Assessing Contextual Factors That Influence Acceptance of Pedestrian Alerts by a Night Vision System

Abstract

Objective:

We investigated five contextual variables that we hypothesized would influence driver acceptance of alerts to pedestrians issued by a night vision active safety system to inform the specification of the system’s alerting strategies.

Background:

Driver acceptance of automotive active safety systems is a key factor to promote their use and implies a need to assess factors influencing driver acceptance.

Method:

In a field operational test, 10 drivers drove instrumented vehicles equipped with a preproduction night vision system with pedestrian detection software. In a follow-up experiment, the 10 drivers and 25 additional volunteers without experience with the system watched 57 clips with pedestrian encounters gathered during the field operational test. They rated the acceptance of an alert to each pedestrian encounter.

Results:

Levels of rating concordance were significant between drivers who experienced the encounters and participants who did not. Two contextual variables, pedestrian location and motion, were found to influence ratings. Alerts were more accepted when pedestrians were close to or moving toward the vehicle’s path.

Conclusion:

The study demonstrates the utility of using subjective driver acceptance ratings to inform the design of active safety systems and to leverage expensive field operational test data within the confines of the laboratory.

Application:

The design of alerting strategies for active safety systems needs to heed the driver’s contextual sensitivity to issued alerts.

Keywords

alert acceptance active safety systems knowledge elicitation driver behavior automation

Introduction

Active safety systems promise significant gains in traffic safety but can have a beneficial impact if and only if drivers use the systems. Driver acceptance of the alerts they issue will therefore be a major development target. In this article, we investigate the influence of context on drivers’ willingness to accept alerts issued by a pedestrian detection active safety system with the aim of informing the development of alerting strategies.

Driver Acceptance of Alerts

A prerequisite for a good pedestrian warning system is reliable detection of pedestrians in a wide variety of traffic contexts (Himanen & Kumala, 1988; Schmidt & Färber, 2009). However, meeting this need is not likely to be sufficient to ensure driver acceptance (Farber & Paley, 1993). Indeed, Farber and Paley (1993) proposed that the ideal warning system would activate in “alarming” situations even if the driver were able to avoid the accident without the alert. Such alerts would give drivers a better appreciation of the function of the system and would tune their expectations for them (Parasuraman & Riley, 1997). Although driver acceptance of alerts is likely tightly coupled to the perception of risk in the traffic context, risk perception may not be associated with the actual risk of collision (Sheridan, 2008). These considerations lead us to suggest that for a pedestrian warning system to become widely accepted, it must implement alerting strategies that take into consideration drivers’ attitudes to the situations in which alerts are being issued.

The perception of risk is hard to measure, but we can measure alert acceptance relatively easily (Källhammer, Smith, Karlsson, & Hollnagel, 2007). We therefore use alert acceptance as a proxy for the perception of risk and to analyze the influence of contextual factors on driver acceptance of active safety systems.

The Research Agenda

In line with Marshall, Lee, and Austria (2007), who found that the rated appropriateness of alerts depends on the driving context, the purpose of this study was to identify factors that influence driver acceptance of pedestrian alerts.

Our selection of factors to consider was based in part on the work of Himanen and Kumala (1988), who developed a model of the interaction between drivers and pedestrians in a variety of traffic contexts. Their purpose was to find factors that influence pedestrian safety. They included factors in their model that describe the locale and road environment, vehicle speed and approach direction, and pedestrian demographics. The pedestrian’s distance from the curb was a central factor in their model. Another factor we investigated was the direction of the pedestrian’s motion relative to the vehicle. Schmidt and Färber (2009) found that pedestrians facing an approaching car are likely to cross the road at greater distances and longer time gaps.

The model by Himanen and Kumala (1988), the observations by Schmidt and Färber (2009), and the fact that we did not have knowledge of pedestrian demographics led us to focus this investigation on five factors: (a) the locale (urban, suburban, or rural), (b) pedestrian location, (c) direction of pedestrian motion, (d) the vehicle’s path, and (e) the curvature of the road. Our aim was to identify factors that influence driver acceptance of alerts and that, by extension, should influence the alerting criteria used by the active safety system.

