The assessment of hazard awareness skills among light rail drivers

Transportation Research Part F 67 (2019) 15–28

Contents lists available at ScienceDirect

Transportation Research Part F journal homepage: www.elsevier.com/locate/trf

The assessment of hazard awareness skills among light rail drivers Avinoam Borowsky a,⇑, Netta Palacci a, Moshe Itzhaki b, David Shinar a a b

Industrial Engineering and Management, Ben-Gurion University of the Negev, Beer-Sheva 8410501 Israel Jerusalem Transportation Master Plan, Jerusalem 91280, Israel

a r t i c l e

i n f o

Article history: Received 18 November 2018 Received in revised form 2 October 2019 Accepted 6 October 2019

Keywords: Light rail drivers Hazard awareness Light rail driving experience Eye movements

a b s t r a c t Light rail (LR) is a popular means of public transportation worldwide, in use in more than 380 cities worldwide. LR drivers must have good hazard awareness: the ability to understand the complexity of the traffic environment and anticipate road events. Yet, no study has examined LR drivers’ ability to anticipate hazards, and this is the purpose of this study. The experimental group included 28 certified LR drivers from the LR in Jerusalem. The control group included 26 licensed drivers, with no experience in LR driving. Participants observed 18 short video clips of typical LR driving that were filmed from the LR driver’s field of view and had to press a response button each time they identified a hazard. Participants’ eye movements and button presses were recorded throughout the experiment. In general, LR drivers were better at identifying hazards compared to the control group. Novice LR drivers with less than 1 year of LR driving experience or under training were more likely to respond to hidden hazards and responded much sooner compared to both the experienced LR drivers and Control drivers. The implications are discussed. Ó 2019 Elsevier Ltd. All rights reserved.

1. Introduction Light rail (LR) is one of the most popular means of public transportation worldwide, currently operated in more than 380 cities all around the world, transferring approximately 13.6 billion passengers daily (Dauby, 2015). LR is a system of electrically powered rail vehicles operating on tracks where passengers board from stations or from track side stops along the street (Neff & Dickens, 2017). Most driving is done by using an electronic throttle control that enables the LR driver to control the acceleration and breaking of the train. Another useful tool is the horn which is used whenever the LR driver needs to notify or alert other road users in the area of the train (Neff & Dickens, 2017). A typical LR cabin’s windshield provides the driver with a visual field of view of about 120 degrees. Although LR driving shares clear similarities with heavy train driving, because they are both limited to rails and require similar throttle manipulation, there are many differences as well. Light rail, for example, is lighter and slower than the train, has more stops and, most importantly, operates in a mixed-traffic environment. These differences impose entirely different demands on the LR driver in order to drive safely. LR driving is complex because it needs to be performed in a dynamic and complex environment and often in congestion and under high levels of mental workload. In terms of the environment, the LR tracks system shares its roads with other road users and can be generally described by three different types of setups: (1) the tracks are separated from the road, (2) the tracks are segregated but still cross through shared intersections, and (3) the roads are entirely shared and the tracks run on ⇑ Corresponding author. E-mail address: [email protected] (A. Borowsky). https://doi.org/10.1016/j.trf.2019.10.003 1369-8478/Ó 2019 Elsevier Ltd. All rights reserved.

16

A. Borowsky et al. / Transportation Research Part F 67 (2019) 15–28

the road in mixed traffic without dedicated lanes (Naweed & Rose, 2015). According to Naweed and Rose (2015) the most risky areas with high potential for a crash include intersections that are used by all road users or areas where the track runs along the road without any segregation. These areas increase the potential for conflicts with other road users that can often be unpredictable. While the LR driver needs to safely navigate through this complex traffic environment and anticipate road hazards there are additional challenges imposed on the driver. One of these challenges is to maintain a high level of customer service by following a strict time schedule and provide a pleasant trip to its passengers (e.g., minimizing emergency breakings; Naweed & Rose, 2015). This argument is supported by an Australian study that used qualitative techniques to investigate what are the primary goals of 10 experienced LR drivers and found that these drivers noted two major goals: ensuring a smooth journey for the passengers and providing a good customer service within the time constrains (Naweed & Moody, 2015). Similarly, Nazning, Currie and Logan (2017) conducted a study where 30 LR drivers (with a range of driving experience between 1.17 and 31 years) participated in a focus group and described seven major challengers for safe driving: (1) safety for all people in and around the light rail including passengers, pedestrians, cyclists and motorists; (2) On-time running pressure - a primary criterion for the evaluation of driving performance and increases the risk taking attitudes among drivers; (3) keeping high level of concentration; (4) Falls of passengers inside the train mostly due to sudden emergency brakes; (5) Anticipating road users’ behaviour; (6) Operational constrains such as long braking distance; and (7) Fatigue and workload. As noted by Naweed, Rose, Singh, and Kook (2017) one of the greatest challenges of LR drivers is to be able to anticipate the behaviour of other road users on the road and around the tracks. According to the authors, good LR driving not only requires the awareness of the various elements in the environment but also the ability to understand the environment’s complexity and be prepared for unexpected events. This ability to anticipate events that have the potential of causing harm to other road users, passengers and the driver has been studied widely in the domain of automobile driving (e.g., Chapman & Underwood, 1998; Horswill & McKenna, 2004; Sagberg & Bjørnskau, 2006; Borowsky, Shinar, & Oron-Gilad, 2010; Borowsky & Oron-Gilad, 2013). In the traffic safety domain, this ability is termed hazard perception, hazard anticipation, or hazard awareness. Hazard perception has many definitions but in general it can be defined as drivers’ ability to ‘‘read” the road and anticipate hazardous situations (Horswill & McKenna, 2004). Horswill, Hill, and Wetton (2015) provide convincing empirical evidence that of the many driving skills that a driver possesses only hazard perception has been found to correlate consistently with traffic crashes. For example, they found that drivers who failed in the Queensland’s official hazard perception test were 25% more likely to be involved in a crash in the preceding year as well as in the year following the test compared to drivers who passed the test successfully. Such findings, among others, have been used to include a hazard perception test as an integral part of the official licensing procedure in the UK since 2002 (Crundall, 2016). While hazard perception tests have been widely studied with respect to car drivers and are being used in some countries as part of the graduate driving licensure system, there seems to be no indication for using this kind of measure to evaluate LR drivers’ performance. One of the most consistent findings with respect to hazard perception is that experienced drivers possess better hazard perception skills compared to young-inexperienced drivers (e.g., Chapman & Underwood, 1998; Horswill & McKenna, 2004; Sagberg & Bjørnskau, 2006; Borowsky et al., 2010; Borowsky & Oron-Gilad, 2013). This superiority is typically reflected by faster response times to hazards (Horswill & McKenna, 2004; although see Borowsky & Oron-Gilad, 2013, for further discussion on this measure), and adopt different scanning patterns for different types of roads compared to novice drivers who apply the same scanning strategies regardless of the type of the road (Chapman & Underwood, 1998). In addition, experienced drivers are also better than young-inexperienced drivers at anticipating hidden hazards (Borowsky et al., 2010; Crundall et al., 2012; Vlakveld et al., 2011), e.g., situations where the hazard instigator is obscured behind a static object in the environment (e.g., a pedestrian who is obscured behind vegetation) or another road user (e.g., a situation where a pedestrian is obscured behind a truck that is stopped behind a midblock crosswalk). To summarize, although one of the main challenges of an LR driver is to be able to anticipate hazardous events there is no study, to the best of our knowledge, in the LR domain that explicitly evaluated hazard perception skills of LR drivers with their unique requirements. The purpose of this study was therefore to generate a valid hazard perception test for LR drivers in order to evaluate their hazard perception performance and possibly define a gold standard for adequate performance that can be later be used for training and screening LR drivers. 2. Material and methods 2.1. Participants Fifty-four participants, 36 males and 18 females, took part in this study. Of those, twenty-eight participants were learners or qualified LR drivers from the Jerusalem LR Transport (JLRT) who were recruited on a voluntary basis via Connect, which is the company that operates the LR in Jerusalem. Of the twenty-eight LR drivers, 3 were learner drivers who were not allowed to drive solo yet, 1 was novice with less than six months of solo LR driving and the rest were experienced LR drivers with more than one year of LR driving experience (of these, one was a female driver). Thus, LR learner or qualified drivers were divided into two groups: (1) twenty-four experienced LR drivers, and (2) novice and learner drivers (labelled novice LR drivers from now on). The third group of participants included twenty-six students from the Industrial Engineering and Management (IEM) department at Ben-Gurion University of the Negev (BGU), who served as the control group. Their age

