Contents lists available at ScienceDirect

Accident Analysis and Prevention

journal homepage: www.elsevier.com/locate/aap

Glass half-full: On-road glance metrics differentiate crashes from near-crashes in the 100-Car data

Bobbie D. Seppelta,b,⁎, Sean Seamana, Joonbum Leeb, Linda S. Angella, Bruce Mehlerb,Bryan Reimerb

a Touchstone Evaluations, Inc., 18160 Mack Avenue, Grosse Pointe, MI 48230, United Statesb Massachusetts Institute of Technology AgeLab &New England Univerity Transportation Center, 77 Massachusetts Avenue, Room E40-289, Cambridge, MA 02139,United States

A R T I C L E I N F O

Keywords:Attention managementDriver distraction100-CarNaturalistic driving studySafety-critical events

A B S T R A C T

Background: Much of the driver distraction and inattention work to date has focused on concerns over driversremoving their eyes from the forward roadway to perform non-driving-related tasks, and its demonstrable link tosafety consequences when these glances are timed at inopportune moments. This extensive literature has es-tablished, through the analyses of glance from naturalistic datasets, a clear relationship between eyes-off-road,lead vehicle closing kinematics, and near-crash/crash involvement.Objective: This paper looks at the role of driver expectation in influencing drivers’ decisions about when and forhow long to remove their eyes from the forward roadway in an analysis that consider the combined role of on-and off-road glances.Method: Using glance data collected in the 100-Car Naturalistic Driving Study (NDS), near-crashes were ex-amined separately from crashes to examine how momentary differences in glance allocation over the 25-s priorto a precipitating event can differentiate between these two distinct outcomes. Individual glance metrics of meansingle glance duration (MSGD), total glance time (TGT), and glance count for off-road and on-road glance lo-cations were analyzed. Output from the AttenD algorithm (Kircher and Ahlström, 2009) was also analyzed as ahybrid measure; in threading together on- and off-road glances over time, its output produces a pattern of glancebehavior meaningful for examining attentional effects.Results: Individual glance metrics calculated at the epoch-level and binned by 10-s units of time across theavailable epoch lengths revealed that drivers in near-crashes have significantly longer on-road glances, and lookless frequently between on- and off- road locations in the moments preceding a precipitating event as comparedto crashes. During on-road glances, drivers in near-crashes were found to more frequently sample peripheralregions of the roadway than drivers in crashes. Output from the AttenD algorithm affirmed the cumulative netbenefit of longer on-road glances and of improved attention management between on- and off-road locations.Conclusion: The finding of longer on-road glances differentiating between safety-critical outcomes in the 100-CarNDS data underscores the importance of attention management in how drivers look both on and off the road. It isin the pattern of glances to and from the forward roadway that drivers obtained critical information necessary toinform their expectation of hazard potential to avoid a crash.Application: This work may have important implications for attention management in the context of the in-creasing prevalence of in-vehicle demands as well as of vehicle automation.

1. Introduction

Drivers today are faced with increased competition for their atten-tion due to the increased presence of in-vehicle information systems(IVIS) and cellular connectivity applications, from satellite navigation,in-car infotainment systems and embedded vehicle systems, to

smartphone applications. In interacting with these devices, drivers mustmanage their attention to and from the roadway, deciding when and forhow long to remove their eyes from the forward roadway − as such,often fracturing attention to the driving task. Driving even in its mostbasic form requires management of visual resources to multiple loca-tions, including the immediate road surroundings, general

http://dx.doi.org/10.1016/j.aap.2017.07.021Received 31 December 2016; Received in revised form 30 May 2017; Accepted 18 July 2017

⁎ Corresponding author.E-mail addresses: bseppelt@touchstoneevaluations.com (B.D. Seppelt), sseaman@touchstoneevaluations.com (S. Seaman), joonbum@mit.edu (J. Lee),

langell@touchstoneevaluations.com (L.S. Angell), bmehler@mit.edu (B. Mehler), reimer@mit.edu (B. Reimer).

Accident Analysis and Prevention 107 (2017) 48–62

0001-4575/ © 2017 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/BY/4.0/).

MARK

surroundings, in-cab instrumentation, passengers, and other objects.Naturally occurring ancillary activities − mind-wandering, roadsideadvertising, etc. − can compete to draw attention away from theroadway. In the past several years, increased vehicle automation (e.g.,power steering, automatic transmissions, adaptive cruise control, lanekeeping aids, etc.), in its affordances for reduced physical and atten-tional demands of the driving task, frees-up attentional resources,which drivers may allocate to non-operational activities. Over thecoming years, further increases in vehicle automation are expected tooccur, potentially amplifying observed changes in the allocation of at-tention. Methods are needed that provide a deeper understanding of thefoundations of attentional strategies to assist in the development ofinterfaces, assistive technologies, and management systems that pro-mote more effective attention allocation under both traditional manualcontrol as well as in drivers’ use of vehicle automation. Given the ex-pected competition for drivers’ attentional resources, it is important tounderstand behaviors that lead to adverse events (those well-docu-mented within the annals of driver distraction work) − the glass half-empty perspective − as well as to study those that are preventive ofsuch outcomes − the glass half-full perspective.

Much of the driver distraction and inattention work has focused onconcerns over drivers removing their eyes from the forward roadway toperform non-driving-related tasks, and its demonstrable link to safetyconsequences when these glances are timed at inopportune moments.This broad and exceptional line of work has established, through theanalyses of glance from naturalistic datasets, a clear relationship be-tween eyes-off-road, lead vehicle closing kinematics, and near-crash/crash involvement (Klauer et al., 2006; Victor et al., 2015). Large-scalenaturalistic datasets such as those collected as part of the 100-CarNaturalistic Driving Study (Dingus et al., 2006) and the SHRP 2 Nat-uralistic Driving Study (Hallmark et al., 2013) contain safety-criticalevents (SCEs) and periods of baseline “normal” driving that have al-lowed study of human behaviors that are assumed to have a con-tributing role to crashes. In previous analyses of these datasets, the roleof driver expectation influencing drivers’ decisions about when and forhow long to remove their eyes from the forward roadway − i.e., anindicator of how drivers allocate their attention − have been discussedbut not directly analyzed. Instead, analyses are primarily constrained tocharacterizing the interacting set of driver, task, and environmentalfactors, i.e., the event dynamics, in the seconds leading up to near-crashes and crashes. In this paper, we focus on how glance behaviorsfurther upstream from the typically considered 0–15 s of time sur-rounding SCEs can provide insight into downstream consequences.Using data collected in the 100-Car Naturalistic Driving Study (NDS),this study separates near-crashes from crashes to examine how mo-mentary differences in glance allocation over the 25 s prior to theevents can differentiate between these two distinct outcomes (i.e.,crashes result in significantly different outcomes in terms of loss of lifeor injury, and property damage as compared to near-crashes, in whichthese consequences are avoided).

1.1. Driver distraction and inattention problem − the glass-half-emptyperspective

Driver distraction is commonly referred to as “the diversion of at-tention away from activities critical for safe driving toward a competingactivity, which may result in insufficient or no attention to activities criticalfor safe driving” (Regan et al., 2009, p. 7). From this definition, attentionrefers to visual, auditory, cognitive, and manual resources, and all mayplay a role in affecting driver performance (e.g., Regan et al., 2009).Expanded considerations of vocal and haptic resources in many moderninterfaces may also need to be considered in this context (Reimer et al.,2016). Diverted attention from the road can lead to deteriorated controlof the vehicle (Crisler et al., 2008; Drews et al., 2009; Hosking et al.,2009), missed events (Fitch et al., 2009; Hosking et al., 2009), anddelayed reaction times (Drews et al., 2009; Horrey and Wickens, 2004;

Lee et al., 2004). Concurrently timed with other contributing factors(Angell et al., 2006), such as the presence of a junction, urban drivingor unexpected events, these lapses in performance can lead to crashes.Considerable research, from in-depth accident analysis to findings fromnaturalistic driving studies, have documented this link between failuresof attention and road crashes (e.g., Dingus et al., 2006; Hickman et al.,2010; Olson et al., 2009; Treat et al., 1977; Wang et al., 1996).

Glance location can be used as a proxy for determining what in-formation a driver is processing; for example, glances (or gaze direc-tion) can be used to infer if the driver’s attention is directed to theforward roadway, or to objects near the road or within the vehicle(Bellenkes et al., 1997; Taylor et al., 2013; Wickens et al., 2003).Though gaze direction and focus of attention are closely linked(Theeuwes et al., 1998; Yantis and Jonides, 1990), they do not have aperfect coupling (Hafed and Clark, 2002; Hunt and Kingstone, 2003;Posner, 1980). Direction of gaze may sometimes be separately locatedfrom cognitive attention, as is represented by glance behaviors such as“vacant staring” (Fletcher and Zelinsky, 2007; Pohl et al., 2007), “lookbut not see” (e.g., Fletcher and Zelinsky, 2009), and gaze concentration(Recarte and Nunes, 2000; Reimer et al., 2012; Victor et al., 2005;Wang et al., 2014; Zhang et al., 2006). Recognizing this limitation ofglance-based analyses, there is, nonetheless, a large body of researchthat has established the close relationship between the location of aglance and the information to which an individual is attending undermany conditions in normal visual perception (e.g., Birrell and Fowkes,2014; Liang et al., 2012; Regan et al., 2013).

Individually, analyses of glance duration and of glance frequencyhave shown a quantifiable link to safety. Single long off-road glances −those greater than 2 s − in controlled experiments are associated withmore frequent lane deviation and slower response to lead vehiclebraking (Dingus et al., 1989). In the study of visual time sharing be-tween the driving task and secondary activities, drivers do not tend tohold glances away from the road for durations beyond 1.6–2.0 s formore difficult or longer tasks (Liang et al., 2014; Sodhi et al., 2002),likely due to the relatively robust performance consequence of de-partures from the lane with glances longer than 2 s (e.g., Senders et al.,1967; Wierwille, 1993b); instead, drivers more typically increase theirnumber of glances away from the road (e.g., Victor et al., 2005; Younget al., 2005). In a recent analysis of the 100-Car NDS, such an increasedfrequency of off-road glances in the approach to near-crash/crashevents is evident (Liang et al., 2014). However, in interpretation of off-road glance frequency, it is necessary to consider its relationship withtask duration, i.e., longer tasks led to more frequent off-road glancesbut individual glance durations tended to be under 2s. Accumulated off-road glances to in-car devices and secondary activities have also beenshown to be detrimental to safety at long durations (e.g., Donmez et al.,2007; Klauer et al., 2010; Senders et al., 1967). A higher percentage ofeyes-off-road time has also been associated with an increased likelihoodof the occurrence of SCEs (Klauer et al., 2006; Olson et al., 2009;Hickman and Hanowski, 2012). Duration of off-road glance, then, bothin the form of a single long duration and in aggregated form across aunit of time provides a robust measure of visual distraction when pre-dicated on safety-relevant outcomes (e.g., Liang et al., 2012; Lianget al., 2014).