Our hypotheses are that each of the these contextual factors influences drivers’ acceptance of alerts to pedestrians. All other factors being equal, we expect alerts to pedestrians in rural locales to receive higher ratings than alerts to pedestrians in urban or suburban settings. We expect ratings to be higher when pedestrians are seen to be within or moving toward the driver’s field of safe travel (Gibson & Crooks,1938). We also expect ratings to alerts to be higher when a pedestrian is encountered while the driver is making a turn or is on a winding road than while driving straight down a straight road.

The study has two parts: a field operational test (FOT) that gathered a set of 57 video clips of pedestrian alerts and a tabletop laboratory exercise in which volunteers rated the acceptance of those alerts.

Field Study

Lerner, Dekker, Steinberg, and Huey (1996) investigated drivers’ annoyance at randomly issued alerts in naturalistic driving conditions. They hypothesized that the frequency with which drivers accept alerts is likely to depend on traffic conditions. They proposed a conceptual field study to test that hypothesis. One of the many aims of the FOT reported in this article was to implement such a field study.

As part of the FOT, we collected drivers’ responses to alerts generated by a prototype pedestrian warning system. We restrict our discussion of the FOT to the drivers’ real-time interaction with the system and a two-button response unit. Drivers used the response unit to indicate whether they accepted the issued alert. We used automatically recorded video clips of the events, with both accepted and rejected alerts, as stimuli in the subsequent experiment.

Participants

For this FOT, 10 male drivers (age, M = 49.2 years, SD = 6.8, range = 40 to 59) participated. All had considerable driving experience (M = 30.9 years, SD = 6.8, range = 22 to 41), had corrected-to-normal vision, and reported driving at least 25,000 km annually (M = 34,800 km, SD = 5,996, range = 25,000 to 40,000). The drivers were recruited at the Autoliv research facilities in Vårgårda, Sweden, and applied voluntarily to the study. Participation conformed to the ethical guidelines established by Vetenskapsrådet (2002), the Swedish Research Council.

Task

Volunteers drove their own instrumented vehicles every day and without restrictions for a period of approximately 2 months. The drivers were familiar with the vehicles and were free not to use the system if they chose. Frequency of use was not collected for privacy reasons. We did not prime the drivers in any way or define criteria for situations that might constitute acceptable alerts. No structured questionnaire was used to collect driver feedback.

While driving, the drivers could press a button to mark a perceived miss, that is, a situation that justified an alert when none was issued.

The Instrumented Vehicles

A prototype long-wave, or far infrared (FIR), night vision system was installed in recent-model-year Volvo S80, Volvo V70, and Saab 9-5 vehicles. The system consists of a camera mounted in the grille of the vehicle and a video display mounted on the upper part of the center console (Figure 1). The system contains integrated pedestrian recognition software that takes a vehicle-centric view: The alerting criteria are based solely on the location and motion of the pedestrians relative to the vehicle and do not consider the pedestrian’s location and motion relative to the street or road. The display screen was updated at 30 Hz with a black-and-white FIR image. The image was augmented by a flashing yellow alert symbol and red rectangle(s) that highlighted the pedestrian(s) whom the system had detected. A snapshot of a pedestrian alert is shown in Figure 2.

Figure 1.

Display at the top and the button response unit at the base of the center stack of a Saab 9-5.

Figure 2.

A typical alert issued by the system.

A PC was mounted in the trunk. A GPS captured the vehicle’s location and provided timing information. The PC recorded the video clips in a time window before and after an alert or a button push and its time stamp to identify the exact location of the pedestrian event in the recorded FIR video stream. No video in the visual spectrum was recorded because of difficulties associated with blooming light sources.

Video Clips

The systems flagged a total of 88 video clips with pedestrian encounters. After the FOT, the alerts were reviewed and a set of 57 video clips selected to be the stimuli in the subsequent tabletop laboratory experiment. Sequences with multiple pedestrians at different locations were eliminated to avoid ambiguity regarding which pedestrian had triggered the alert. Bicyclists were also excluded, as not enough cases were recorded to allow any separate statistical analysis on them. All remaining video clips were used. Each clip shows approximately 30 s of images from the FIR camera (like that shown in Figure 2) roughly 20 s before and 10 s after the recorded alert. The 30-s length is a compromise designed to provide sufficient context for the alert while limiting the time of the experiment.