A. Borowsky et al. / Transportation Research Part F 67 (2019) 15–28

17

range was 25–28 years-old with an average car-driving experience of 7.4 years. Experienced LR drivers age range was 26–62 years-old (Mean age = 41.88, SD = 8.08) and Novice LR drivers age range was 30–54 years (Mean age = 44, SD = 10). The age range of the control group was between 25 and 28 years (mean age = 25.5, SD = 1.10). This group had no prior knowledge or experience in LR driving. All participants had uncorrected Snellen visual acuity of 6/9 (20/30) or better, and normal contrast sensitivity based on the FACT contrast sensitivity test (Ginsburg, 1984). In addition, all participants filled a demographic questionnaire before the experiment began. The experiment was approved by the IRB of BGU. 2.2. Apparatus Apparatus. 2.3. Eye tracking laboratory The experiment was conducted in the eye tracking laboratory at BGU. A 2000 LCD wide screen with 1360 * 768 pixels (width = 41 cm, height = 25.8 cm), connected to a Pentium 4 PC, was used to display the movies. Participants sat at an average distance of 65 cm from the LCD, which provided them with an average visual field of 22 degrees vertically and 35 degrees horizontally. Another PC, behind the participant, was used by the experimenter to operate the eye tracking software interface and to control the participant’s computer. There was a partition between the experimenter area and the participant to ensure that the participant could not see the experimenter (the experimenter was able to see the participant’s face via a video camera). Finally, an external data cable was used to synchronize the stimuli (movie frame number and button presses) running on the participant’s computer with the eye tracking sampling on the experimenter’s computer. Participants’ eye movements were recorded with an Eye Tracking System (ETS; Applied System Laboratories, Model D6), sampling the visual gaze at 60 Hz, with a nominal accuracy of 0.5 degrees of visual angle. The D6 facial recognition algorithm allows head free eye tracking without putting any equipment on the participant. 2.3.1. Hazard perception movies of real-world LR driving The preparation of the movies for the experiment included several steps. (1) Camera installation. A Panasonic Lumix G DMC-G7 camera with an Olympus 7–14 mm wide lens was used to record the video footage. The camera was mounted in the middle on the internal side of the driver’s cabin’s windshield directed toward the rails. This setting allowed capturing the external environment from a driver’s perspective. (2) Video recordings. Recordings of the drive from both directions of the route were taken at different days and hours with different drivers. A single recording took about 50 min and covered the whole route along one travel direction. The Jerusalem light rail is currently working on a single urban route 13.8 km long. In total, all video recordings included 6 h of continuous driving. (3) Movies editing. The movies were edited into dozens of relatively short movies (between 24 and 42 s) with a rate of 50 fps. A representative sample of these movies was presented to a group of experts (LR drivers’ qualified trainers) at the JLRT in order to get a grasp of the typical hazardous situations that LR drivers might experience when driving along the route at different days and hours. Based on the experts’ comments and ideas, eighteen short representative driving movies were selected for the experiment. In order to avoid unwanted background sounds (such as honking) effects on participants behaviour, the original soundtrack of each movie was removed and replaced with the same sound of a moving train but without other noises (like horn of the LR or of other vehicles). Of these 18 movies, two movies were used for practice purposes only and were not included in the analysis. The other sixteen movies were used for testing. Of these, thirteen were recorded during day-time and three during night time. Each movie contained a different number of hazardous events that are defined next. 2.3.2. Definition and classification of hazards Typical hazards that appeared in the movies included situations such as pedestrians walking parallel to and near the tracks, pedestrians and bicycles who were crossing the tracks, a passing train in the opposite tracks, other vehicles crossing at crossroads, etc. Examples for typical hazards in daytime and night time can be seen in Fig. 1.

Fig. 1. Two examples of typical hazards in night-time (left – a pedestrian standing on the left side of the track) and daytime (right – an intersection with a truck on the left potentially obscuring other traffic).