In addition to single metrics of off-road glance behavior, a numberof glance-based algorithms have considered the combined effects ofduration and frequency over time in predicting safety outcomes (e.g.,Ahlström et al., 2009, 2013; Donmez et al., 2007, 2008; Klauer et al.,2006; Liang, 2009; Pohl et al., 2007; Rydström, 2007; Senders, 1967;Tian et al., 2013; Victor et al., 2005; Zhang and Smith, 2004). Morecomprehensive overviews of driver distraction detection algorithmsexist within the volumes of driver distraction research (e.g., Dong et al.,2011; Lee et al., 2017; Moeckli et al., 2013; Liang et al., 2012).

Behind this focus on off-road glances is the distraction-basedgrounding: when drivers divert their gaze from the forward roadway(often the result of a visual-manual activity) and allow themselves to

B.D. Seppelt et al. Accident Analysis and Prevention 107 (2017) 48–62

49

become less aware of the driving environment, they increase their riskof a safety-critical event (Klauer et al., 2006; Olson et al., 2009;Hickman et al., 2010; Victor et al., 2015). While off-road glance metricshave definitive links to safety outcomes, these measures do not considerthe attentional precursors to a driver’s decision to initiate long glancesaway from the forward roadway. A proposed alternative perspective tothe more traditional distraction-related study of off-road glance beha-vior is to take an attention management perspective and also considerhow drivers allocate their glances on-road over time − analyzing bothsingle on-road glance behaviors and patterns of glance on- and off-roadwell upstream of a precipitating event. Such an analysis is expected tohelp develop deeper insight into the previously-attended informationmotivating perceived relative risk probabilities that underlie a driver’sdecision to redirect gaze.

1.2. Importance of on-road sampling behavior − the glass-half-fullperspective

On-road glance behavior provides a complementary and importantmirrored perspective to off-road glance behavior in understanding theeffects of visual demand on attention management.

Drivers must sample the road environment long enough in-betweenoff-road glances to build and maintain an accurate representation of thedriving scene to maintain basic vehicle control, and to detect cues in theenvironment that may signal potential hazards (e.g., Angell, 2007;Taylor et al., 2011). A useful concept to characterize the effect of an on-road glance on a driver’s awareness of the meaning of dynamic changesin their environment is situation awareness (SA), defined as “the per-ception of the elements in the environment within a volume of time andspace, the comprehension of their meaning, and the projection of their statusin the near future” (Endsley, 1995, p. 36). A higher level of SA − that inwhich projection of environmental state occurs − depends upon per-ception and comprehension of relevant cues. How long, when, andwhere drivers look on-road dictates the visual information used to formexpectations for how variables used to perform the driving task willevolve in the next moments of time. Cues such as the vehicle’s positionin the lane, relative to surrounding traffic, and on the location of re-levant features of the environment relative to the path of travel are usedto generate a moment-to-moment situation model, which affects deci-sion-making and action behaviors used to perform the driving task.

From vision science work on scene perception we know that the“gist” of individual scenes can be acquired very rapidly − well withinthe duration of a single fixation, and typically less than a second(Biederman et al., 1982; Chun and Jiang, 1998; Friedman, 1979;Greene and Oliva, 2009; Potter, 1999; Schyns and Oliva, 1994; Thorpeet al., 1996). However, when knitting together dynamic scenes and inprojecting the locations and trajectories of events and objects requiredas part of the hazard identification and response task of driving, sceneawareness is more complex.

The type of cuing influences the amount and type of processingresources deployed for event detection and response. Salient visualproperties induce automatic, reflexive, bottom-up attention selection.However, this reflexive selection effect is attenuated with increasedeccentricity of the stimuli from the direction of gaze (Anstis, 1974), or,if masked by a visual disruption such as a blink, saccade (fast eyemovement), occlusion, or an initial onset outside of the visual field ofview (an effect well-cited within a large body of research on changeblindness; e.g., Rensink, 2002; Stelmach et al., 1984). In addition toeffects from such data limits due to missing or occluded information,event detection and response can also be affected by resource limits onbottom-up perceptual processes (e.g., Angell, 2007).

Event detection and response also depends on top-down facilitation(and/or inhibition), i.e., supervisory control processes (e.g., Angell,2007). These processes function to “prime” or “cue” a driver to a re-levant event’s possible occurrence at a spatial location, and enable adriver to ignore irrelevant stimuli competing for his/her attention

(Lavie, 1995, 2005). With experience, drivers learn to anticipate ha-zards (Crundall and Underwood, 1998; Falkmer and Gregersen, 2005).The role of knowledge or expectation is evident in comparisons be-tween novice and experienced drivers, in which poor hazard anticipa-tion performance for novices is attributed to their lack of knowledge forhow to navigate them (e.g., McKnight and McKnight, 2003) and of theinformation that should be sampled in the driving scene. In one ex-ample, novice and experienced drivers were placed in situations inwhich latent threats (those difficult to detect either because they are notvisible until the last moment, e.g., a pedestrian emerging from behind astopped truck in the right parking lane of an un-signaled midblockcrosswalk, or because they are visible but not yet fully materialized,e.g., a car that is stopped in a long line of cars in a left turn lane, butwhich might pull out in front of a driver who is in the adjacent, rightlane of travel; Taylor et al., 2011) were presented at multiple locations.In this study, experienced drivers were more likely to gaze at locationsindicative of their anticipation of the latent threats as compared tonovice drivers, and consequently detected more of the resultant hazards(82% detected for experienced drivers versus 47% for novice drivers).

The duration of a glance to the road importantly reflects the pro-cessing required to resolve information needed to reduce uncertaintyabout evolving road/traffic dynamics and environmental state (e.g.,Henderson, 2003). If cued to return gaze to the forward roadway, ittakes drivers 2–4 s to verify hazard presence and to then initiate a re-sponse (Glaser et al., 2016; note: this range reflects a dependency on theproperties and dynamics of individual hazards, and the time required toboth detect and respond to a hazard following from an off-road glance).In his meta-analysis of simulator studies, controller road studies, andnaturalistic observation, Green (2000) cited a response range as low as0.7 s for expected signals, 1.25 s for unexpected signals, and 1.5 s forsurprise events. All of the included studies in his meta-review, however,used overt cuing to signal the event, and measured responses con-strained to ideal detection conditions − i.e., daylight, good weather,clear visibility, etc. The combined range of reported hazard response(0.7 s–4 s) points out that not all hazards are created equal: some aremore difficult to correctly recognize as hazards because they are furtherafield, have low salience in contrast to the surrounding environment, ordue to the absence of bottom-up cuing to signal their presence.

To detect the presence of potential hazards that require top-down(or schema-driven) processing of the driving environment, Samuel andFisher (2015) recently reported that drivers need 4 s of on-road glancetime at a minimum, but potentially on the order of 7 s or more forreliable detection; notably, this time estimate does not factor in theadditional time required for response selection and action. In this study,drivers were asked to perform simulated vehicle tasks, alternating be-tween views of the forward roadway and an in-vehicle location inwhich the length of provided on-screen information varied from 1 to 4 sin between 2 s off-road glance intervals (thus providing drivers amaximum, fixed 2 s of viewing time to the in-vehicle task location).During this alternating glance sequence, multiple potential (or latent)threats were presented. A glance towards the region of a potentialthreat at a specified location in advance of its unfolding was used todefine detection. Drivers detected significantly more events when pro-vided with the longest view of the forward road (4s). Further, there wasevidence that drivers’ top-down processing of events − or lack of ha-zard anticipation as defined from the lack of a glance to the specifiedregion − was inhibited at the shorter on-road glance durations, sup-porting the hypothesis that longer on-road viewing time provides anopportunity for drivers to sample relevant cues, enabling them to formadequate SA, and to in turn guide their expectancy of events. In sum,when events are expected, detection is speeded by top-down attentionalfacilitation, and when events are unexpected, detection occurs in abottom-up, stimulus-driven way − which is a different attentionalprocess − one slower on the order of hundreds of milliseconds toseconds. To build and maintain an accurate representation of thedriving scene to support expectancy of events, drivers must sample the

B.D. Seppelt et al. Accident Analysis and Prevention 107 (2017) 48–62

50

road environment long enough to both detect predictive cues and tohave an understanding of the locations in the driving environment re-levant for detecting potential hazards.

In addition to the importance of the length of an individual on-roadglance towards providing a driver with adequate time to perceive re-levant cues to then develop appropriate expectations about upcomingevents, the pattern of glances off-road may produce a cumulative effecton a driver’s awareness of roadway context (Liang et al., 2012). Instudying the distribution of off-road glances leading up to near-crash/crash events in the 100-Car dataset as compared to baseline conditions,Liang et al. (2014) noted that “not only are crashes and near-crashes morelikely to occur when drivers look away from the forward road, but there is adistinct pattern of glance behavior (compared to baseline) occurring as earlyas 30 seconds before a critical event” (p. 2105); the authors extrapolatedthese results to infer a likely cumulative role of neglect of the forwardroadway in contributing to crash risk.

Of the algorithms proposed to estimate the effects of off-roadglances (e.g., Liang et al., 2012; Lee et al., 2013), the AttenD algorithmis the only one that considers the effect of on-road and off-road glancesover time in either building or depleting, respectively, a reserve of in-formation about the roadway (akin to scene awareness or SA; Kircherand Ahlström, 2009, 2012). This algorithm takes existing glance data−multiple, discrete variables of glance duration, frequency, and location− and combines them into a single continuous hybrid metric. Usingsimple rules for how these variables affect this scalar value; the outputvalue is a real-time estimate of information decay and recovery − alsoreferred to as a “buffer”. The initial buffer value is set at 2 and for eachsecond of off-road glance decreases at a rate of one unit until aminimum value of 0 is reached, at which point it does not decreasefurther but stays at 0 until the driver glances back to the roadway. Thevalue increases at the same rate for each second of on-road glance aftera latency period of 0.1s, which reflect a psychological refractory periodfrom the cost of switching attention back to the roadway, until it re-turns to 2, at which point it does not increase further. If a driver glancesat the mirrors or the speedometer (regions that inform a driver’s abilityto perform the primary vehicle control tasks relative to surroundingtraffic dynamics), there is a latency of 1 s before the scalar value de-creases. By design, the AttenD algorithm was intended by its authors topredict distraction potential in which a 0 value demarks a distractedstate (Kircher and Ahlström, 2009, 2012). However, in its ability torepresent glance history in the form of a profile that dynamicallychanges to reflect the amount of stored information about the roadway,it produces a pattern of glance behavior meaningful for examining at-tentional effects. The output value offers the potential to delineate si-tuations in which drivers have appropriately sampled on-road in-formation to avoid a crash versus those situations in which they haveinsufficiently maintained an awareness of the roadway. It is from thisperspective − of its output value providing an indication of how ef-fectively attention is allocated to the roadway across time − fromwhich the AttenD algorithm is analyzed within this paper.