Experiment

Participants

Two groups of participants took part in the experiment. The first was the group of 10 drivers from the field study. The second was a group of 25 volunteers (age, M = 43.5 years, SD = 10.4, range = 30 to 66) recruited from the same facility as the drivers. Most of the 25 had considerable driving experience (M = 24.2 years, range = 10 to 46). There was a wide spread of reported current annual mileage (M = 19,200 km, range = 2,000 to 50,000 km). None of the 25 had experience with the pedestrian alert system in his or her personal vehicle. Again, participation conformed to the ethical guidelines established by Vetenskapsrådet (2002), the Swedish Research Council.

Tabletop Equipment and Procedure

After reporting individually to the laboratory, the participants were told that the purpose of the study was to study factors that influence driver acceptance of issued alerts. The instructions covered the experimental procedure and equipment and the voluntary basis of participation and provided the opportunity to ask questions. Participants were informed that they would be asked to watch and rate a collection of infrared video clips, that the collection of clips included both issued alerts and situations perceived as not requiring an alert, and that they could quit or interrupt the experiment at any time for any reason at all. They were not provided with any feedback about the ratings of the video clips.

The laboratory setup consisted of a PC laptop connected to a video projector that presented the set of video clips on the wall at a distance of approximately 3 m and a horizontal field of view of approximately 40°. Immediately following the presentation of each clip, the projector screen showed the frozen last frame of the video clip, and the PC presented the response screen, which contained a scale bar and two buttons labeled Repeat and Next. No information about the collected traffic context other than the FIR video clips was provided to the participants.

The experiment was self-paced. As it was the participant who started each video clip, the experiment afforded opportunities to pause and exit the room. Each of the 35 participants rated all 57 clips. The clips were presented in random order. None of the participants used the repeat option. To avoid response bias, we did not query them on their thoughts regarding their criteria for acceptance.

Rating Method

Our approach to assess driver acceptance of alerts builds on the hazard perception test used in U.K. driving tests (Jackson, Chapman, & Crundell, 2009). We presented raters with a set of video clips of pedestrian events that they may or may not judge to warrant an alert. The video clips were shown without the alert icon. The flashing alert symbol was suppressed to avoid any indication about whether (and when) the event was triggered by the system or was flagged by the driver as a miss.

Following van der Laan, Heino, and De Waard (1997), we quantified the relative level at which raters indicated they would accept an alert to the events in the video clips. To achieve a single measure of driver acceptance, as in the U.K. hazard perception test, we used a single scale anchored by completely reject and completely accept to condense the two components usefulness and satisfying used by van der Laan et al. By using a single measure, we sought to avoid any confound posed by individual differences in drivers’ interpretation of the different components of the van der Laan et al. metric. The scale was presented as a slider bar to obtain a continuous, but not necessarily interval, measure. The participant used the mouse to indicate the relative level of acceptance of an alert.

Analysis

We cannot know whether the raters used the scale as an ordinal or an interval scale. We provided only two anchor points, no interval marks on the slider bar, and no instructions or other means to define a set of equal intervals between the anchors. Our intent was to treat the ratings as ordinal data and to follow the procedure prescribed by Siegel and Castellan (1988) for the analysis of ordinal data. This procedure converts the raw ratings into ranks within raters, assesses the concordance in those ranks across raters, and uses the Kruskal-Wallis ANOVA for ranks to assess the statistical significance across categories of events. As there were 57 clips in total, the ranks varied in the range of 1 to 57, with 1 being the rank for the clip with the least accepted alert and 57 the rank for the clip with the most accepted alert. The motivation for the scale bar and other aspects of the ratings procedure is discussed by Källhammer et al. (2007).

On the other hand, meaningful interpretations can be made with the use of parametric tests, such as ANOVA, even when the scale may not fulfill the requirements of an interval scale. As parametric analysis may facilitate broader understanding of the results, we analyzed the data using both parametric ANOVA on the raw ratings, which assumes the raters treated the slider bar as an interval scale, and the Kruskal-Wallis ANOVA on the rank of the rating, which treats the raw ratings as ordinal data.

