for example, wheelchair users) or who have had past experiences in traffic collisions. We expect differences in how those groups of people would interact with AVs compared to those without such experiences.

As pointed out in the previous section, PBQ and BFI are expected to vary across cultures. Cultural differences can be compared to get an in-depth understanding of how they influence crossing intentions in front of an AV, for example, conducting a France-UK-Australia comparison of PBQ’s internal consistency.

Road crossing is affected by a multitude of factors, not only the four factors we investigated in this questionnaire. Different sets of factors and their interactions can be examined in the future. Furthermore, they can adopt Virtual Reality (VR) or simulators to enhance immersion and study pedestrian behaviours beyond intentions.

9. Conclusions and Implications

This study highlights that pedestrians’ crossing decisions vary not only with stimulus changes but also with their interactions. The four stimuli examined—road gradient, weather, pedestrian–AV distance, and vehicle type—all influenced road crossing decisions. Our study was conducted to understand which scenarios are perceived to be riskier from a pedestrian’s perspective as an AV approaches them, thus aiding the safe integration of AVs into road networks.

A novel finding was that pedestrian road-crossing behaviour was influenced by whether the AV was travelling uphill or downhill, a factor not explored in existing literature. While previous research was inconclusive about weather effects, our results showed a preference for crossing in sunny conditions over rainy ones. Additionally, pedestrians were more likely to cross when a small car was 80 m away than when a lorry was 40 m away. All findings aligned with the initial hypotheses.

Pedestrians’ psychological traits can be used to effectively predict their behaviour around AVs. Risk-averse pedestrians exhibited more positive behaviours, higher agreeableness, and lower violations and openness. Resolute pedestrians were more aggressive, open, and less neurotic, with fewer positive behaviours. Behavioural intentions were linked to these traits, categorising participants into three profiles. Consistent with Hulse et al. [66], AV perception varies with individual risk-taking tendencies. Indecisive pedestrians made more errors, were extraverted, and scored lower on positive behaviour and agreeableness. These profiles can inform road safety initiatives targeting high-risk individuals and those most likely to make errors at road crossings. Future work could integrate insights from AV driving strategies (e.g., overtaking; [67]) with models of pedestrian behaviour, to better capture how both traffic flow and human behaviour shape AV performance.

In this way, the study was conducted to understand how pedestrians interact with AVs at road crossings and what factors affect these interactions. This study further validates the application of scenario-based methodology [44,45,46,48] in the context of AVs and expands its applicability to pedestrians’ road crossing decisions whilst interacting with an AV.

The takeaway message from this study and the implications on safe pedestrian–AV interactions are summarised as follows:

• The results from this study can be used to inform the design of AVs. For example, AVs could be designed to have a more cautious driving style in the conditions considered the riskiest by pedestrians, i.e., while travelling downhill in rainy weather.
• Pedestrians preferred to cross when there was the largest gap (i.e., 80 m, not 40 m) between them and the incoming AV. Thus, AVs should be designed to start decelerating as far away from the pedestrian as possible.
• The knowledge of decision-making in different scenarios can be used in pedestrian profiling and safety initiatives and policies. Three distinct profiles were identified using clustering analysis: risk-averse, resolute, and indecisive.
• Furthermore, participants’ crossing decisions were associated with self-reported behaviour (positive behaviour, violations, lapses, errors and aggressive behaviour) and psychological traits (openness, conscientiousness, extraversion, agreeableness and neuroticism). Thus, pedestrian profiling helps inform traffic simulation studies by furthering the understanding of the road environment and geometry on pedestrian behaviour when exposed to AVs. It also advises practitioners to develop measures to mitigate or avoid the occurrence of human errors from the pedestrian perspective.
• The results from our experiment can be used to design further studies that compare pedestrian interactions with either AVs or HDVs. Furthermore, the findings on pedestrian profiling can be used to develop studies that deal with trust and/or familiarity with AVs.