1.3. Implications for analyzing glance behavior in SCEs

Previous analyses of naturalistic data have focused on tasks whichelevate crash risk (using odds-ratios), and on eyes-off-road glancemeasures in predicting distraction potential, focusing on the few sec-onds, typically 0–15 s, prior to near-crash/crash or SCEs. This commonfocus around the seconds leading up to SCEs may not provide a largeenough analysis window to evaluate the role of expectancy in informingdownstream glance behavior. Previous reports of the 100-Car NDS haveshown that drivers are typically looking away from the roadway at thetime of near-crash/crash events (e.g., Liang et al., 2014; Victor et al.,2015) − but, the unanswered question is why? Were drivers simplyinopportune in their glance away from the roadway at the moment anevent happened (Victor et al., 2015)? Or, were these off-road glances inthe moments before a crash an outcome of poorly-formed expectations

of the driving environment due to insufficient on-road sampling? Ananalysis of glance behavior in the seconds prior to safety-critical out-comes is expected to reveal breakdowns in upstream information pro-cesses, i.e., a breakdown in a proactive barrier − those mismatchesbetween proactive schema selection and the actual situation (Engströmet al., 2013), or from a depleted situation model (e.g., Horswill andMcKenna, 2004) to the extent there are significant differences in thepattern and length of on-road sampling between near-crash and crashevents.

Commonly, near-crash and crash events in naturalistic drivinganalyses are combined in assessments of risk associated with glancebehaviors, placed under a presumed equivalence label of “safety-cri-tical”. While challenges to this approach have appeared (Knipling,2015), this combining of near-crashes and crashes is typically due tolow sample sizes of crash events (for the purpose of computing oddsratios) − along with what has been argued to be a similarity in thekinematics of the two categories in the final moments prior to a pre-cipitating factor. However, in this study, we hypothesize that the way inwhich near-crashes differ from crashes is, in fact, an important differ-ence. Namely, in a near-crash, the driver successfully avoids a crash −presumably by shifting attention at the “right time” and to the rightplace in a way that enables him/her to take an avoidance action andprevent a crash. A key question is this: what pattern of glancing orattending enables a near-crash to be different from a crash, and can areliable and distinct pattern be identified that differentiates these twooutcomes? In this paper, consequently, crashes were analyzed sepa-rately from near-crashes due to their inherent inequality in causalityand harm (Knipling, 2015) to study temporal differences in on-roadvisual sampling.

It is in recognizing the important role that hazard awareness andanticipation plays in crash risk (e.g., Horswill and McKenna, 2004) thatthis paper proposes to analyze on-road glance behavior (both singleglance duration and the pattern of glances over time) using the existingpublicly-available 100-Car NDS dataset. The central thesis of this paperis that the ability to form appropriate expectations of events (informinga driver when to time off-road glances in the moments prior to a pre-cipitating events) depends on the amount of information sampled fromthe forward roadway − its frequency and duration. In differentiatingbetween near-crash and crash events in the 100-car dataset, it is hy-pothesized that allocation of on-road glances over time induces a cu-mulative a priori for ill-timed off-road glances near to SCEs.

2. Method

2.1. The 100-Car naturalistic driving study data

The 100-Car Naturalistic Driving Study continuously recorded on-road data from 100 vehicles between January 2003 and June 2004,collecting approximately 2 million vehicle miles or nearly 43,000 h ofdata from 241 primary and secondary drivers in unobtrusively in-strumented vehicles (Dingus et al., 2006). The data contains periods ofnormal, daily driving; epochs of fatigue, impairment, judgment error,risk taking, engagement in non-driving related tasks, aggressivedriving, and traffic violations −behaviors and performance outcomesrepresentative of real-world driving.

Two databases with video and electronic driver and vehicle per-formance data were constructed from the data by the Virginia TechTransportation Institute (VTTI): a baseline database and an event da-tabase. The baseline database was created by stratifying the entire da-taset based upon the number of crashes and near-crashes each vehiclewas involved in and then randomly selecting 20,000 6.1-s segmentsfrom the driving data. This stratification of the baseline epochs wasperformed to create a case-controlled dataset in which there weremultiple baseline epochs per each crash or near-crash event to allow forcalculation of odds ratios. The baseline database for the 20,000 epochsincludes vehicle, environmental, and driver state variables. Eye glance

B.D. Seppelt et al. Accident Analysis and Prevention 107 (2017) 48–62

51

analyses were performed for 5000 of these baseline epochs.The event database includes vehicle and glance data for a set of

near-crashes and crashes for a period of time surrounding each event:up to 30 s prior to and until 10 s after the precipitating factor: in total,68 crashes and 760 near-crashes. Glance sequences for each event inthese two databases were scored by a single trained reductionist at VTTIfrom the video in a frame-by-frame manner, i.e., human reductionistsviewed video of each driver’s face and coded for each frame (at 10 Hz)the location of gaze among a set of 13 locations: “cell phone”, “centerstack”, “instrument cluster”, “interior object”, “passenger”, “left for-ward”, “left mirror”, “left window”, “rearview mirror”, “right forward”,“right mirror”, “right window”, and “forward”. Two additional glancecodes were possible for recording instances of missing data: “eyesclosed” and “no video”.

The glance sequences and location codes from the 100-car databasewere downloaded from the online repository (http://forums.vtti.vt.edu/index.php?/files/category/3-100-car-data/) for the full period ofavailable data for baseline epochs and for events. Key operational de-finitions are listed in Table 1.

As detailed above, for each near-crash (NC) and crash (C) event,additional glance data were coded after the identified precipitatingfactor (Klauer et al., 2006). For the analyses reported in this paper, allof the epochs were truncated at the precipitating factor (i.e., the sec-onds after the precipitating event were not included in the analysiswindow). Setting the precipitating event as the cutoff point for theepochs allowed us to isolate our analysis to driver glance behaviorleading up to a critical event. We assumed that after the precipitatingfactor, glance behavior would fundamentally change—moving from a“proactive” state to a “reactive” state (Engström et al., 2013), in thatthe precipitating event is expected to induce different attentional pro-cesses (those triggered from a precipitating event’s exogenous stimuliversus from endogenous cues, i.e., changes in the driver’s internally-active goals or processes). The salient properties of this event in theform of transient onsets or looming are expected to dictate attention toNC/C conditions at or in the time following the precipitating factor(Markkula et al., 2016). In particular, driver response following thispoint would depend on kinematic urgency and reaction speed to initiatean avoidance maneuver. More practically, the precipitating factor is anormalized point at which to cut the epochs to create a cleaner datasetfor comparison of near-crashes and crashes. Epochs with missing eye-glance data were filtered from the downloaded data, reducing thenumber of events for analysis within this paper to 55 crashes and 724near-crashes (from the original set of 68 crashes and 760 near-crashes).

2.2. Epoch filtering

Epochs in the event dataset reflected a mixture of pre-crash beha-viors. However, the objective of this research was to examine the re-lationship between attention and glance behavior prior to crash. It wasimportant to filter out any epochs of crash and near-crash that mayhave arisen from factors that would confound this relationship, or alterits interpretation in a fundamental (and perhaps unknowable) way.Therefore, filters were applied in a controlled manner to distill the set ofepochs to a set which would provide a clean and clear test of hypotheses

− and to allow effects of filtering to be examined and reported.Epochs were filtered by applying four criteria:

2.2.1. Aggressive drivingEpochs containing aggressive driving, as coded in the 100-car da-

taset, were not included in the filtered dataset. It was hypothesized thatthe relationship between glance behavior and attention would be dif-ferent in an aggressive driving epoch than a non-aggressive drivingepoch.

2.2.2. Parking lotsEpochs that were recorded in driving contexts that were classified as

being in a parking lot were not included in the filtered dataset. It washypothesized that glance behavior in a parking lot would be differentthan glance behavior outside of a parking lot, with more driving-re-levant glances away from the windshield.

2.2.3. DrowsinessEpochs that were coded as containing drowsy driving were not in-

cluded in the filtered dataset. It was hypothesized that the relationshipbetween glance behavior and attention could be categorically differentin drowsy driving epochs, with either more frames coded as “eyesclosed,” or more coded as “eye forward,” but “inattention,” glancesforward”.

2.2.4. Missing videoEpochs that contained frames coded as “missing video” were not

included in the filtered dataset. Because we were ultimately interestedin the effect of glance behavior over time on the outcome of SCEs, wedid not want to remove “missing” frames while retaining the largerepoch, since doing so would either have resulted in treating the framesas missing data or would have forced non-contingent glances intotemporally-contingent locations. Therefore, we excluded epochs con-taining missing video.

These filtering choices represented a compromise between reason-ably eliminating confounds in the relationship between attention andglance behavior, and preserving an adequate number of epochs for astatistical analysis. The numbers of epochs removed at each stage offiltering are presented in Table 2, further broken down by the amountof data available from the online 100-car event database in each tem-poral bin (see next section). The bolded final row of Table 2 representsthe filtered set of epochs used for the remaining analyses. For any rowof the table, the number of epochs with glances in a bin increases acrossthe row, moving from those with at least 25 s of data to those with≤5 sof data, reflecting the varying length of epochs in the event dataset.

To reduce the effects of missing data on the crash/near-crash con-trast, we restricted our analysis to the 20 crash epochs and 368 near-crash epochs with at least 25 s of data, and cropped the onset of eachepoch to 25s, leaving us with a substantially smaller dataset of identicalduration epochs.

Notably, this filtered set of epochs does not represent a set of in-dependent observations, as multiple near-crash epochs may originatefrom one driver, and some drivers were associated with epochs in bothnear-crash and crash conditions. To adjust for this lack of independence

Table 1Key Operational Definitionsa.

Precipitating factor The driver behavior or state of the environment that initiates the crash, near-crash, or incident and the subsequent sequence of actions that results in anincident, near-crash, or crash.

Near-crash Any circumstance that requires a rapid, evasive maneuver by the subject vehicle, or any other vehicle, pedestrian, cyclist, or animal to avoid a crash. Arapid, evasive maneuver is defined as a steering, braking, accelerating, or any combination of control inputs that approaches the limits of the vehiclecapabilities.

Crash Any contact with an object, either moving or fixed, at any speed in which kinetic energy is measurably transferred or dissipated. Includes other vehicles,roadside barriers, objects on or off of the roadway, pedestrians, cyclists, or animals.

a Definitions from Dingus et al. (2006).