Results

Concordance in alert acceptance

Internal consistency across raters lends credence to the aggregation of the raw ratings or their rankings that is required for the analysis of whether context matters. We tested the internal consistency of the ratings provided by the 35 tabletop raters (the10 FOT drivers and the 25 who did not participate in the FOT) with both parametric and nonparametric tests. The parametric test was correlation. We found the mean rating for each event and found the correlations for each rater with those means. The median correlation was +.75. The nonparametric test was Kendall’s coefficient of concordance of the ranks of the raw ratings. This test of interjudge reliability assesses the degree of agreement in the rank ordering of a set of items (e.g., the 57 video clips) by n judges (Siegel & Castellan, 1988). It imposes no categorical dimensions of similarity on rated items. After correcting for the numerous ties in the intrajudge ranks, we found them highly consistent, W = 0.55, χ2(56) = 1247, p < .0001. On average, the raters, whether they had experience with the events or not, differentiated between them in a similar way. Both measures of internal consistency support using the mean values of either the ratings or the ranks in the analysis of contextual variables.

Contextual variables

We investigated the influence of the five categories of contextual information listed in Table 1. The video clips were classified into three locales: rural, suburban and urban. We differentiated urban (n = 17) from suburban (n = 22) by requiring continuous blocks of buildings. The presence of streetlights differentiated suburban from rural areas (n = 18).

Table 1:

Summary of Statistical Tests of the Influence of Contextual Variables on Driver Acceptance of Alerts

		Kruskal-Wallis ANOVA by Ranks			ANOVA
Variable	Categories	df	KW	ϵ²	df	F	ES
Locale	Urban, suburban, rural	2	4.25	.08	2, 54	2.00	.26
Pedestrian location	In street, left, right edge, right side	3	27.06***	.48	3, 53	15.73***	.64
Pedestrian motion	Same, opposite, into street, standing	3	14.11***	.25	3, 53	5.08**	.62
Vehicle direction	Straight, turning	1	0.00	.00	1, 55	0.00	.00
Road direction	Straight, turning	1	1.72	.03	1, 55	1.71	.15

Note. Thirty-five participants rated the acceptance of alerts to 57 traffic events on a continuous scale (0 = reject completely, 100 = accept completely). Nonparametric ANOVA was calculated for the mean rank of the rating in each category and parametric ANOVA for the mean rating in each category.

p < .05. **p < .01. ***p < .001.

The clips were also classified on the basis of location of the pedestrian: in street, left, right edge, and right side (n = 9, 10, 20, and 18, respectively). The category right edge was used for pedestrians walking on the edge of the road, whereas the category right side was used for pedestrians beyond the right edge of the road or on a sidewalk. There were insufficient cases on the left side and left edge to allow analysis of both categories separately.

Pedestrian motion was partitioned into four categories: same, opposite, into street, and standing (n = 12, 10, 10, and 25, respectively). The pedestrians classified as same and opposite were walking in a direction predominately parallel to the vehicle’s path in either the same or the opposite direction as the driver’s vehicle. Into street implied that the pedestrian was walking perpendicularly to the direction of vehicle travel and toward the center of the street. An example is a pedestrian approaching a zebra crossing. There were no cases in which the system issued an alert as a pedestrian walked perpendicularly away from the direction of vehicle travel.

The two categories of vehicle direction were vehicle straight (n = 47) and vehicle turning (n = 10). Similarly, road directions were classified as either straight road (n = 30) or turning road (n = 27). A vehicle turning in an intersection was classified as vehicle turning and straight road, whereas a vehicle traveling on a curved road was classified as vehicle straight and turning road.

Nonparametric Kruskal-Wallis one-way ANOVA by ranks and one-way parametric ANOVA of the raw ratings were calculated to ascertain whether drivers’ acceptance of alerts varies across the various subcategories of traffic context. Both tests reject the null hypothesis of no differences across the four categories of both pedestrian location and pedestrian motion. Figure 3 shows the mean ratings for these contextual variables. As predicted by the model of Himanen and Kumala (1988), our participants responded much more positively to alerts to pedestrians in the street and on the right edge of the street than to alerts to pedestrians to the left of the vehicle and on the right sidewalk. These differences are supported by both parametric and nonparametric post hoc comparisons (see Table 2).

Figure 3.

Mean and standard errors of ratings for the four subcategories of (a) pedestrian location and (b) pedestrian motion.

Table 2:

Summary of Post Hoc Comparisons of Subcategory Means

	Wilcoxon Rank Sum Test	Tukey’s HSD Test

Category	W	D	s(D)
Pedestrian location
In street, left	2.16*	20.5*	7.3
In street, right edge	−0.92	−2.3	6.4
In street, right side	3.63***	29.9***	6.5
Left, right edge	−2.71**	−22.8**	6.1
Left, right side	1.32	9.4	6.3
Right edge, right side	4.60***	32.2***	5.2
Pedestrian motion
Same, into street	−3.53***	−29.6***	8.2
Same, standing	−2.80**	−23.1**	8.2
Same, opposite	−1.64	−12.7	6.7
Into street, standing	1.40	6.5	8.6
Into street, opposite	2.17*	16.9*	7.2
Standing, opposite	1.08	10.4	7.2

p < .05. **p < .01. ***p < .001.

