Accident Analysis and Prevention 74 (2014) 107–117

Contents lists available at ScienceDirect

Accident Analysis and Prevention journal homepage: www.elsevier.com/locate/aap

A multinomial choice model approach for dynamic driver vision transitions Shih-Hsuan Huang 1, Jinn-Tsai Wong * Department of Transportation and Logistics Management, National Chiao-Tung University, 4F, 118, Section 1, Chung Hsiao W. Road, Taipei 10044, Taiwan, ROC

A R T I C L E I N F O

A B S T R A C T

Article history: Received 1 November 2013 Received in revised form 7 October 2014 Accepted 8 October 2014 Available online xxx

Exploring the continual process of drivers allocating their attention under varying conditions could be vital for preventing motor vehicle crashes. This study aims to model visual behaviors and to estimate the effects of various contributing factors on driver’s vision transitions. A visual attention allocation framework, based on certain contributing attributes related to driving tasks and environmental conditions, has been developed. The associated logit type models for determining driver choices for focal points were successfully formulated and estimated by using naturalistic glance data from the 100-car event database. The results offer insights into driver visual behavior and patterns of visual attention allocation. The three focal points that drivers most frequently rely on and glance at are the forward, left and rear view mirror. The sample drivers were less likely to demonstrate troublesome transition patterns, particularly in mentally demanding situations. Additionally, instead of shifting vision directly between two non-forward focal points, the sample drivers frequently had an intermediate forward glance. Thus, seemingly unrelated paths could be grouped into explanatory patterns of driver attention allocation. Finally, in addition to the vision-transition patterns, the potential pitfalls of such patterns and possible countermeasures to improving safety are illustrated, focusing on situations when drivers are distracted, traveling at high speeds and approaching intersections. ã 2014 Elsevier Ltd. All rights reserved.

Keywords: Renewal cycle Visual attention Vision transition Naturalistic driving

1. Introduction Perceiving information from the environment, 90% of which is visual (Ho and Spence, 2008), is the fundamental step of comprehending driving situations, making decisions and performing actions (Endsley, 1995). Drivers, to be aware of situations, must allocate visual attention resources to areas of interest. In addition, drivers experiencing problems in visual attention allocation will result in decision errors and actions under an insufficient understanding of the driving environment and thus possibly increased crash risk (De Waard et al., 2008, 2009; Marmeleira et al., 2009). In the United States, recognition errors, including inattention, distraction and inadequate surveillance, contributed to 41% of human-factor-related crashes (NHTSA, 2008). The US Department of Transportation’s National Highway Traffic Safety Administration also estimated that there were at least 3000 deaths annually from crashes attributed to distraction, specifically due to the use of in-vehicle devices (NHTSA, 2012). Thus, exploring visual attention

* Corresponding author. Tel.: +886 2 2349 4959; fax: +886 2 2349 4953. E-mail addresses: [email protected] (S.-H. Huang), [email protected] (J.-T. Wong). 1 Tel.: +886 2 2349 4995; fax: +886 2 2349 4953. http://dx.doi.org/10.1016/j.aap.2014.10.010 0001-4575/ ã 2014 Elsevier Ltd. All rights reserved.

allocation, which is defined as conscious vision transitions, for drivers under varying conditions is vital for preventing crashes. A model capturing vision transition among various focal points is needed. Previous studies have intensively analyzed driver visual attention allocation and extracted several factors related to attention demand (Underwood et al., 2002a,b, 2003; Martens and Fox, 2007; Levin et al., 2009; Borowsky et al., 2010; Konstantopoulos et al., 2010). Wickens et al. (2003, 2007); Wickens et al., (2003, 2007) proposed the concept of the SEEV model, summarizing factors into four constructs, salience, effort, expectancy, and value (Wickens et al., 2003, 2007; Horrey et al., 2006Werneke and Vollrath, 2012). Drivers pay more attention to the target that is more relevant to safety (value), or threats expected (expectancy) or salient. Meanwhile, drivers are more likely to shift vision to focal points closer to the current gazed point (effort). These studies have provided useful ways for driver visual behavior analyses; however, they show only the aggregated results and have not taken the dynamic attention-allocation behavior into consideration. Therefore, an enhanced understanding of information perception and situational awareness in naturalistic driving requires a refined model to capture vision transition among various focal points.