B.D. Seppelt et al. Accident Analysis and Prevention 107 (2017) 48–62

52

of samples, we evaluated three methods to account for the unevencrossing of drivers by SCE condition: by looking at those fully-crossed(drivers with both crash and near-crash epochs), looking at those fully-uncrossed (drivers with only epochs in either SCE condition) or a mix(in which all crash epochs [crossed or otherwise], and only near-crashepochs from drivers who did not crash, were included). The later ap-proach was selected in a final filtering of the epochs because it yieldedthe highest number of crashes, which were particularly sparse ascompared to the filtered set of near-crashes (Table 2). The final datasetfor analysis therefore included glance data points for 77 drivers, in-cluding 16 drivers involved in crashes and 61 drivers involved in near-crashes. Reported findings therefore represent independent observa-tions, calculated from exclusive crash or near-crash epochs of identicallengths.

3. Results

Analysis of the pre-precipitating event glance data was first com-puted using the following metrics for both off- and on-road glance lo-cations: Mean single glance duration (MSGD); a count of glances; andtotal glance time (TGT).

3.1. Mean single glance duration (MSGD)

The duration of individual glances was computed as the number offrames subtended by an individual glance. Individual glances werecategorized using the coding format of the publicly available 100-cardata (Table 3): an individual glance was defined from when the glancelocation code of an individually-coded frame of glance data changed inthe glance log, indicating the onset of a new glance, to when the glancelocation code changed again, indicating the offset of that same glance.Glances within “off-road” locations were coded separately for the pur-poses of these glance metrics; two consecutive glances to locations ca-tegorized as “off-road” but coded as having been made to differentspecific locations (such as “center stack” and “right window”) werecomputed as two individual glances when computing MSGD andnumber of glances. This same convention was used to designate in-dividual “on-road” glances. In the publicly-available 100-car data

dictionary, definitions for glance location to the forward roadway weresplit into three regions, defined as being “out the ‘left’, ‘straight’, or‘right’ windshield”, respectively. As defined in Table 3, left and rightforward locations were included in the “on-road” category, and allother locations were included in the “off-road” category in order toassess high-level on-/off- road attention distribution over the studiedepochs. Left and right forward locations were included in the on-roadcategory as these locations contain information used for detecting ha-zards peripheral to forward view (Lamble et al., 1999; Recarte andNunes, 2000).

MSGD was computed for crash and near-crash epochs, and wasfurther subdivided by off-road and on-road locations (see Table 4 andFig. 1).

The effects of glance location (on-road vs. off-road) and epoch-type(crash vs. near-crash) on MSGD were evaluated using a 2-way ANOVA,with epoch-type as a between-subjects factor, and glance location andthe interaction between epoch-type and glance location using a re-peated measures error term. The main effect of epoch-type was sig-nificant, F(1,75) = 4.65, p < 0.05, partial η2 = 0.02; crash epochshad significantly shorter MSGD than near-crash epochs overall. Themain repeated-measures effect of glance location was also significant, F(1,76) = 111.82, p < 0.0001, partial η2 = 0.43; on-road glances hadsignificantly longer MSGD than off-road glances. The interaction be-tween epoch-type and glance location was significant, F(1,76) = 4.47,p < 0.05, partial η2 = 0.02; notably, the difference between near-crash and crash MSGD was especially pronounced for on-road locations,with on-road glances being, on average, 3.91 s longer for near-crashdrivers than crash drivers. In other words, near-crash epochs were as-sociated with longer glances out the windshield − to the roadway −than crash epochs, although both made similarly short glances awayfrom the roadway.

It is important to note that the maximum length of any glance islimited by the window over which the glance data were coded by VTTI(and, hence, available for analysis in this study). Because drivers innear-crash and crash epochs appear to be making long glances on-road,which are interrupted by shorter glances off-road, we chose to evaluatethe effect of this as a function of when in the pre-precipitating eventepoch the glance was initiated. This was accomplished by “binning”glances as a function of how far from the onset of the precipitating

Table 2Number of Epochs Remaining After Each Stage of Filtering, In Aggregate and By Bin.

Number and Percentage of Epochs with Glances in Bin (Seconds Before Precipitating Event)

Stage of Filtering C/NCa 30-25s 25-20s 20-15s 15-10s 10-5s 5-0s All

N % N % N % N % N % N %1. All Event Epochs C 30 54.5 33 60.0 41 74.5 45 81.8 46 83.6 55 100 55

NC 480 66.3 586 80.9 676 93.4 710 98.1 718 99.2 724 100 7242. Filtered Aggressive Driving C 30 54.5 33 60.0 41 74.5 45 81.8 46 83.6 55 100 55

NC 479 66.3 585 80.9 675 93.4 709 98.1 717 99.2 723 100 7233. Filtered Parking Lot Epochs C 29 55.8 32 61.5 40 76.9 44 84.6 45 86.5 52 100 52

NC 469 66.1 574 81.0 662 93.4 696 98.2 704 99.3 709 100 7094. Filtered Drowsiness C 22 56.4 24 61.5 29 74.4 31 79.5 32 82.1 39 100 39

NC 396 65.0 488 80.1 565 92.8 596 97.9 604 99.2 609 100 6095. Filtered Epochs With Missing Video:

Final CountsC 20 54.1 22 59.5 27 73.0 29 78.4 30 81.1 37 100 37NC 368 64.9 455 80.2 526 92.8 556 98.1 564 99.5 567 100 567

a ‘NC’ indicates near-crash epochs, and “C” indicates crash epochs.

Table 3Glance Location Categorization.

Glancecategory

Glance locations included within category

On-road forward; left forward; right forwardOff-road center stack; instrument cluster; interior object; cell phone; left

mirror; left window; rearview mirror; right mirror; rightwindow; passenger; eyes closed

Table 4MSGD by Epoch-Type and Glance Location. Cell entries are MSGD (SEM) in seconds.

SCE Type

Glance Location Crash Near-CrashAll Locations 3.51 (0.61) 5.42 (0.61)On-Road 5.98 (0.80) 9.89 (0.90)Off-Road 1.03 (0.29) 0.95 (0.10)

B.D. Seppelt et al. Accident Analysis and Prevention 107 (2017) 48–62

53

event the glance was initiated.

3.2. The temporal effects of glance behavior: binned glance statistics

To evaluate the effect of time on MSGD for on-road glances, wedivided glances into three bins, using the temporal distance betweenthe onset of a glance and the onset of the precipitating event of eachepoch. For example, a glance initiated within a window between 20 sprior to the precipitating event and 25 s prior to the precipitating event

was placed into a “20-25 s” window, and its associated glance durationwas compiled into that window’s MSGD statistics. These were com-puted for each bin, for on-road glances (i.e., glances falling withinwindshield glance locations, namely forward, left forward, and rightforward), for the filtered set of epochs (see Table 2 for the number ofepochs with data in each bin for the filtered dataset).

As hypothesized, the greatest nominal difference in MSGD for on-road glances is at the earliest bin (20–25 s upstream from the pre-cipitating event), with near-crashes being associated with 12.39 s

Fig. 1. Mean single glance duration (MSGD) as a function of glancelocation and SCE type. Error bars represent 95% confidence intervals.

Fig. 2. Mean single glance duration (MSGD) as a function of time from event onset and epoch-type for on-road glances (left panel) and off-road glances (right panel). Error bars represent95% confidence intervals.

B.D. Seppelt et al. Accident Analysis and Prevention 107 (2017) 48–62

54

(SD = 8.02s) on-road glances, and crashes being associated with 9.68 s(SD = 5.08s) on-road glances. Within the 20–25 s bin, a mixed-designANOVA revealed a significant main effect of glance location, F(1,76)= 136.50, p < 0.0001, partial η2 = 0.50, but only a marginal effect ofSCE type, F(1,75) = 2.94, p = 0.09, η2 = 0.01, and no interaction(p = 0.41). Aggregated over SCE type, an ANOVA comparing bin andglance location found a significant main effect of bin, F(2,152)= 62.86, p < 0.0001, η2 = 0.15, of glance location, F(1,76)= 291.80, p < 0.0001, η2 = 0.28, and a significant interaction effect,F(2,152) = 53.54, p < 0.0001, η2 = 0.13. In decomposing the inter-action effect, there was a significant main effect of bin for on-roadglances, F(2,152) = 60.07, p < 0.0001, η2 = 0.37, but a non-sig-nificant effect of bin for off-road glances, F(2,152) = 2.20, p = 0.12,η2 = 0.01, indicating that it is the longer upstream length of on-roadglances (as opposed to an increase length of off-road glance nearer tothe precipitating event) that drive differences in SCE outcomes seen atthe summed epoch level (Fig. 1).

While Fig. 2 depicts a clear trend, in which the greatest differencesbetween near-crash and crash epochs occur as a result of earlier longon-road glances being ended earlier in crash epochs than in near-crashepochs, the use of MSGD to statistically evaluate this trend was un-successful− due principally to the great deal of variability in the MSGDvalues, and the small number of crashes available. Other metrics wereanalyzed to reveal the attentional mechanisms behind this trend. If longglances made earlier in pre-precipitating event epochs ended earlier incrash epochs, as the MSGD analysis indicated, then other glance me-trics, such as number of glances or total glance time, are expected toreveal significant differences in subsequent bins.

3.3. Glance count

A count of glances within any given bin was computed following theprocedure used by Krause et al. (2015): glances that were initiatedwithin a bin incremented that bin’s glance count by 1; glances thatsubtended later bins incremented those bins’ glance counts by a fractionrepresenting the proportion of the bin subtended. For example, a glancethat was initiated in the 20–25 s bin incremented that bin by 1; if it thensubtended eight of the 10 s of the 10–20 s bin (i.e., it ended at 12 s) itwould increment the 10–20 s bin by 8/10, or 0.80. This method ensuredthat long glances initiated in earlier bins were reflected in subsequentbins, and that every bin for every epoch for which there was availabledata would get a glance count of at least 1. To account for the differencein the number of epochs available for each crash and near-crash driver,glance counts were averaged within bins across all epochs for a driverrather than summed.

Fig. 3, right panel, depicts the count of off-road glances for each binfor near-crash and crash epochs using the filtered set of epochs. Whilenone of the observed differences are statistically significant, the sourceof shorter on-road MSGD for crash drivers can be reasonably identified:more off-road glances, especially pronounced in 10–20 s and 0–10 sbins. The number of glances on-road (Fig. 3, left panel) appears to alsobe greater for crashes in the 0–10 s bin, suggesting that it is moreglancing, not long off-road glancing, that is associated with driverscrashing. Critically, this trend appears, however subtly, in the earliestbin (20–25 s before the onset of the precipitating event linked tocrashing).

Additionally, the data suggests that crashes were associated with agreater standard deviation of glance counts (1.55 for off-road and 1.31for on-road) than near-crashes (0.99 for off-road and 1.04 for on-road).These values were computed across drivers, and while not statisticallytested, supports the trend evident in the confidence intervals in Fig. 3:of greater variability in glancing behavior for epochs ending in crashesthan in near-crashes.