18

A. Borowsky et al. / Transportation Research Part F 67 (2019) 15–28

Based on the general theme of the experts’ comments, 42 hazards were predefined by the research team and in cases of disagreements they were settled through discussion. After running the experiment and analysing the response data, eight additional hazards were defined post hoc; ultimately yielding 50 hazardous situations across all sixteen movies. A new hazard was added post hoc to the hazards database only if at least 30% of all participants responded to that hazard. The length of a hazardous situation could range from 1.64 to 14.36 s (mean = 5.55 s, SD = 2.59 s). The classification of the 50 hazardous situations follows the classification procedure previously used by Borowsky and Oron-Gilad (2013). One of the main classification factors is Hazard type, which is defined according to a combination of two possible dimensions of the hazard. The first dimension refers to the visibility of the hazard instigator. This dimension defines whether the hazard instigator can be seen (visible) or is obscured by the environment or other road users (hidden). For example, a hidden hazard might include a crossing pedestrian who is obscured by a passing train in front of the LR driver and a visible hazard might include a bicyclist who is riding near the tracks and is visible to the LR driver throughout the hazardous situation. The second dimension of a hazard is the hazard state (materialized vs. unmaterialized hazard), which refers to whether the hazard instigator is on a collision course with the LR driver (materialized hazard), a situation that requires an evasive manoeuvre in order to prevent a crash (e.g., brake), or whether the hazard instigator has not yet entered the driver’s path (unmaterialized hazard), a situation that only requires monitoring of the hazard instigator without taking any evasive response. Three types of hazards emerge from the combination of these two dimensions: (1) Visible-Potential (VP) hazard. This type of hazard includes situations where the hazard instigator is visible, but the hazard instigator had not entered the driver’s path. This type of hazard requires the driver to monitor the hazard, but it does not require any evasive action. A potential hazard can become a materialized hazard at any given moment. For example, a pedestrian walking in parallel next to the tracks who starts to cross the tracks. (2) Hidden-Potential (HP). This type of hazard includes situations where the hazard instigator is hidden behind other road users or behind environmental objects and therefore it cannot be seen by the LR driver. In this type of hazard, the hazard instigator does not enter the driver’s path. This type of hazard requires the driver to monitor the area from where a hazard instigator might appear in order to detect the hazard when and if it becomes visible. For example, a passing train in the opposing tracks that can obscure pedestrians who are standing behind it. (3) VisibleMaterialized (VM) hazard. This type of hazard is similar to visible potential hazards that materializes. For instance, a pedestrian that darts into the tracks very close to the front of the train. Although theoretically the two dimensions, hazard type and hazard state can yield four options, the fourth option, that is, a hidden-materialized hazard is not feasible. In fact it is a hidden-potential hazard that turned into a visible-materialized hazard. In total, there were 30 visible-potential hazards, 10 hidden-potential hazards and 10 visible-materialized hazards. The second classification factor was the source of the hazard or the type of the hazard instigator. Hazard sources were pedestrians, bicyclists, other vehicles, other light rails or road constructions workers. 2.3.3. Hazard perception test and software Eighteen driving scene movies were displayed randomly to the participant on a 2000 LCD and he or she was asked to press a designated response button whenever they identified a hazardous situation. The participant’s response had no effect on the movie and it continued playing until the movie ended. Each button press was marked by a ding sound. At the end of each movie the participant indicated the reason for each of his or her responses and typed it into a designated text-box on the screen. The number of text boxes presented on the screen matched the number of button presses in each movie. A fixation screen (a grey screen with a small black circle in the middle) was presented for 500 ms before the beginning of the next movie. The purpose of this screen was to make sure that all participants started scanning each movie from the centre of the screen. An in-house software was used to display the movies and to record all button presses and their associated hazard descriptions as well as the eye movements’ data. 2.3.4. Questionnaires In addition to a demographics questionnaire, the multidimensional driving style inventory (MDSI, Ben-Ari et al., 2003) was administered to the participants at the end of the hazard perception testing. This questionnaire evaluates drivers’ driving styles and is based on a self-report scale assessing eight different driving styles based on which a single dominant style is determined for each driver. 2.4. Experimental design and variables The experiment had a 2 * 3 * 3 mixed design and included a practice session consisting of two movies and a HPT session consisting of 16 movies. Each participant observed the movies in a random order. The between-subjects independent variables were drivers’ group: experienced LR drivers, novice LR drivers, or Control (students from the IEM department at BGU) and Gender. The within-subjects independent variable was hazard type, which included 3 levels: visible-potential, hiddenpotential, and visible-materialized (as explained earlier in Section 2.3.2). The dependent variables that are detailed in the results section are based on two different types of data: (1) button presses, and (2) eye movements. While the behavioural response data (i.e., button presses) did not require much preparation for the analysis eye movements data did require several steps before they could be used.

A. Borowsky et al. / Transportation Research Part F 67 (2019) 15–28

19

2.5. Procedure The participant arrived at the lab and received a short explanation about the purpose of the study and its importance, after which he signed a consent form agreeing to participate in the experiment. Then, he was given a visual acuity test and a contrast sensitivity test. Afterwards, the participant was asked to sit in front of the display and to fill a demographics questionnaire. Once finished, he went through a short eye calibration process followed by a detailed instructions page presented on the screen explaining to the participant that he or she is about to observe a sequence of LR driving movies that were taken from a LR driver’s perspective and that his task is to press a response button as quickly as possible whenever he identifies a hazard. The instruction page also contained the following hazard definition: ‘‘any object, situation, occurrence or combination of these that introduce the possibility of the individual road user experiencing harm” (Haworth, Symmons and Kowadlo, 2000). Next, two movies were presented as practice to allow the participant to get familiarized with the experimental task. At the end of the practice, the participant could ask for further clarifications and explanations if necessary. Then, the participant was asked to start the HPT. As described above, in this phase 16 driving scene movies were displayed to the participant and he or she was asked to press a response button whenever he (she) identified a hazard. At the end of each movie the participant indicated the reasons for each button press. When this phase ended the participant was asked to fill the MDSI questionnaire and the experiment ended. 3. Analysis and results 3.1. Data preparation For each participant two types of data files were generated: (1) an eye movements’ data file containing raw data of gaze position at a rate of 60 Hz, and (2) a TXT file including all button presses and their exact time stamp along each movie as well as verbal descriptions of each hazard. Each type of raw data required several preparation steps before it could be analysed. The information on these preparation steps are described in the following sub-sections. 3.1.1. Fixations and regions of interest (ROIs) First, with respect to the raw eye movements’ data file, these data were used for extracting fixations for each participant during each movie. Fixation extraction was done using the dispersion methodology applied by Gitelman (2002) in ILAB. The dispersion algorithm has three parameters: minimum fixation duration (milliseconds), minimum dispersion considered a fixation, and maximum consecutive sample loss. These parameters were set to 100 ms, 1 visual degree, and infinity (default), respectively. Fixation position was set as the mean position of all samples considered in a single fixation. Before the extraction of fixations’ data into a summary table, a validation phase was made in order to ensure data validity. This validation phase compared the eye movements’ raw data file with the video of the superimposed eye movements recorded in the experiment. Whenever the experimenter identified problems with the eye movements’ recordings (e.g., a large amount of missing data, or non-calibrated data), that specific trial was omitted from the analysis. This procedure resulted in poor eye movements’ data of 13 participants from the control group and of 8 participants from the test group. These participants were removed from the eye movements’ analysis altogether. Therefore, the eye-movements data analysis is based on 13 participants from the control group and 20 participants from the test group. Second, in order to determine whether a participant perceived the hazard it was necessary to define an area that surrounds the hazard and observe the number of fixations that fell inside this area and their duration during the hazardous situation. This area is termed region of interest (ROI). A ROI was defined as a rectangle surrounding the target including an additional 15 pixels on each side to account for the 0.5 degrees of visual angle nominal error reported by the eye tracker manufacturer. For some of the hazards there were two ROIs depending on the number of elements that were considered as a hazard. The definition process of the ROIs included the following steps: (1) first, we have identified the start and end time of the hazard situation in terms of frames (and seconds). (2) Then, each time window of a given hazard was segmented into several short segments depending on the total duration of the hazard. The purpose of this segmentation was to adjust the size and position of the ROIs to the context of the dynamic situation during the whole-time window of the hazard. (3) Next, for each segment a representative screenshot was extracted from the middle of the segment and on that screenshot a rectangle was drawn around the hazard. Then, the X and Y values of two points of the rectangle creating the ROI were extracted and recorded. The X’s and Y’s ROI data of all hazards were used later for the analysis of the eye-movement data. To provide an example for how a ROI was defined consider Fig. 2. In this figure there is a passing train on the opposing tracks. The opposing train obscures potential pedestrians that might be hidden behind it. Therefore, in this type of situation we defined a ROI (red rectangle) behind the train in the area where we expect LRDs to monitor for potential pedestrians that might cross the tracks. 3.1.2. Behavioural response With respect to the behavioural response, button presses and verbal descriptions of the hazards, were consolidated into a preliminary summary table. For each button press in the summary table, the experimenter examined whether it matched a predefined hazard or not. A button press was classified as a response to a predefined hazard only if both the verbal descrip-