As shown in Figure 3b, the ratings for pedestrians walking in the same direction as the vehicle were much lower than they were for the other three categories of pedestrian motion. This finding may be discrepant with the observations of Schmidt and Färber (2009), who found that pedestrians facing a vehicle cross the road at greater distances. We were surprised that tests for the variable locale proved to be nonsignificant at an alpha level of .05. However, if we were to adopt an alpha of .10, the post hoc tests found that pedestrians in rural settings did elicit significantly higher levels of acceptance than did alerts in both the urban and suburban settings.

Discussion

The results suggest that as expected, pedestrian location and motion influence driver acceptance of alerts issued by our pedestrian detection system. Alerts are more accepted when pedestrians are close to or moving toward the vehicle’s path, that is, when they can be seen to infringe on the field of safe travel (Gibson & Crooks, 1938). The exercise of safe travel involves navigating through complex environments and, among other things, detecting and avoiding pedestrians. The driver’s perception of risk will direct his or her attention, resulting in greater attention to salient objects, such as pedestrians in the road.

Our data are equivocal in their support of the hypothesis that alerts to pedestrian encounters are more accepted when the driver is on rural roads. In coming studies, we will revisit the locale category and investigate whether locale information can be used to achieve alerting criteria with greater driver acceptance. Our data do not support the hypotheses that vehicle direction (turning or straight) and road geometry (curving or straight) influence driver acceptance. We do not have sufficient data at this time to investigate how the interaction of these factors may influence alert acceptance.

These findings support the argument that designers of the alerting criteria to be used by an active safety system should consider the contextual dependency of alert acceptance. Indeed, we suspect that the failure to consider contextual dependencies may explain why some emerging active safety systems have met with driver resistance. We argue that the acceptance of an alerting system is likely to be relatively high when, and only when, it issues alerts at times when and in situations in which drivers are likely to expect them (Källhammer et al., 2007). It should come as no surprise that these situations are contextually dependent (Himanen & Kumala, 1988; Schmidt & Färber, 2009) and involve violations of the field of safe travel.

Pedestrians in the field of safe travel put themselves and the driver at risk. The quality of the driver’s risk assessment depends on the adequacy of the available information (Williams & Noyes, 2007). Our findings lead us to suggest that drivers will assess the adequacy of a pedestrian alerting system by whether it contains alerting criteria that involve consideration of the field of safe travel. Such a system would issue alerts that drivers will be likely to both expect and welcome. By issuing alerts that drivers expect and welcome, the system would not only influence the drivers’ attention to potentially threatening situations but also shape their expectations and risk assessment (Neisser, 1976).

A major challenge to FOT studies is that most of the observed events are unique in various ways. The everyday context makes it difficult to experimentally control and accurately repeat trials (Walker, Stanton, & Young, 2008).

In this article, we have shown how we can efficiently investigate contextual factors that influence drivers’ acceptance of alerts to pedestrians by leveraging expensive FOT data within the confines of the laboratory. The laboratory review and rating of FOT data provides consistent measures of acceptance in a controlled environment and is a useful tool for developers of the alerting criteria used by active safety systems.

The FIR image used in this study provided participants with more contextual information than was available to the drivers looking through the windscreen. Not only are pedestrians more visible in the FIR image, but other ambient information is also clearly evident in the image. The FIR videos therefore preserve the context as well as a brief exposure can.

It remains an empirical question whether the length of our video clips (30 s) was sufficient to provide a full sense of the traffic context to our participants. From the standpoint of contextualization, nothing can match the FOT experience. However, it would be unrealistic and unproductive to expect participants in an experiment to sit through the video of an extended driving trip to set the stage for a random encounter with a pedestrian. The main limitation to our study may be that the 30-s duration of our clips deprives participants of useful contextual information.

Additional studies with participants with less experience and other demographic backgrounds are needed to test whether the results generalize to other nationalities and age groups of drivers.