3.4. Total glance time

Total glance time (on- and off-road) was computed as the totalamount of time within a bin that glances were coded as either on-road(i.e., within the set of forward, left forward, and right forward glancelocations) or off-road (all other locations). Because not all bins were thesame size (namely, the 20–25 s bin was half the size of the others) weconverted raw times into percentages. Also, because every frame iscoded as either on-road or off-road, total off-road glance times are di-rectly dependent upon on-road glance times; consequently, only sta-tistics for on-road glance times are reported.

Across all three bins, crash and near-crash epochs were associatedwith similar total glance time on-road (86.45% and 88.60%, respec-tively), F(1,75) = 0.46, p = 0.50. However, there was a main effect ofbin, F(2,150) = 3.61, p < 0.05, partial η2 = 0.02, with bins closer tothe precipitating event having less total on-road glance time, and aninteraction between bin and type of epoch, F(2,150) = 3.99, p < 0.05,partial η2 = 0.02. As is shown in Fig. 4, the downward slope of therelationship between time until precipitating event and total on-roadglance time for crash epochs drives this interaction effect. This con-tinues to support the hypothesis that increased glancing in approach tothe onset of a precipitating event is linked to both the large decrease inmean on-road glance duration and to the failure to anticipate a pre-cipitating event in crashes as opposed to near-crashes.

3.5. Distribution of glance across on-road regions

The cumulative findings of MSGD, glance count, and total glancetime point to the importance of on-road glance length and the threadingof these glances in between off-road glances. To examine in more depththe role of on-road glance behavior and its effects on downstreamsafety-critical outcome, sampling to peripheral road regions were se-parately considered from the central road region in a secondary analysisof mean single glance duration. In particular, this analysis aimed touncover if drivers were differentially accessing information from per-ipheral regions (those areas of the driving scene cited as relevant forhazard and event detection) in near-crash compared to crash situations.

Glance durations were examined for left forward, center forward,and right forward locations. As is evident in Fig. 5, there was a tendencyin the peripheral locations (i.e., left forward and right forward) forMSGD to be longer in near-crash epochs than crash epochs, regardlessof bin. Right forward glances were significantly longer for near-crashepochs than crash epochs within the 10–20 s bin, F(1,75) = 4.26,p < 0.05, η2 = 0.05. While differences between near-crash and crashMSGD were in the same direction for other bins for both peripherallocations, differences were not significant.

A subsequent analysis of the probability of glance within each binindicates the differences in MSGD are likely due to an increased prob-ability for drivers to initiate glances to peripheral road locations innear-crash situations; these drivers were significantly more likely toinitiate left forward glances than those in crash situations 10–20 s be-fore a precipitating event, F(1,27) = 6.52, p < 0.05, η2 = 0.08 andmarginally more likely 0–10 s before a crash, F(1,75) = 3.43, p = 0.07,η2 = 0.04. Near-crash drivers were also marginally more likely to in-itiate right forward glances 20–25 s before a precipitating event, F(75,1) = 3.82, p = 0.054, η2 = 0.05, 10–20 s before a precipitatingevent, F(75,1) = 3.77, p = 0.06, η2 = 0.05, and significantly morelikely 0–10 s before a precipitating event, F(75,1) = 4.67, p < 0.05,η2 = 0.06, than those involved in crashes. Differences between centerforward MSGD and between near-crash and crash epochs were notsignificant, although glances initiated to the center forward region20–25 s before the precipitating event were 3 s longer for near-crashthan crash epochs (MSGD = 12.68 s vs. 9.67s, respectively).

To evaluate reported differences in glance behavior between near-crashes and crashes at a higher temporal resolution than in 10 s periods,we evaluated glance behavior using the AttenD algorithm, which

B.D. Seppelt et al. Accident Analysis and Prevention 107 (2017) 48–62

55

produces a continuous measure indicative of how attention is allocatedover time; notably, considering the effect of on-road and off-roadglances in either building or depleting, respectively, a reserve of in-formation about the roadway.

3.6. The AttenD or “Buffer” algorithm

The AttenD algorithm was used to annotate each frame of glancedata with an associated “buffer” value, based on the glance history ofeach epoch. Fig. 5 depicts aggregated buffer values for near-crash andcrash epochs, for each frame across the possible 25 s of pre-pre-cipitating event glance data for the filtered set of epochs. Because of thebuffer design, where the buffer is initialized as “full” (i.e., having ascore of 2.0) and declines as off-road glances are made, both the near-crash and crash buffer lines begin at 2.0 at the 25 s marker. For thecrash epochs, it drops significantly lower, and, in aggregate, fluctuates

markedly until terminating at a value roughly one tenth of a buffervalue lower than near-crash buffers. For the near-crash epochs, itfluctuates, in aggregate, more moderately above a value of 1.9 until thetime of the precipitating event (Fig. 6). These patterns suggest thebuffer is sensitive to the effects of glancing that were identified in theother, more typical reported glance metrics. Further, they highlight thebuffer’s ability to reveal cumulative effects of making more/longer on-road glances to the road over time. On-road glances act to refill thebuffer, i.e., to restore a driver’s awareness of scene information criticalto hazard anticipation and response.

As with the other measures, buffer values were broken down by bin(Fig. 7). Overall, the near-crash epochs are associated with higherbuffer values (1.93 vs. 1.86, respectively), F(1,75) = 4.25, p < 0.05,partial η2 = 0.02. Additionally, there was a marginally significant maineffect of bin, F(2,150) = 3.02, p= 0.052, partial η2 = 0.02, and asignificant interaction between bin and epoch type, F(2,150 = 4.69,

Fig. 3. Mean count of off-road glances (left panel) and on-road glances (right panel) as a function of bin and epoch-type. Error bars represent 95% confidence intervals.

Fig. 4. Mean total on-road glance time (left panel) and off-road glance time (right panel) as a function of bin and epoch-type, using the filtered set of epochs. Error bars represent 95%confidence intervals.

B.D. Seppelt et al. Accident Analysis and Prevention 107 (2017) 48–62

56

p < 0.05, partial η2 = 0.03. As with total on-road glance time, thebuffer value associated with crash epochs drops as the bins get tem-porally closer to the precipitating event, with a maximal difference atthe 10–0 s bin, immediately before the precipitating event. Ad-ditionally, the confidence intervals around the buffer mean values growsubstantially for the crash epochs, in support of the reasoning from theglance count analysis of an increased rate of glancing as the

precipitating event draws nearer. An analysis of the average standarddeviation of buffer value (Fig. 6, right panel) supports this reasoning.Overall, the main effect of epoch type was significant for buffer stan-dard deviation, F(1,75) = 4.37, p < 0.05, partial η2 = 0.03, as wasthe main effect of bin, F(2,150) = 4.30, p < 0.05, partial η2 = 0.03,and the interaction between the two was marginally significant, F(2,150) = 2.92, p = 0.06, partial η2 = 0.02. Crash epochs have a

Fig. 5. Mean single glance duration (MSGD) as a function of time from event onset and epoch-type for left forward (on-road) glances (left panel), center forward (on-road) glances (centerpanel), and right forward (on-road) glances (right panel). Error bars represent 95% confidence intervals. Note: The scale for the plot of center forward glances is different from the plots ofleft and right forward glances due to the large difference in glance durations to these regions.

Fig. 6. Mean buffer value as a function of time until precipitatingevent and epoch-type, using the filtered set of epochs.

B.D. Seppelt et al. Accident Analysis and Prevention 107 (2017) 48–62

57

higher standard deviation of buffer value, with the two bins closest tothe precipitating event producing the highest deviations.

3.7. Baseline vs. near-crash vs. crash

Because AttenD buffer values appeared to maximally differ betweencrash and near-crash epochs in the moment before the precipitatingevent, we also looked at the short baseline epochs available from the100-car study to evaluate whether the AttenD algorithm could alsodifferentiate SCE epochs from baseline epochs during this short timewindow.

There were 4935 baseline glance-coded epochs of 6.1 s available. Asin previous analyses, we applied a filter to the epochs, removing 246epochs that contained missing video or parking lot driving (because thebaseline epochs were not scored for aggressive driving or drowsinesswithin the available online dataset, filters for these behaviors were notapplied). Buffer values were averaged across the 6.1 s of available data

for each epoch in the three conditions (baseline, near-crash, and crashepochs), and summary values are plotted in Fig. 8. Overall, buffer valuesignificantly varied by epoch type, F(2,5693) = 17.24, p < 0.0001,η2 < 0.01, with baseline having a significantly higher buffer valuethan near-crashes (p < 0.0001, Holm-Bonferroni-corrected) and cra-shes (p < 0.05, Holm-Bonferroni-corrected). Crashes and near-crashmean buffer values were statistically equivalent (p= 0.19). Standarddeviation of buffer value also significantly varied by epoch type, F(2,5693) = 8.95, p < 0.001, η2 < 0.01, with baseline having sig-nificantly lower buffer standard deviations than near-crashes(p < 0.001, Holm-Bonferroni-corrected) and crashes (p < 0.05,Holm-Bonferroni-corrected). Crash and near-crash buffer standard de-viations were also statistically equivalent (p = 0.18).

4. Discussion

This analysis revealed that naturalistic driving epochs ending in

Fig. 7. Mean buffer value (left panel) and mean standard deviation of buffer value (right panel) as a function of epoch type and bin. Error bars represent 95% confidence intervals.

Fig. 8. Mean buffer value (left panel) and standard deviation of buffer value (right panel) as a function of epoch type. Error bars represent 95% confidence intervals.

B.D. Seppelt et al. Accident Analysis and Prevention 107 (2017) 48–62

58

crash vs. near-crash differ in critically important ways that suggest newinsights into driver behavior and crash avoidance. In a near-crash, thedriver may successfully avoid a crash by more effectively balancingattention on and off the road at earlier points in the driving sequenceprior to the occurrence of a precipitating event, in particular byspending more time during individual glances looking at the road.When shifting their gaze off-road, drivers in near-crash situations do soat a more consistent duration and rate. This pattern − in which forwardattention to the road is interleaved with off-road glances, and for longeramounts of time to encode critical information from the forwardroadway, may equip drivers with the information they need to avoid acrash. This distinct pattern of glancing at the road and of attending, asrevealed from the AttenD buffer plot (Fig. 6), appears as a significantdifferentiation between near-crash and crash epochs. These findings, ifreplicated through other work, have the potential to change the para-digm used for thinking about crash risk and crash avoidance.

4.1. Effects of common individual glance metrics

Common glance metrics of mean single glance duration, glancecount, and total glance time across the available epoch lengths point tothe significance of longer on-road glances and more frequent glancesbetween on-road and off-road locations in differentiating near-crashfrom crash outcomes. Looking at glance patterns across the 25 s of pre-precipitating data examined in the 10 s bin analysis, there are cleartrends indicating that the period of time 10–25 s in advance of theprecipitating event plays a critical role − a time period, which has,until now, been largely overlooked in the analysis of SCEs. Driversspend less time looking at the road as they move closer to the pre-cipitating event − an effect that was more pronounced for crashes thannear crashes, as evident from mean glance duration and total glancetime results. Concomitant with reduced on-road glance times, driversengage in increased glancing between on- and off-road locations, withincreased variability in this switching for crash epochs − an effect thatalludes to a destabilized allocation of attentional resources (Hancockand Warm, 1989). Finally, as evident from the breakout of central andperipheral road regions, drivers neglect information in the periphery inthe moments prior to crashes (as far back as 25s) to a larger extent thanin near-crash situations.