20

A. Borowsky et al. / Transportation Research Part F 67 (2019) 15–28

Fig. 2. An example of a ROI of a typical hidden potential hazard.

tion of the hazard and the temporal location of the button press matched the attributes of that hazard. All responses that were not classified as predefined hazards were also examined. These responses that were not associated with predefined hazards yielded eight newly defined hazards (according to the criterion in Section 2.3.2). Finally, for each button press the response time was also computed by subtracting the beginning of each hazard from the moment when the button press was initiated. The final hazards that were included in the analyses and their characteristics can be found in Appendix A. 3.2. Results Data collection and preparation included two types of information. The first refers to participants’ button presses and the second to their fixations data. The results section refers to both types of information as well as to the MSDI questionnaire data and is organized as follows. The first sub-section describes the analyses with respect to differences in drivers’ abilities to detect different types of hazards. The second sub-section describes the analyses with respect to the time it took drivers to identify the various types of hazards and to respond to them. All analyses were done with an alpha of 0.05. Whenever post hoc pairwise multiple comparisons were done the alpha was corrected with the sequential Bonferroni procedure. The final section describes the driving styles profiles of the LR drivers according to the MSDI questionnaire analysis. 3.3. Hazard detection (probability to identify a hazard) The first analysis in this section refers to the probability that a participant will press the response button during the hazard’s allotted time window. For this analysis, the dependent variable is binary distributed indicating whether a participant pressed the response button within the allotted time window of the hazard (‘‘1”) or not (‘‘0”). Three independent variables were included in the analysis. The first independent variable was group (control, novice LR drivers, and experienced LR drivers), the second was the type of hazard (visible-materialized, visible-potential, hidden-potential) and the third was gender. A binary logistic regression model was utilized within the framework of General Linear Mixed Models (GLMM), with a logit link function. The independent variables and their second-order interactions were included as fixed effects and participants were included as a random effect. Applying a backwards elimination procedure, the final model revealed two significant main effects for group and hazard type and a significant interaction between group and hazard type (see Table 1). Table 1 A summary of the fixed effects of the final hazard detection model. Source

F

DF1

DF2

Sig

Estimated means (SE) of the probability to respond to a hazard

Odds ratio

Group

39.78

2

2690

<0.01

Hazard type

49.01

2

2690

<0.01

NLR/C = 6.31 ELR/C = 2.24 VP/VM = 0.32HP/VM = 0.13

Gender Group * HazardType

5.90 14.1

1 4

2690 2690

=0.015 <0.01

C = 0.3 (0.02), NLR = 0.73 (0.04), ELR = 0.49 (0.03) NLR > ELR > C (Padj = 0.01; Padj = 0.03 respectively) VP = 0.49 (0.02), HP = 0.28 (0.03), VM = 0.75 (0.03) VM > VP > HP (Padj < 0.01; Padj < 0.01 respectively) Males = 0.47(0.02), Females = 0.549(0.03) See Fig. 2

F/M = 1.37

Note. C, LRN and LRE stands for control, novice LR drivers, and experienced LR drivers. VP, HP and VM stands for visual-potential, hidden-potential and visual-materialized hazards respectively.