108

S.-H. Huang, J.-T. Wong / Accident Analysis and Prevention 74 (2014) 107–117

To represent the process of drivers’ visual attention allocation, previous studies have used the focal point glance probability (Wickens et al., 2003, 2007; Horrey et al., 2006Horrey et al., 2006; Wong and Huang, 2011) or the proportion of time spent on specific targets (Underwood et al., 2003; Levin et al., 2009; Borowsky et al., 2010; Konstantopoulos et al., 2010) to describe visual attention allocation. With these types of representation, the research results commonly showed that drivers spend most of their time looking forward. However, no detailed vision shifts among non-forward focal points are depicted. It would be difficult to find the sequential connection between two non-forward focal points, such as the path of shifting vision from the right to the front and finally to the left (Wong and Huang, 2013a). Thus, developing an effective method for representing the vision transition process is the first challenge in analyzing visual attention allocation. In this study, the renewal cycle concept developed by (Wong and Huang, 2013a) is adopted to model the vision transition process. The next step is to identify the factors affecting vision transitions. The factors considered in the widely recognized SEEV model provide good references for this study. Lastly, an econometric model is needed to capture drivers’ visual attention allocation behavior. In other words, the objective of this study is to model visual behaviors and estimate the effects of various contributing factors on drivers’ vision transitions under limited data from realistic field settings 2. Vision transition process Wong and Huang proposed the renewal cycle concept to represent visual attention allocation (Wong and Huang, 2013a,b). The forward glance, where drivers spend most of their time looking, is set as the reference point. A renewal cycle is defined as the process of glances starting from the reference point (forward), shifting to other focal point(s), and back to the reference point (forward) again. Treating the entire glance sequence of off-road focal points as a basic component avoids the creation of an analysis of visual attention allocation overly concentrated on forward glances; and it enables observations of the interaction between glances towards forward and non-forward focal points. In particular, renewal cycles containing various non-forward focal points can be identified and analyzed as needed. The most frequently found renewal cycles, reflecting general visual behavior to prevent loss of awareness, involved drivers glancing from the front to only one non-forward focal point. Among them, certain renewal cycles can recur. Such repetitious behavior represents how drivers divide a long glance at a target into several shorter glances and repeatedly shift vision from the front to the intended focal point (Metz et al., 2011; Wong and Huang, 2013a). Moreover, instead of shifting vision along two nonforward focal points consecutively, drivers generally shift vision back to the front before shifting to another non-forward focal point. However, Wong and Huang (2013a) showed a substantial proportion of renewal cycles, approximately 10%, contained more than one non-forward focal point. Some of the multiple off-road glances were planned and may have no risks (Dukic et al., 2012). Despite the potential risk of losing awareness about leading traffic, drivers frequently shift vision directly between two non-forward focal points as needed under specific conditions, such as when approaching intersections. Still, a large portion of these renewal cycles containing more than one non-forward focal point were dangerous, particularly under complex road conditions and were likely the primary cause of crashes (Wong and Huang, 2013b). To capture the vision transition paths between any two nonforward focal points, this study investigated the focal point choices after each non-forward glance. The conceptualized focal point choices follow a loop process in which drivers must choose a focal

Model for choosing focal points after glancing at nonforward focal point NFa Non-forward glance (NFb or others) Non-forward glance (NFa)

Non-forward glance (NFa) Forward glance Non-forward glance (NFb or others)

Fig. 1. Major types of vision transition.