4.2. Glance patterns over time

The findings of this work are in line with the proposition that safety-related behavior lies in the pattern of glances on and off the road.Simple single metrics (such as mean single glance durations to one lo-cation or another, or total glance time off-road) only partially capturethis pattern. The important finding from this work is that it is howglances on- and off-road are interleaved that supports drivers in de-veloping adequate information about the situation that is vital. Thistype of pattern may be best captured by a hybrid measure.

Viewed from the output of the AttenD algorithm in Fig. 6, driversengaged in shorter mean on-road glances for crashes compared to near-crashes in the 25 s prior to the precipitating event, but, in aggregate,starting in the 20–10 s period prior to this event initiated comparativelymore frequent glances on-road and off-road, together producing thetemporary spike centered around 10s. The precipitous drop in buffervalue following this period for crashes is reflected in the longer totalglance time spent off-road (in contrast to near-crashes) in the 10–0 sprior to the precipitating event (see Fig. 4, right panel).

When considering the temporal effects of glance behavior across thefull 25 s of data for both near-crashes and crashes, it is evident thatdrivers generally demonstrated more attenuated attention across time.The lower AttenD buffer values in the 6.1 s epoch analysis for SCEscompared to the baseline epochs confirm this reduced attention to theforward roadway from that seen during normal (non-safety-critical)periods of driving.

4.3. A glass half-full perspective indicates a need to consider on-road glancebehavior in analyses of SCEs

Overall, drivers in the safety-critical situations showed reducedglance duration to on-road locations and an increase in glances betweenon- and off-road locations in the approach to the precipitating event.This glance pattern is evident to a larger extent in the lower, morevariable buffer profile in the seconds leading up to crashes compared tonear-crashes. A downward sloping profile signals degrading awarenessof the roadway environment.

Ample on-road glance time is needed to form a robust situationawareness and to trigger relevant schemata in order to accuratelyforecast the probabilities, rate, and consequences concomitant to un-folding event dynamics (Horswell and McKenna, 2004). It is in theperception of scene information relevant to task goals that drivers areable to prime, through top-down mechanisms, the attentional resourcesneeded to facilitate fast reactions to looming cues (Engström et al.,2013). Reduced on-road glance durations and decreased sampling ofperipheral road regions for drivers in the crash epochs in the momentsleading up to the precipitating event likely left these drivers ill-equipped in their decision for when to time an off-road glance. Analysesreported in this paper suggest that just a few more seconds of on-roadglance time attending to cues in the driving environment that foretell ofhazard potential can indeed make the difference as to whether a drivercrashes or averts a crash. In total, glance behaviors that differentiate the100-car SCEs can be interpreted as supporting the view that ill-timedoff-road glances in the moments before a crash are in fact an outcome ofpoorly-formed expectations of the driving environment (i.e., depletedsituation models) consequent from under-sampling of predictive cues.Or, stated another way, drivers in crash epochs, to a larger extent thanthose in near-crash epochs, experienced breakdowns in their proactivebarrier (Engström et al., 2013).

4.4. Role of non-driving related task activities on glance behaviors in the100-Car event data

Increasingly, the presence of driver assistance systems and otherhigher-level automated systems may amplify the need to consider at-tention management and to adopt metrics such as AttenD over moretraditional glance metrics. The need to understand and support atten-tion management when drivers are engaged in non-driving related tasksis particularly crucial to the extent they have a degraded awareness ofhow distraction affects their driving performance and/or do not stra-tegically plan when to initiate non-driving related tasks (e.g., Horreyet al., 2008; Lerner et al., 2008; Perez et al., 2011). Initial designs ofsmart systems that encourage on-road glances have shown benefits tohow drivers allocate their visual resources and counter distraction ef-fects (e.g., Birrell and Fowkes, 2014).

There are a wide variety of activities drivers engaged in prior to theSCEs in the 100-car event database (e.g., Klauer et al., 2006, p. 24;Owens et al., 2015). For SCEs that involve non-driving related tasks,inattention, or high workload as a factor, divergence of attention fromthe road in pre-crash/near-crash period is expected. To fully examinethe attentional mechanisms at play, the preexisting categories of in-attention types, as defined within Klauer et al. (2006) − secondarytasks, driving-related inattention, non-specific eye glances, and drowsiness− could be examined between the near-crash and crash epochs to de-termine their effects on the observed glance patterns. While thesebreakdowns of epochs into inattention categories have proved mean-ingful in the past for calculating odds ratios associated with the risks ofNC/C outcomes (Klauer et al., 2006), for this work, there is an in-sufficient number of crashes per inattention type to perform statisticalanalyses in comparison with the near-crash events. Due to the relativelylow number of crashes compared to near-crashes, more in-depth ana-lyses on the role of secondary task engagement on glance behaviors thatdistinguish between safety-critical outcomes is an important area of

B.D. Seppelt et al. Accident Analysis and Prevention 107 (2017) 48–62

59

future work; the NEST dataset, which was recently made publicly ac-cessible, provides opportunities for such efforts (Seaman et al., 2017).

5. Conclusion

Taken together with recent findings (e.g., Samuel and Fisher, 2015),this work may have important implications for attention managementin the context of the increasing prevalence of in-vehicle demands andan increasing level of vehicle automation. Current human-machine in-terface (HMI) evaluation guidelines largely focus on minimizing off-road glance durations. While this perspective is supported by the datain the seconds leading up to a crash, a broader “glass half full” per-spective that looks further in advance of a precipitating event suggeststhat time on-road is potentially an equally, if not more important,metric deserving deeper consideration. Evident from this analysis of the100-Car NDS data, drivers better protect their ability to anticipate ha-zards if they more effectively acquire information from their roadwaysurroundings. Appropriately interspaced and sufficiently sustainedscanning of the road mitigates the loss of awareness of the roadwayenvironment that occurs during glances away from the road. In situa-tions where a driver does not allocate enough attention to the roadbetween off-road glances, effective recovery of an understanding of thedriving situation may be flawed since subsequent glances will likely bebased on incomplete situation knowledge, and perhaps timed in-appropriately as well as located inappropriately. Hybrid measures suchas the AttenD algorithm that consider the combined effects of on- andoff-road glances in a temporal relationship offer a unique perspectiveinto a driver’s capacity to develop and maintain an understanding of thedriving environment that bears on their capacity to appropriately re-spond to emergent events. The implication of this is that a hybridmeasure like the AttenD algorithm can be interpreted as indicating the‘amount of situation knowledge’ that a driver maintains and updates (aprecursor for situation awareness) while driving.

In its use of a set of rules that determine how individual parametersproduce changes in a buffer value to reflect a driver's management ofattention over time, the AttenD algorithm affords building in furtherparameters and rule conditions to account for complex and interactingeffects from secondary task load. Recent work applied a modified formof the AttenD algorithm, which built in additional rules for the temporalthreading of central and peripheral glances to the forward roadway, todifferentiate between periods of driver engagement with tasks ofvarying cognitive load (Seppelt et al., 2017). Recent application of theAttenD algorithm to alternative sources of naturalistic data (NEST da-taset; cf. Owens et al., 2015) demonstrate replicability for findings re-ported in this paper, and began to apportion the role of an attributablepresence and type of secondary task load and their differing patterns ofon- and off-road glance behavior to safety-relevant outcomes (Seamanet al., 2017). For future work, it would be desirable to have longerepochs of glance data for a wider range of SCEs that reflect vehicleswith more modern, advanced features, including multi-modal interfacecharacteristics − such as voice input, smartphone integration, etc.,than have been available in datasets to date. Work may also be neededto consider the relationships between crash, near-crash, and baselineepochs of lower-risk drivers (notably, the 100-Car dataset characterizesthe behavior profile of high-risk drivers and must therefore be inter-preted with caution in how observed effects extrapolate to low-riskdrivers; cf. Klauer et al., 2009). The consideration of attentional stra-tegies in lower-risk drivers (e.g. those involved in fewer SCEs per year)may lead to an increased understanding of “successful” avoidance ofthreats offered by a precipitating event, thereby enhancing the “glasshalf-full” perspective.

Finally, work by our group is ongoing in investigating and refiningmore advanced methods for combining information gained from on-and off-road glances over time as well as other supporting metrics ofcognition into an “attention buffer” that aims, in part, to provide agreater degree of sensitivity to predict real-time failures in attention

allocation that impact situation awareness and crash risk. It is hy-pothesized that a more optimized model can lead to an in-vehicle in-terface assessment methodology that builds-upon and advances thestate of practice from current methods (e.g., NHTSA, 2016, Alliance,European Statement of Principles, JAMA, etc.) to one that considers theallocation of attention over time for task interactions interleaved withdriving. Such an assessment perspective may offer a predictive safetyvalidity not found in earlier distraction-based methods, and shifts thefocus to positive design goals. Implemented in the context of a real-timedriver state sensing system the assessment model can be applied to theactive management of driver attention in the context of environmentalcharacteristics, vehicle automation, and other inter-related factors.Theapproaches, i

Acknowledgements

Support for the this work was provided by the Advanced HumanFactors Evaluator for Automotive Demand (AHEAD) Consortium, theUS DOT’s Region I New England University Transportation Center atMIT, and the Toyota Class Action Settlement Safety Research andEducation Program. The views and conclusions being expressed arethose of the authors, and have not been sponsored, approved, or en-dorsed by Toyota or plaintiffs’ class counsel.

References

Ahlström, C., Kircher, K., Kircher, A., 2009. Considerations when calculating percent roadcentre from eye movement data in driver distraction monitoring. Proceedings of theFifth International Driving Symposium on Human Factors in Driver Assessment,Training and Vehicle Design 132–139.

Ahlström, C., Kircher, K., Kircher, A., 2013. A gaze-based driver distraction warningsystem and its effect on visual behavior. IEEE Trans. Intell. Transp. Syst. 14 (2),965–973.

Angell, L.S., Auflick, J., Austria, P.A., Kochhar, D.S., Tijerina, L., Biever, W., et al., 2006.Driver Workload Metrics Task 2 Final Report (No. HS-810 635).

Angell, L.S., 2007. Effects of secondary task demands on drivers’ responses to eventsduring driving: surrogate methods and issues. Driving Assessment 2007: 4thInternational Driving Symposium on Human Factors in Driver Assessment, Training,and Vehicle Design.

Anstis, S.M., 1974. A chart demonstrating variations in acuity with retinal position. VisionRes. 14 (7), 589–592.