A. Borowsky et al. / Transportation Research Part F 67 (2019) 15–28

21

As shown in Table 1 (row 1) the estimated probability that novice LR drivers will respond to a given hazard was significantly larger than that of both the experienced LR drivers and the control drivers. In addition, experienced LR drivers were significantly more likely to respond compared to the control drivers. Furthermore, participants were more likely to respond to visible materialized hazards compared to all other types of hazards (row 2). There was also a gender effect such that females tended to respond more often compared to males (row 3). Next, Fig. 3 describes the interaction between group and hazard type. Post hoc pairwise comparisons analysis revealed several patterns. First, novice LR drivers and experienced LR drivers were significantly more likely to respond to visible-potential hazards compared to the control group (NLR > C, Padj < 0.01; ELR > C, Padj < 0.01). Novice LR drivers were also more likely to respond to visible-potential hazards compared to experienced LR drivers (Padj < 0.01). Novice LR drivers were significantly more likely to respond to hidden-potential hazards compared to experienced LR drivers (Padj < 0.01) and compared to the control group (Padj < 0.01). In addition, experienced LR drivers were significantly more likely than the control to respond to hidden-potential hazards (Padj = 0.01). The visible-materialized type of hazard did not yield significant differences between the groups. Second, while novice LR drivers had a similar response probability for all types of hazards, both the control group and experienced LR drivers were significantly more likely to respond to visible-materialized hazards, followed by visible-potential hazards followed by hidden-potential hazards (ELR: VM > VP, Padj < 0.01, VP > HP, Padj < 0.01; Control: VM > VP, Padj < 0.01, VP > HP, Padj < 0.01). The second analysis in this section refers to the probability that a participant will have at least one fixation on the hazard during its allotted time window. This analysis was aimed at examining whether different groups of drivers detected different types of hazards, information that might partially explain the differences in their behavioural responses. For this analysis, the dependent variable is binary distributed indicating whether a participant had at least one fixation inside an ROI of a given hazard (‘‘1”) or not (‘‘0”). For this analysis, the independent variables group, gender and type of hazard were included. Since the dependent variable is binary distributed we utilized a binary logistic regression model within the framework of GLMM, with a logit link function. The independent variables and their second-order interactions were included as fixed effects and participants were included as a random effect. Notably, in order to make sure that only participants with valid eye data during the hazard are included we have calculated the average number of fixations of all participants during a hazard window and removed from the analysis participants who had a number of fixations that was smaller than 2 standard deviations from the mean. This procedure resulted in a total of 1026 valid cases (out of the possible 1050 that were considered originally as valid). We used the same procedure for all eye data analyses. Applying a backwards elimination procedure, the final model revealed a marginally significant main effect for gender and a significant effect for hazard type (see Table 2). As shown in Table 2 (row 1) the estimated probability that females will have at least one fixation on a hazard was significantly larger than that of male drivers. In addition, the estimated probability of having at least one fixation on the hazard was significantly smaller for hidden potential hazards than for both visible-materialized and visible-potential hazards. Finally, the estimated probability to have at least one fixation on the hazard was not significantly different between the groups of drivers (C = 0.594(0.06), NLR = 0.39(0.13), ELR = 0.49 (0.12)). 3.4. Normalized response time This variable measures the time interval between the beginning of the hazardous situation and the time when the button press was initiated divided by the duration of the time window of that hazard. For this analysis, three independent variables were included into the model as well as their second-interactions. The independent variables were similar to those in the previous analysis. Since the dependent variable is normalized, it is not normally distributed and it is bounded between 0

Fig. 3. Interaction between group and hazard type. Note. Error bars represent standard error.

22

A. Borowsky et al. / Transportation Research Part F 67 (2019) 15–28

Table 2 A summary of the fixed effects of the final model of having at least one fixation on a hazard. Source

F

DF1

DF2

Sig

Estimated means (SE) of the probability to have at least one fixation on the hazard

Odds ratio

Gender Hazard Type

3.79 9.3

1 2

1020 1020

=0.052 <0.01

Males = 0.431(0.09), Females = 0.548(0.1) VP = 0.56 (0.09), HP = 0.37 (0.09), VM = 0.54 (0.1) HP < VM = VP (Padj < 0.01; Padj < 0.01 respectively)

F/M = 1.6 VP/VM = 1.08 HP/VM = 0.5

and 1. Thus, we applied a logit transformation on it. We then utilized a linear regression model within the framework of GLMM. The independent variables and their second-interactions were included as fixed effects and participants were included as a random effect. Applying a backwards elimination procedure, the final model revealed three significant main effects for group, gender and hazard type and a significant interaction between group and hazard type (See Table 3). The values of the estimated means were transformed back to the original values of the dependent variable (that is, before the logit transformation, i.e., normalized response time). Because anti-log of the transformed dependent variable does not produce the means of the original dependent variable these values represent the medians of the original dependent variable. This argument is valid because for a normal distribution the mean is equal to the median. As shown in Table 3 (row 1), Experienced LR drivers as well as the Control drivers were slower to respond to a given hazard compared to the novice LR drivers. No differences were found between the control group and the Experienced LR drivers. In addition, response time for hidden hazards was much faster than for visible-potential hazards (row 2). Next, Fig. 4 describes the statistically significant interaction between group and hazard type. Post hoc pairwise comparisons analysis revealed several patterns. First, experienced LR drivers were significantly slower to respond to hidden-potential hazards compared to novice LR drivers (Padj < 0.01) and compared to the control group (Padj < 0.01). In addition, novice LR drivers were significantly faster to respond to hidden-potential hazards compared to the Control (Padj = 0.016). No differences between the groups were found for materialized hazards. Second, while experienced LR drivers displayed significantly slower response time for hidden hazards compared to visible hazards (HP > VP: Padj < 0.01; HP > VM: Padj < 0.01), novice LR drivers displayed an opposite pattern, that is, a faster response time towards hidden hazards compared to both visible-potential hazards and visible-materialized hazards (HP < VP: Padj < 0.01; HP < VM: Padj < 0.01). Third, both experienced LR drivers and Novice LR drivers were faster to respond to visible-potential hazards compared to the control (Padj < 0.01 for both comparisons). Forth, females identified the hazards significantly faster compared to males (row 3). Finally, Fig. 5 depicts the estimated median response time of each group to a typical hidden hazard ‘‘limited field of view due to a train in the opposing direction” (6 out of 9 hazards that were classified as hidden potential hazards). Post hoc analysis of the interaction between group and hazard type revealed several patterns. First, all participants identified visible materialized hazards at about the same time. Second, for visible-potential hazards the control drivers were slower to respond compared to novice and experienced LR drivers (Padj < 0.01 for both comparisons) who were not different from one another. Third, for hidden-potential hazards, experienced LR drivers were slower to respond compared to both the control drivers and the novice LR drivers (Padj < 0.01 for both comparisons). The control drivers were slower to respond compared to the novice LR drivers (Padj < 0.01). The second analysis in this section refers to the first fixation on the hazard from the moment the hazard first appeared in the movie. The dependent variable was defined as the normalized time interval until first fixation, that is the time interval between the appearance of the hazard and the first fixation on the hazard divided by the duration of the hazard. This measure can be seen as response time in terms of eye movements. Since the dependent variable is normalized, it is not normally distributed and it is bounded between 0 and 1. Thus, we applied a logit transformation on it. We then utilized a linear regression model within the framework of GLMM. The independent variables and their interaction were included as fixed effects and participants were included as a random effect. Applying a backwards elimination procedure, the final model revealed a significant main effect of hazard type (F(2,796) = 14.41,P < 0.01). All other effects were not statistically significant. VM hazards were identified later (Estimated NRT = 0.2, CI (0.13–0.27)) compared to both VP (0.09, 0.06–0.11) and HP hazards

Table 3 A summary of the fixed effects of the final normalized response time model. Source

F

DF1

DF2

Sig

Estimated medians (lower-upper limits of the confidence interval) of the normalized RTs

Group

13.55

2

1118

<0.01

Hazard Type

5.06

2

1118

<0.01

Gender Group*HazardType

13.86 16.07

1 4

1118 1118

<0.01 <0.01

C = 0.47 (0.37–0.58), NLR = 0.2 (0.13–0.31), ELR = 0.57 (0.47–0.66) ELR > NLR; C > NLR (Padj < 0.01), C = ELR VP = 0.52 (0.45–0.59), HP = 0.29 (0.19–0.42), VM = 0.42 (0.33–0.52) VP > HP (Padj < 0.01); VM = VP, VM = HP Males = 0.52(0.45–0.59), Females = 0.3(0.22–0.4) See Fig. 3

Note. C, NLR and ELR stands for control, novice LR drivers, and experienced LR drivers. VP, HP and VM stands for visual-potential, hidden-potential and visual-materialized hazards respectively.