Key Points

Pedestrian location and motion influence driver acceptance of alerts issued by our pedestrian detection system.

Alerts are more accepted when pedestrians can be seen to infringe on the field of safe travel.

The design of the alerting criteria to be used by an active safety system should include consideration of the contextual dependency of alert acceptance.

The study demonstrates a useful method to leverage expensive field operational test data within the confines of the laboratory.

Footnotes

Acknowledgements

We are grateful to the Human Factors and Ergonomics Society editors and three anonymous reviewers for their helpful comments to improve this article.

Jan-Erik Källhammer is head of active safety research at Autoliv Research in Vårgårda, Sweden. He received his MSc in mechanical engineering from Luleå Technical University in Sweden, his MSc in electrical engineering from Duke University in Durham, North Carolina, and his PhD from the Department of Information and Computer Science at Linköping University, Sweden. His current research focuses on the assessment of driver acceptance of advanced driver assistance systems.

Kip Smith is a senior lecturer of human systems integration at the Naval Postgraduate School in Monterey, California, and principle scientist at Cognitive Engineering and Decision Making, Inc., in Des Moines, Washington. He received his MSc in geophysics from the University of New Mexico and his PhD in management from the University of Minnesota. His work focuses on the influence of context and culture on individual and group decision making in dynamic environments.

References

Farber

Paley

(1993). Using freeway traffic data to estimate the effectiveness of rear-end collision countermeasures. In Proceedings of the 1993 Annual Meeting of IVHS America (pp. 260–268). Washington, DC: IVHS America.

Gibson

J. J.

Crooks

L. E.

(1938). A theoretical field-analysis of automobile driving. American Journal of Psychology, 51, 453–471.

Himanen

Kumala

(1988). An application of logit models in analyzing the behavior of pedestrians and car drivers on pedestrian crossings. Accident Analysis & Prevention, 20, 187–197.

Jackson

Chapman

Crundall

(2009). What happens next? Predicting other road users’ behavior as a function of driving experience and processing time. Ergonomics, 52, 154–164.

Källhammer

J. E.

Smith

Karlsson

Hollnagel

(2007). Shouldn’t the vehicle react as the driver expects? In Proceedings of the 4th International Driving Symposium on Human Factors in Driver Assessment, Training, and Vehicle Design (pp. 9–15). Iowa City, IA: The University of Iowa.

Lerner

N. D.

Dekker

D. K.

Steinberg

G. V.

Huey

R.W.

(1996). Inappropriate alarm rates and driver annoyance (Tech. Rep. DOT HS 808 533). Retrieved from http://www.itsdocs.fhwa.dot.gov/JPODOCS/REPTS_TE/3757.pdf

Marshall

D. C.

Lee

J. D.

Austria

P. A.

(2007). Alerts for in-vehicle information systems: Annoyance, urgency, and appropriateness. Human Factors, 49, 145–157.

Neisser

(1976). Cognition and reality: Principles and implications of cognitive psychology. San Francisco, CA: Freeman Press.

Parasuraman

Riley

(1997). Humans and automation: Use, misuse, disuse, abuse. Human Factors, 39, 230–253.

10.

Schmidt

Färber

(2009). Pedestrians at the kerb: Recognising the action intentions of humans. Transportation Research–Part F, 12, 300–310.

11.

Sheridan

T. B.

(2008). Risk, human error, and system resilience: Fundamental ideas. Human Factors, 50, 418–426.

12.

Siegel

Castellan

N. J.

Jr. (1988). Nonparametric statistics for the behavioral sciences (2nd ed.). New York, NY: McGraw-Hill.

13.

van der Laan

J. D.

Heino

De Waard

(1997). A simple procedure for the assessment of acceptance of advanced transport telematics. Transportation Research–Part C: Emerging Technologies, 5, 1–10.

14.

Vetenskapsrådet (2002). Forskningsetiska principer inom humanistisk-samhällsvetenskaplig forskning [Ethical principles for research using human participants]. Stockholm, Sweden: Author.

15.

Walker

G. H.

Stanton

N. A.

Young

M. S.

(2008). Feedback and driver situation awareness (SA): A comparison of SA measures and contexts. Transportation Research–Part F, 11, 282–299.

16.

Williams

D. J.

Noyes

J. M.

(2007). How does our perception of risk influence decision-making? Implications for the design of risk information. Theoretical Issues in Ergonomics Science, 8, 1–35.