Bellenkes, A.H., Wickens, C.D., Kramer, A.F., 1997. Visual scanning and pilot expertise:the role of attentional flexibility and mental model development. Aviat. SpaceEnviron. Med. 48 (7), 569–579.

Biederman, I., Mezzanotte, R.J., Rabinowitz, J.C., 1982. Scene perception: detecting andjudging objects undergoing relational violations. Cognit. Psychol. 14 (2), 143–177.

Birrell, S.A., Fowkes, M., 2014. Glance behaviours when using an in-vehicle smart drivingaid: a real-world: on-road driving study. Transp. Res. Part F: Traffic Psychol. Behav.22, 113–125.

Chun, M.M., Jiang, Y., 1998. Contextual cueing: implicit learning and memory of visualcontext guides spatial attention. Cognit. Psychol. 36 (1), 28–71.

Crisler, M., Brooks, J., Ogle, J., Guirl, C., Alluri, P., Dixon, K., 2008. Effect of wirelesscommunication and entertainment devices on simulated driving performance.Transp. Res. Rec.: J. Transp. Res. Board 2069, 48–54.

Crundall, D.E., Underwood, G., 1998. Effects of experience and processing demands onvisual information acquisition in drivers. Ergonomics 41 (4), 448–458.

Dingus, T.A., Antin, J.F., Hulse, M.C., Wierwille, W.W., 1989. Attentional demand re-quirements of a moving map navigation system. Transp. Res. A 23 (4), 301–315.

Dingus, T.A., Klauer, S.G., Neale, V.L., Petersen, A., Lee, S.E., Sudweeks, J.D., et al., 2006.The 100-car Naturalistic Driving Study, Phase II-results of the 100-car FieldExperiment (No. HS-810 593).

Dong, Y., Hu, Z., Uchimura, K., Murayama, N., 2011. Driver inattention monitoringsystem for intelligent vehicles: a review. IEEE Trans. Intell. Transp. Syst. 12 (2),596–614.

Donmez, B., Boyle, L.N., Lee, J.D., 2007. Safety implications of providing real-timefeedback to distracted drivers. Accid. Anal. Prev. 39 (3), 581–590.

Donmez, B., Boyle, L.N., Lee, J.D., 2008. Mitigating driver distraction with retrospectiveand concurrent feedback. Accid. Anal. Prev. 40 (2), 776–786.

Drews, F.A., Yazdani, H., Godfrey, C.N., Cooper, J.M., Strayer, D.L., 2009. Text messagingduring simulated driving. Hum. Factors: J. Hum. Factors Ergon. Soc. 51, 762–770.

Endsley, M.R., 1995. Toward a theory of situation awareness in dynamic systems. Hum.Factors: J. Hum. Factors Ergon. Soc. 37 (1), 32–64.

Engström, J., Victor, T., Markkula, G., 2013. Attention selection and multitasking ineveryday driving: a conceptual model. In: In: Regan, M., Lee, J.D., Victor, T.W. (Eds.),Driver Distraction and Inattention: Advances in Research and Countermeasures, vol.1 Ashgate Publishing (pp. 27–54).

Falkmer, T., Gregersen, N.P., 2005. A comparison of eye movement behavior of in-experienced and experienced drivers in real traffic environments. Optom. Vis. Sci. 82

B.D. Seppelt et al. Accident Analysis and Prevention 107 (2017) 48–62

60

(8), 732–739.Fitch, G.M., Lee, S.E., Klauer, S., Hankey, J., Sudweeks, J., Dingus, T., 2009. Analysis of

Lane-Change Crashes and Near-Crashes. US Department of Transportation, NationalHighway Traffic Safety Administration.

Fletcher, L., Zelinsky, A., 2007. Driver state monitoring to mitigate distraction.Proceedings of the Internal Conference on the Distractions in Driving (pp. 487–523).

Fletcher, L., Zelinsky, A., 2009. Driver inattention detection based on eye gaze—Roadevent correlation. Int. J. Rob. Res. 28 (6), 774–801.

Friedman, A., 1979. Frame theory, comprehension: and picture memory. J. Exp. Psychol.Gen. 108, 316–355.

Glaser, Y.G., Llaneras, R.E., Glaser, D.S., Green, C.A., 2016. Relationship Between DriverEyes-Off-Road Interval and Hazard Detection Performance Under Automated Driving(No. 2016-01-1424). (SAE Technical Paper).

Green, M., 2000. How long does it take to stop? methodological analysis of driver per-ception-Brake times. Transp. Hum. Factors 2 (3), 195–216.

Greene, M.R., Oliva, A., 2009. The briefest of glances the time course of natural sceneunderstanding. Psychol. Sci. 20 (4), 464–472.

Hafed, Z.M., Clark, J.J., 2002. Microsaccades as an overt measure of covert attentionshifts. Vision Res. 42 (22), 2533–2545.

Hallmark, S., McGehee, D., Bauer, K.M., Hutton, J.M., Davis, G.A., Hourdos, J., et al.,2013. Initial Analyses from the SHRP 2 Naturalistic Driving Study: Addressing DriverPerformance and Behavior in Traffic Safety (No. SHRP 2 Safety Project S08).

Hancock, P.A., Warm, J., 1989. A dynamic model of stress and sustained attention. Hum.Factors: J. Hum. Factors Ergon. Soc. 31 (5), 519–537.

Henderson, S.G., 2003. Input modeling: input model uncertainty: why do we care andwhat should we do about it? Proceedings of the 35th Conference on WinterSimulation: Driving Innovation (pp. 90–100). Winter Simulation Conference.

Hickman, J.S., Hanowski, R.J., 2012. An assessment of commercial motor vehicle driverdistraction using naturalistic driving data. Traffic Inj. Prev. 13 (6), 612–619.

Hickman, J.S., Hanowski, R.J., Bocanegra, J., 2010. Distraction in Commercial Trucksand Buses: Assessing Prevalence and Risk in Conjunction with Crashes and Near-crashes (No. FMCSA-RRR-09-042).

Horrey, W.J., Wickens, C.D., 2004. Driving and side task performance: the effects ofdisplay clutter, separation, and modality. Hum. Factors: J. Hum. Factors Ergon. Soc.46 (4), 611–624.

Horrey, W.J., Lesch, M.F., Garabet, A., 2008. Assessing the awareness of performancedecrements in distracted drivers. Accid. Anal. Prev. 40 (2), 675–682.

Horswill, M.S., McKenna, F.P., 2004. Drivers' hazard perception ability: situationawareness on the road. In: Banbury, S., Tremblay, S. (Eds.), A Cognitive Approach toSituation Awareness: Theory and Application, 1st ed. Ashgate Publishing, Ltd., UK(pp. 155–175).

Hosking, S.G., Young, K.L., Regan, M.A., 2009. The effects of text messaging on youngdrivers. Hum. Factors: J. Hum. Factors Ergon. Soc. 51 (4), 582–592.

Hunt, A.R., Kingstone, A., 2003. Covert and overt voluntary attention: linked or in-dependent? Cognit. Brain Res. 18 (1), 102–105.

Kircher, K., Ahlström, C., 2009. Issues related to the driver distraction detection algo-rithm AttenD. In: First International Conference on Driver Distraction andInattention. Gothenburg, Sweden.

Kircher, K., Ahlström, C., 2012. The impact of tunnel design and lighting on the perfor-mance of attentive and visually distracted drivers. Accid. Anal. Prev. 47, 153–161.

Klauer, S.G., Dingus, T.A., Neale, V.L., Sudweeks, J.D., Ramsey, D.J., 2006. The Impact ofDriver Inattention on Near-Crash/Crash Risk: An Analysis Using the 100-CarNaturalistic Driving Study Data.

Klauer, S.G., Dingus, T.A., Neale, V.L., Sudweeks, J.L., Ramsey, D.J., 2009. ComparingReal-World Behaviors of Drivers with High Vs. Low Rates of Crash and Near-Crashes.NHTSA, Washington, D.C (USDOT Report # DOT HS 811 091).

Klauer, S.G., Guo, F., Sudweeks, J., Dingus, T.A., 2010. An Analysis of Driver InattentionUsing a Case-Crossover Approach on 100-Car Data: Final Report (No. HS-811 334).

Knipling, R.R., 2015. Naturalistic driving events: no harm, no foul, no validity. In Drivingassessment 2015. International Symposium on Human Factors in Driver Assessment,Training, and Vehicle Design 196–202.

Krause, M., Conti, A.S., Henning, M., Seubert, C., Heinrich, C., Bengler, K., et al., 2015.App analytics: predicting the distraction potential of in-vehicle device applications.Procedia Manuf. 3, 2658–2665.

Lamble, D., Laakso, M., Summala, H., 1999. Detection thresholds in car following si-tuations and peripheral vision: implications for positioning of visually demanding in-car displays. Ergonomics 42 (6), 807–815.

Lavie, N., 1995. Perceptual load as a necessary condition for selective attention. J. Exp.Psychol. Hum. Percept. Perform. 21 (3), 451.

Lavie, N., 2005. Distracted and confused?: Selective attention under load. Trends Cogn.Sci. 9 (2), 75–82.

Lee, J.D., Hoffman, J.D., Hayes, E., 2004. Collision warning design to mitigate driverdistraction. Proceedings of the SIGCHI Conference on Human Factors in ComputingSystems (pp. 65–72). ACM.

Lee, J., Moeckli, J., Brown, T.L., Roberts, S., Victor, T., Marshall, D., et al., 2013.Detection of driver distraction using vision-based algorithms. Proceedings of the 23rdEnhanced Safety Vehicle Conference.

Lee, J., Sawyer, B.D., Mehler, B., Angell, L., Seppelt, B.D., Seaman, S., Reimer, B., 2017.Linking the detection response task and the AttenD algorithm through the assessmentof human-machine interface workload. Transportation Research Record (No. 17-06664).

Lerner, N., Singer, J., Huey, R., 2008. Driver Strategies for Engaging in Distracting TasksUsing In-Vehicle Technologies (No. HS-810 919).

Liang, Y., Lee, J.D., Yekhshatyan, L., 2012. How dangerous is looking away from theroad? Algorithms predict crash risk from glance patterns in naturalistic driving. Hum.Factors: J. Hum. Factors Ergon. Soc. 54 (6), 1104–1116.

Liang, Y., Lee, J.D., Horrey, W.J., 2014. A looming crisis: the distribution of off-roadglance duration in moments leading up to crashes/near-crashes in naturalisticdriving. Proceedings of the Human Factors and Ergonomics Society Annual Meeting58 (No. 1), 2102–2106 (Sage Publications).

Liang, Y., 2009. Detecting Driver Distraction. Unpublished PhD Dissertation. TheUniversity of Iowa, Iowa City.

Markkula, G., Engström, J., Lodin, J., Bärgman, J., Victor, T., 2016. A farewell to brakereaction times? Kinematics-dependent brake response in naturalistic rear-end emer-gencies. Accid. Anal. Prev. 95, 209–226.