23

A. Borowsky et al. / Transportation Research Part F 67 (2019) 15–28

Fig. 4. Estimated NRT medians of the interaction between group and hazard type.

Fig. 5. An example of average RT for each group for a typical hidden potential hazard. The left (0.12 s from the hazard first appearance), middle (0.52 s) and right (4.3 s) images represent the average RT of 2 novice LR drivers, 1 control, and 6 experienced LR drivers respectively.

(Padj < 0.01). VP hazards were identified significantly later compared to HP hazards (0.04, 0.03–0.07; Padj < 0.01 for both comparisons). All drivers were similarly likely to respond after 10% of the hazard’s time window on average. To summarize the results, Table 4 presents the main findings focusing on the group of drivers and the types of hazards. 3.5. MSDI questionnaire analysis. The questionnaire analysis follows the method of van Huysduynen, Terken, Martens, and Eggen (2015). The questionnaire is consisted of 44 questions divided into 8 categories. Each question is consisted of a score between 1 and 6 on a Likert scale. The way the participants’ scores were analysed are as follows: for each category the average score of all participants was calculated. Then to this average 1 standard deviation was added. This value represents a threshold such that a participant

Table 4 A summary of the study’s main findings. Performance measure

Data type

Group of drivers

Hazard type

Gender

Group*hazard type

Hazard identification

Button press

NLR > ELR > C

VM > VP > HP

F>M

Eye movements

N.S.

HP < VM, VP

F > M Marginally significant

VM: ELR = NLR = C VP: NLR > ELR > C HP: NLR > ELR > C NLR: VM = VP = HP ELR,C: VM > VP > HP N.S.

Button press

ELR,C > NLR

VP > HP; VM = VP, HP

M>F

Eye movements

N.S.

VM > VP > HP

N.S.

Response time

VM: ELR = NLR = C VP: C > NLR, ELR HP: ELR > C > NLR N.S.

Note. C, NLR and ELR stands for control, novice LR drivers, and experienced LR drivers. VP, HP and VM stands for visual-potential, hidden-potential and visual-materialized hazards respectively. F and M stands for females and males respectively. N.S. stands for non-significant.

24

A. Borowsky et al. / Transportation Research Part F 67 (2019) 15–28

Fig. 6. LR drivers’ MSDI profiles.

which his score was higher than the threshold received 1 and 0 otherwise. Fig. 6 presents the driving style profiles of all LR drivers. It can be seen that the majority of drivers considered themselves as patient and careful. 4. Discussion This study was aimed at examining LRDs ability to anticipate road hazards as this is one of the greatest challenges of LRDs (Naweed et al., 2017). While hazard perception has been widely examined in the traffic safety domain there is no study that examined these abilities among LRDs. Thus, for the purpose of this study several traffic scene movies were taken from an LRD’s perspective on a designated LR route in Jerusalem in order to generate a hazard perception test for LRDs. Official LRDs at different levels of LR driving experience (experienced and novices) as well as students who had no experience in LR driving were invited to the eye tracking lab and were asked to observe sixteen movies from an LRD perspective and to press a response button each time they identify a road hazard. Participants’ eye movements were recorded throughout the experiment. The results of this study revealed several patterns. Because the types of hazards according which the hazards were classified (visible-materialized, visible-potential, and hidden potential) were a prominent factor in affecting participants’ responses and eye movements the discussion will be organized accordingly. First, with respect to materialized hazards, there were no significant differences between experienced LR, novice LR and control drivers either in terms of the probability to respond or in the time to respond. In addition, this type of hazard was identified more often than both visible-potential and hidden potential hazards and did not yield any differences in terms of eye movements. This pattern is consistent with previous studies showing that even novice drivers have no problems in detecting materialized hazards and respond to them at the same speed as experienced drivers (Vlakveld et al., 2011; Borowsky et al., 2010). This pattern is not surprising because responding to a materialized hazard where the hazard instigator is visible (e.g., a pedestrian crossing the tracks in front of the LR) is independent of driving experience and do not require any anticipation capabilities. In general, these types of hazards typically induce the largest number of responses among all types of hazards and our results show the same pattern as well (See Table 4). Second, with respect to visible-potential hazards, two patterns were found. First, experienced LRD (0.56) and novice LRD (0.69) were significantly more likely to respond to a visible-potential hazard compared to the control (0.24). Novice LR drivers were significantly more likely to respond to this type of hazard compared to experienced LR drivers. This pattern is very interesting because it shows that this type of hazard is unique to the LR driving context and LR drivers who are operating in this unique environment are more sensitive to this type of hazard compared to the control. That is, the control group were students who had no experience in LR driving and thus were not sensitive for visible potential hazards. In other words, an observer who is not familiar with the context of the driving environment cannot really judge whether a pedestrian who is walking along the tracks or a construction worker who stands beside the tracks will enter the LR path of travel. Although