McKnight, A.J., McKnight, A.S., 2003. Young novice drivers: careless or clueless? Accid.Anal. Prev. 35 (6), 921–925.

NHTSA, 2016. Early Estimate of Motor Vehicle Traffic Fatalities for the First Nine Months(Jan-Sep) of (DOT HS 812 240). U.S. Department of Transportation, NationalHighway Traffic Safety Administration, Washington D.C.

Olson, R.L., Hanowski, R.J., Hickman, J.S., Bocanegra, J.L., 2009. Driver Distraction inCommercial Vehicle Operations (No. FMCSA-RRR-09-042).

Owens, J.M., Angell, L., Hankey, J.M., Foley, J., Ebe, K., 2015. Creation of the naturalisticengagement in secondary tasks (NEST) distracted driving dataset. J. Safety Res. 5433–e29.

Perez, L.S., Angell, J., Hankey, R.K., Deering, R.E., Llaneras, C.A., Green, M.L., Neurauter,J.F., 2011. Advanced Crash Avoidance Technologies (ACAT) Program–Final Report ofthe GM-VTTI Backing Crash Countermeasures Project. No. HS-811 452.

Pohl, J., Birk, W., Westervall, L., 2007. A driver-distraction-based lane-keeping assistancesystem proceedings of the institution of mechanical engineers, part I. J. Syst. ControlEng. 221 (4), 541–552.

Posner, M.I., 1980. Orienting of attention. Q. J. Exp. Psychol. 32 (1), 3–25.Potter, M.C., 1999. Understanding Sentences and Scenes: The Role of Conceptual Short-

Term Memory Fleeting Memories: Cognition of Brief Visual Stimuli. pp. 13–46.Recarte, M.A., Nunes, L.M., 2000. Effects of verbal and spatial-imagery tasks on eye

fixations while driving. J. Exp. Psychol.: Appl. 6 (1), 31.Regan, M.A., Young, K.L., Lee, J.D., 2009. Introduction. In Driver Distraction: Theory,

Effects, and Mitigation. CRC Press, Boca Raton, FL (pp. 3–7).Regan, M.A., Lee, J.D., Victor, T.W. (Eds.), 2013. Driver Distraction and Inattention:

Advances in Research and Countermeasures, vol. 1 Ashgate Publishing, Ltd..Reimer, B., Mehler, B., Wang, Y., Coughlin, J.F., 2012. A field study on the impact of

variations in short term memory demands on drivers’ visual attention and drivingperformance across three age groups. Hum. Factors 54 (3), 454–468.

Reimer, B., Mehler, B., Reagan, I., Kidd, D., Dobres, J., 2016. Multi-modal demands of asmartphone used to place calls and enter addresses during highway driving relative totwo embedded systems. Ergonomics 59 (12), 1565–1585.

Rensink, R.A., 2002. Change detection. Annu. Rev. Psychol. 53 (1), 245–277.Rydström, A., 2007. The Effect of Haptic Feedback in Visual-Manual Human-Machine

Interaction. Luleå University of Technology, Luleå.Samuel, S., Fisher, D.L., 2015. Evaluation of the minimum forward roadway glance

duration critical to latent hazard detection. Transportation Research Board 94thAnnual Meeting (No. 15–1707).

Schyns, P.G., Oliva, A., 1994. From blobs to boundary edges: evidence for time-andspatial-scale-dependent scene recognition. Psychol. Sci. 5 (4), 195–200.

Seaman, S., Lee, J., Seppelt, B., Angell, L., Mehler, B., Reimer, B., 2017. It’s all in thetiming: using the AttenD algorithm to assess texting in the NEST naturalistic drivingdatabase. 9th International Driving Symposium on Human Factors in DriverAssessment, Training, and Vehicle Design.

Senders, J.W., Kristofferson, A.B., Levison, W.H., Dietrich, C.W., Ward, J.L., 1967. Theattentional demand of automobile driving. Highw. Res. Rec. 195.

Seppelt, B., Seaman, S., Angell, L., Mehler, B., Reimer, B., 2017. Differentiating cognitiveload using a modified version of AttenD. 9th International Conferences onAutomotive User Interfaces and Interactive Vehicular Applications.

Sodhi, M.S., Reimer, B., Llamazares, I., 2002. Glance analysis of driver eye movements toevaluate distraction behavior research methods. Instrum. Comput. 34 (4), 529–538.

Stelmach, L.B., Bourassa, C.M., Di Lollo, V., 1984. Detection of stimulus change: thehypothetical roles of visual transient responses. Percept. Psychophys. 35 (3),245–255.

Taylor, T., Masserang, K.M., Pradhan, A.K., Divekar, G., Samuel, S., Muttart, J.W.,Pollatsek, A., Fisher, D.L., 2011. Long term effects of hazard anticipation training onnovice drivers measured on the open road. In: Proceedings of Driving Assessment2011, Lake Tahoe, Nevada. Public Policy Center, University of Iowa, Iowa City.

Taylor, T., Pradhan, A.K., Divekar, G., Romoser, M., Muttart, J., Gomez, R., Fisher, D.L.,et al., 2013. The view from the road: the contribution of on-road glance-monitoringtechnologies to understanding driver behavior. Accid. Anal. Prev. 58, 175–186.

Theeuwes, J., Kramer, A.F., Hahn, S., Irwin, D.E., 1998. Our eyes do not always go wherewe want them to go: capture of the eyes by new objects. Psychol. Sci. 9 (5), 379–385.

Thorpe, S., Fize, D., Marlot, C., 1996. Speed of processing in the human visual system.Nature 381 (6582), 520–522.

Tian, R., Li, L., Chen, M., Chen, Y., Witt, G.J., 2013. Studying the effects of driver dis-traction and traffic density on the probability of crash and near-crash events innaturalistic driving environment. IEEE Trans. Intell. Transp. Syst. 14 (3), 1547–1555.

Treat, J.R., Castellan, N.J., Stansifer, R.L., Mayer, R.E., Hume, R.D., Shinar, D., Tumbas,N.S., et al., 1977. Tri-level study of the causes of traffic accidents: final Report. CausalFactor Tabulations and Assessments, vol. I.

Victor, T.W., Harbluk, J.L., Engström, J.A., 2005. Sensitivity of eye-movement measuresto in-vehicle task difficulty. Transp. Res. Part F: Traffic Psychol. Behav. 8 (2),167–190.

Victor, T., Dozza, M., Bärgman, J., Boda, C.N., Engström, J., Flannagan, C., Markkula, G.,et al., 2015. Analysis of Naturalistic Driving Study Data: Safer Glances, DriverInattention and Crash Risk (No. SHRP 2 Report S2-S08A-RW-1).

Wang, J.S., Knipling, R.R., Goodman, M.J., 1996. The role of driver inattention in crashes:

B.D. Seppelt et al. Accident Analysis and Prevention 107 (2017) 48–62

61

new statistics from the 1995 Crashworthiness Data System. 40th Annual Proceedingsof the Association for the Advancement of Automotive Medicine (Vol. 377, p. 392).

Wang, Y., Reimer, B., Dobres, J., Mehler, B., 2014. The sensitivity of different meth-odologies for characterizing drivers’ gaze concentration under increased cognitivedemand. Transp. Res. Part F: Traffic Psychol. Behav. 26, 227–237.

Wickens, C.D., Goh, J., Helleberg, J., Horrey, W.J., Talleur, D.A., 2003. Attentionalmodels of multitask pilot performance using advanced display technology. Hum.Factors: J. Hum. Factors Ergon. Soc. 45 (3), 360–380.

Wierwille, W.W., 1993b. An initial model of visual sampling of in-car displays and con-trols. In: Gale, A.G., Brown, I.D., Haslegrave, C.M., Kruysse, H.W., Taylor, S.P. (Eds.),Vision in Vehicles 4, (pp. 271–282). Amsterdam: North-Holland.

Yantis, S., Jonides, J., 1990. Abrupt visual onsets and selective attention: voluntary versusautomatic allocation. J. Exp. Psychol. Hum. Percept. Perform. 16 (1), 121.

Young, R.A., Aryal, B.J., Muresan, M., Ding, X., Oja, S., Simpson, S.N., 2005. Road-to-lab:validation of the static load test for predicting on-road driving performance whileusing advanced in-vehicle information and communication devices. In: Proceedingsof the Third International Driving Symposium on Human Factors in DriverAssessment, Training and Vehicle Design. Rockport, Maine.

Zhang, H., Smith, M.R., 2004. A Literature Review of Visual Distraction Research. SAfetyVEhicles Using Adaptive Interface Technology (SAVE-IT) –Task 7 Final Report.

Zhang, H., Smith, M.R., Witt, G.J., 2006. Identification of real-time diagnostic measuresof visual distraction with an automatic eye-tracking system. Hum. Factors: J. Hum.Factors Ergon. Soc. 48 (4), 805–821.

Bobbie D. Seppelt is research scientist at Touchstone Evaluations, Inc. Prior to this po-sition, she served as a postdoctoral research associate in the Industrial and SystemsEngineering department at the University of Wisconsin-Madison. She received her Ph.D.in Industrial Engineering in 2009 from the University of Iowa and her M.S. in EngineeringPsychology in 2003 from the University of Illinois at Urbana-Champaign.

Sean Seaman is a research scientist at Touchstone Evaluations, Inc. He is also researchfaculty in the Department of Psychiatry and Behavioral Neurosciences at the Wayne StateUniversity School of Medicine in Detroit. He received his Ph.D. in Psychology in 2010from Wayne State University.

Joonbum Lee is a postdoctoral associate at the MIT AgeLab. He received his Ph.D. inIndustrial Engineering from the University of Wisconsin-Madison in 2014. His researchinterests include driver distraction, visual attention allocation, and evaluation of in-ve-hicle human-machine interfaces.

Linda S. Angell is a Principal Scientist at Touchstone Evaluations, Inc. −and itsPresident. She holds a PhD (in Experimental/Cognitive Psychology) and has over 35 yearsof research experience, spanning both academic and applied positions.

Bryan Reimer is a Research Engineer at the Massachusetts Institute of TechnologyAgeLab and the Associate Director of the New England University Transportation Center.He is an author on over 200 technical contributions in transportation and related humanfactors areas, and is a graduate of the University of Rhode Island with a Ph.D. in Industrialand Manufacturing Engineering.

Bruce Mehler is a Research Scientist at the Massachusetts Institute of Technology AgeLaband the New England University Transportation Center. He served as Director ofApplications & Development at NeuroDyne Medical, Corp. prior to assuming his currentposition at MIT. He has an MA in Psychology from Boston University, and completedadditional studies in the neurophysiological study behavior within multiple programs,including the Marine Biological Laboratory at Woods Hole.

B.D. Seppelt et al. Accident Analysis and Prevention 107 (2017) 48–62

62