A. Borowsky et al. / Transportation Research Part F 67 (2019) 15–28

25

novice LR drivers tended to respond more often to this type of hazard compared to experienced LR drivers we believe this is due to their over-sensitivity to respond to any type of hazard, which is a result of their extensive training. This argument will be elaborated later. Notably, there was no difference between the groups in terms of their likelihood to have at least one fixation on the hazard. This eye movements’ pattern suggest that the issue for not pressing the button was not because drivers did not fixate on it but rather they either did not interpret it correctly or did not think it was hazardous. Response time analysis showed that not only did the control group identify fewer hazards of this type but they were also much slower to respond to these hazards compared to experienced LRD and novice LRD who were not different from one another. Again, the control group late response was not because they fixated on the hazard later but rather because it took them more time to decide whether a visible-potential hazard is indeed hazardous. This tendency may be attributed to a term known as response bias (e.g., Wallis & Horswill, 2007; Crundall, 2016) Third, and perhaps the most interesting findings with respect to the differences between experienced LRDs and novice LRDs, was demonstrated for hidden-potential hazards. Of the nine hazards that were classified as hidden potential, six (67%) included situations where a LR in the opposing track was approaching the driver. The other three scenarios included situations of a vehicle that might obscure other traffic who might cross the tracks. It was expected that drivers will indicate these situations as hazardous because the tail of the train in the opposing lane obscures possible pedestrians that might dart into the tracks. It was also hypothesized, as in previous studies in traffic safety (e.g., Borowsky & Oron-Gilad, 2013) that experienced LR drivers will identify these types of hazards better than novice LRDs. Our findings, however, revealed a different pattern. In fact, we found an opposite pattern. Novice LRDs were more likely than any other group to respond to hidden-potential hazards (0.66). Experienced LRDs (0.22) were significantly more likely than the control (0.09) to respond to hidden-potential hazards. In addition, the RT results revealed that experienced LRDs were significantly slower (Median RT = 0.85) to respond to hidden hazards compared to both the control (0.22) and the novice LR drivers (0.04). Appendix B provides an example of drivers’ fixations dispersion during a 2 s window of a typical hidden hazard – an approaching train in the opposing lane. This 2 s window is considered as a critical point where the train in the opposing lane is about to pass and potential pedestrians may be obscured behind it. Before resolving this apparent contradiction, it should be noted that due to the relatively low number of novice drivers this is only an assumption that should be further investigated. That said, note that of the four novice LRDs who participated in the study, 3 were learner drivers. This means that 3 LRDs were still during their training and one driver had finished his training less than a year prior to the experiment. According to LRD instructors, during their training, drivers are taught what typical hazards exist when driving along the track especially when a train is approaching the tracks in the opposing lane. Learner drivers are usually taught to use their horn whenever a LR is passing in the opposing lane in order to indicate possible pedestrians to be aware of their upcoming LR. One reasonable explanation to this apparent contradiction is that unlike typical studies that investigate differences between novice and experienced drivers, in this study novice LR drivers were learners who receive professional training. This means that these drivers are relying on declarative knowledge rather than on driving experience and that they are much more sensitive to respond to hazards that were emphasized during their training. This pattern of over-sensitivity to hazards after training was also reported by Meir, Borowsky, and Oron-Gilad (2014). Consistent with our argument, looking at the response pattern of novice LR drivers (Table 4 row 2) it can be seen that novice LR drivers were similarly likely to respond at high rates to all type of hazards and this pattern was different from the other two groups. Looking at Fig. 5 (left panel), one can see that learner LR drivers responded to this type of hazard from a far distance compared to both the control and experienced LRDs. Pressing the button when the LR in the opposing track is so far away may reflect their reliance on declarative knowledge rather than understanding the situation. In fact, when we discussed this pattern with LRDs instructors they all indicated that this was too far away from the train that if they would have used their horn in that moment it is most likely that pedestrians standing behind the LR in the opposing track would not have heard it. Experienced LRDs on the other hand (Fig. 5 right panel) did not respond to the hazard until the last moment, when using their horn would have been effective. These drivers relied on their driving experience to monitor the situation until they collected enough evidence that required a response that would also be effective. Finally, this study has also found gender differences in terms of the likelihood of responding to a hazard and in response time. Nevertheless, since the focus of this study was LR drivers and since vast majority of females were from the control group it would be inappropriate to reach any meaningful conclusions regarding gender differences with respect to LR driving. To summarize, this study was aimed to design a hazard perception test for LR drivers. The hazards that were used in the HPT were generated on the basis LR driving experts’ recommendations. This study is an important step in establishing a HPT for LRDs that will hopefully be used by LR trainers as well as other stakeholders. Since the number of LRDs was limited, especially with respect to novice LRD, future studies should further investigate how LR driving experience is affecting LR driving hazard perception abilities.

Acknowledgement We would like to thank the Paul Ivanier Center for Robotics and Production and Management and to the Jerusalem Transportation Master Plan management for their support.

26

A. Borowsky et al. / Transportation Research Part F 67 (2019) 15–28

Appendix A. A summary of the hazards in each movie classified according to hazard type and source

Movie Hazard # Hazard

Hazard source Hazard Type

Hazard start time

Hazard end time

Hazard Duration (sec)

1

pedestrian road construction road construction road construction

VP VP

0.02 20.68

2.7 27.32

2.68 6.64

VP

27.24

31.6

4.36

VP

31.96

34.92

2.96

1 2

closeness to tracks road construction

3

road construction

4

road construction

2

1 2 3 4 5

closeness to tracks closeness to tracks tracks crossing closeness to tracks tracks crossing

pedestrian pedestrian pedestrian pedestrian pedestrian

VP VP VM VP VM

3.8 12 14.82 25.2 25.8

9.2 14.8 19.4 29.08 29.4

5.4 2.8 4.58 3.88 3.6

3

1 2 3

limited field of view train closeness to tracks pedestrian closeness to tracks pedestrian

HP VP VP

2.56 22.6 32.6

8.86 34.2 40.2

6.3 11.6 7.6

4

1 2

hazard on tracks Vehicles at intersection

bicycle vehicle

VP VP

5.5 6.4

9.72 11

4.22 4.6

5

1 2 3 4

limited field of view hazard on tracks intersection closeness to tracks

train bicycle vehicle pedestrian

HP VP VP VP

0.84 5.12 14.48 26.16

5.56 8.32 21 30.4

4.72 3.2 6.52 4.24

6

1 2

limited field of view vehicle limited field of view vehicle

HP HP

1.4 15.06

6 23.8

4.6 8.74

7

1 2

HP VP

7.68 7.68

15.2 15.28

7.52 7.6

3

limited field of view train vehicles at vehicle intersection closeness to tracks pedestrian

VP

15.3

20.44

5.14

1

road construction

VP

1.92

16.28

14.36

VP

16.88

27.2

10.32

HP VM

27.2 34.2

33.96 38.2

6.76 4

8

2 3 4

road construction road construction road construction limited field of view train tracks crossing pedestrian

9

1

vehicles at intersection

vehicle

HP

8.6

12.9

4.3

10

1 2

tracks crossing closeness to tracks

bicycle pedestrian

VM VP

13.4 13.4

22 20.3

8.6 6.9

11

1 2 3 4 5

limited field of view tracks crossing closeness to tracks closeness to tracks tracks crossing

train pedestrian pedestrian pedestrian pedestrian

HP VM VP VP VM

7.48 11.94 11.94 17.24 21.6

11.66 16.8 16.8 23.2 28.08

4.18 4.86 4.86 5.96 6.48

12

1 2 3 4

closeness to tracks closeness to tracks limited field of view tracks crossing

vehicle pedestrian train pedestrian

VP VP HP VM

2.08 2.08 18.76 23.8

4.3 5.76 23.6 28.6

2.22 3.68 4.84 4.8

27

A. Borowsky et al. / Transportation Research Part F 67 (2019) 15–28 Appendix A (continued)

Movie Hazard # Hazard

Hazard source Hazard Type

Hazard start time

Hazard end time

Hazard Duration (sec)

13

1 2

intersection hazard on tracks

vehicle bicycle

HP VP

11.92 20

19.4 26.2

7.48 6.2

14

1 2

closeness to tracks busy station platform

pedestrian pedestrian

VP VP

14.6 16.84

18 28.8

3.4 11.96

15

1 2 3 4

tracks crossing closeness to tracks closeness to tracks tracks crossing

pedestrian pedestrian pedestrian bicycle

VM VP VP VM

0.6 5 9.56 14.86

3.36 9.6 14.8 16.5

2.76 4.6 5.24 1.64

16

1 2 3

closeness to tracks tracks crossing closeness to tracks

pedestrian pedestrian vehicle

VP VP VM

10.32 16 17.5

17.44 20.88 19.5

7.12 4.88 2

Note: VP, VM and HP stand for visible-potential, visible-materialized, and hidden potential hazards. Appendix B. . An example for fixations dispersion for all drivers during a critical phase of a hidden hazard – A train approaching from the opposing direction.

The green circles represent the fixations of 16 experienced LR drivers. The red squares represent the fixations of 9 control drivers and the blue diamond represent the fixations of one novice LR driver. All fixations were taken from the same time window of 2 s right before the critical moment when the light rail on the opposing lane ends and possible pedestrians may appear behind it. It can be noticed that the centre of the green cluster is much more to the left compared to the red cluster. This example suggest that experienced LR drivers were much more attuned to the end of the opposing rail compared to the control group. Nevertheless, this is only an example that should be further explored. References Borowsky, A., & Oron-Gilad, T. (2013). Exploring the effects of driving experience on hazard awareness and risk perception via real-time hazard identification, hazard classification, and rating tasks. Accident Analysis & Prevention, 59, 548–565. Borowsky, A., Shinar, D., & Oron-Gilad, T. (2010). Age, skill, and hazard perception in driving. Accident Analysis & Prevention, 42(4), 1240–1249. Chapman, P. R., & Underwood, G. (1998). Visual search of driving situations: Danger and experience. Perception, 27(8), 951–964. Crundall, D. (2016). Hazard prediction discriminates between novice and experienced drivers. Accident Analysis & Prevention, 86, 47–58.

28

A. Borowsky et al. / Transportation Research Part F 67 (2019) 15–28

Crundall, D., Chapman, P., Trawley, S., Collins, L., Van Loon, E., Andrews, B., & Underwood, G. (2012). Some hazards are more attractive than others: Drivers of varying experience respond differently to different types of hazard. Accident Analysis Prevention, 45, 600–609. Dauby, L. (2015). Light rail in figures: Statistics brief report. International Association of Public Transport. Ginsburg, A. P. (1984). Visual form perception based on biological filtering. Sensory experience, adaptation, and perception: Festschrift for Ivo Kohler, 53–72. Gitelman, D. R. (2002). ILAB: A program for post experimental eye movement analysis. Behavior Research Methods, Instruments & Computers, 34(4), 605–612. Haworth, N., Symmons, M., & Kowadlo, N. (2000). Hazard perception by inexperienced motorcyclists (Report 179). Melbourne: Monash University Accident Research Centre. Horswill, M. S., Hill, A., & Wetton, M. (2015). Can a video-based hazard perception test used for driver licensing predict crash involvement? Accident Analysis & Prevention, 82, 213–219. Horswill, M. S., & McKenna, F. P. (2004). Drivers’ hazard perception ability: Situation awareness on the road (pp. 155–175). A cognitive approach to situation awareness: Theory and application. Meir, A., Borowsky, A., & Oron-Gilad, T. (2014). Formation and evaluation of act and anticipate hazard perception training (AAHPT) intervention for young novice drivers. Traffic Injury Prevention, 15(2), 172–180. Naweed, A., & Moody, H. (2015). A streetcar undesired: Investigating ergonomics and human factors issues in the driver–cab interface of australian trams. Urban Rail Transit, 1(3), 149–158. Naweed, A., & Rose, J. (2015). ‘‘It’s a frightful scenario”: A study of tram collisions on a mixed-traffic environment in an Australian metropolitan setting. Procedia Manufacturing, 3, 2706–2713. Naweed, A., Rose, J., Singh, S., & Kook, D. (2017). Risk factors for driver distraction and inattention in tram drivers. In Advances in Human Aspects of Transportation (pp. 257–268). Springer, Cham. Naznin, F., Currie, G., & Logan, D. (2017). Key challenges in tram/streetcar driving from the tram driver’s perspective – A qualitative study. Transportation Research Part F: Traffic Psychology and Behaviour, 49, 39–48. Neff, J., & Dickens, M. (2017). 2016 Public Transportation Fact Book (67 Ed.). American Public Transportation Association. Sagberg, F., & Bjørnskau, T. (2006). Hazard perception and driving experience among novice drivers. Accident Analysis & Prevention, 38(2), 407–414. Taubman-Ben-Ari, O., Mikulincer, M., & Gillath, O. (2004). The multidimensional driving style inventory—scale construct and validation. Accident Analysis Prevention, 36(3), 323–332. van Huysduynen, H. H., Terken, J., Martens, J. B., & Eggen, B. (2015). September). Measuring driving styles: A validation of the multidimensional driving style inventory. In In Proceedings of the 7th International Conference on Automotive User Interfaces and Interactive Vehicular Applications (pp. 257–264). ACM.. Vlakveld, W., Romoser, M., Mehranian, H., Diete, F., Pollatsek, A., & Fisher, D. (2011). Do crashes and near crashes in simulator-based training enhance novice drivers’ visual search for latent hazards? Transportation Research Record: Journal of the Transportation Research Board, 2265, 153–160. Wallis, T. S., & Horswill, M. S. (2007). Using fuzzy signal detection theory to determine why experienced and trained drivers respond faster than novices in a hazard perception test. Accident Analysis & Prevention, 39(6), 1177–1185.