point at which to glance based on the current glanced point. With the concept of renewal cycles, three types of vision transitions can occur. In the developed loop process, Fig. 1 illustrates an example of vision transition after a non-forward glance, NFa. The first type of vision transition is shifting from focal point NFa directly to another non-forward focal point (NFb or others). In this case, the current renewal cycle is incomplete because the driver has not shifted vision back to the front. For other types of vision transition, a driver who ends the current renewal cycle and begins a new one may choose to look at the non-forward focal point NFa again. This second type of vision transition is referred to as the repeated renewal cycle. The third type of vision transition involves drivers shifting vision to another non-forward focal point (NFb or others) after glancing at the non-forward focal point NFa and the front sequentially. This type of vision transition requires drivers to determine a non-forward focal point at which to glance in the new renewal cycle, which forms a vision transition from one nonforward focal point to the forward side, and then to another nonforward focal point. 3. Model and specification 3.1. Model framework Visual attention allocation is a continual process of choosing focal points and is described according to the three types of vision transition defined in Fig. 1. Intuitively, vision transition could be analyzed by a transition (duration) model or, alternatively, by a multinomial logit (MNL) model. The MNL is widely applied in various studies and considered a suitable tool for choice behavior; thus, it is adopted in this study. To represent the path connecting two non-forward focal points, the model consists of several submodels; each represents the vision being shifted from a specific non-forward focal point. Fig. 2 shows the conceptual framework of an MNL sub-model representing vision transition from the specific non-forward focal pointNFa. Focal point choices involve two sequential steps, represented by the two MNLs shown in Fig. 2. In the first layer – modeling types of renewal cycle – vision shifts after a non-forward glance can be one of the following three types: shifting vision back to forward to begin a new renewal cycle, shifting vision back to forward to repeat the current renewal cycle, or continuing the current renewal cycle by shifting vision directly to another non-forward focal point. The results of the layer 1 model are used to calculate the probabilities of new renewal cycles, repeated renewal cycles, and multiple-glance renewal cycles. Once the alternative of starting a new renewal cycle is chosen in the layer 1 model, another model for calculating the probability of choosing a specific non-forward focal point other than the point NFa is required. Thus, this study formulated the second layer MNL model to derive the probability of a path connecting two renewal cycles. The results from the second layer model reveal the

S.-H. Huang, J.-T. Wong / Accident Analysis and Prevention 74 (2014) 107–117

Glance to forward area Vision transition model for non-forward focal point (NFa)

Glance at NFa

Vision shift back to the frontal side Start a new renewal cycle

Vision shift to another non-forward area

Repeated renewal cycle (NFa) Modeling Types of Renewal Cycle

NFb

...

NFc

Path of connecting two non-forward points

Vision transition model for non-forward focal point (NFb)

Vision transition model for non-forward focal point (NFc)

...

Fig. 2. Framework of the vision transition model.

probabilities of driver vision transitions in connecting various nonforward focal points indirectly. 3.2. Model specification Determining the factors affecting driver visual attention allocation is crucial for model development. Based on the constructs proposed by the SEEV model and the available attributes provided by the 100-car naturalistic driving study, Table 1 illustrates the factors used in the model specification of this study. The alternative specific constant (ASC) reflects the difference in the utility of an alternative from that of the base alternative when all other conditions are equal. Thus, the ASC reflects the relative preference of the respective focal point or type of renewal cycle chosen. For choosing another non-forward focal point in the second layer, the estimated constants represent the potential paths that drivers are more likely to follow. Such path results can be used to reflect the effort of shifting vision from one focal point to another, which is determined by the visual angle difference

109

between two focal points (Wickens et al., 2003, 2007 Horrey et al., 2006). The drivers were expected to more likely shift vision to objects close to the current glanced point than farther objects. Attributes related to salience, such as color and conspicuity, represent the ease of differentiating and identifying targets (Gershon et al., 2012; Koustanaï et al., 2012; Mcintyre et al., 2012). Unfortunately, the 100-car event database did not provide the salience information of each glanced target. Weather conditions (rainfall) were used as a surrogate to represent the possible effect of the ease of identifying targets on shifts in driver vision. The lack of available salient attributes is recognized a limitation of this study. The third type of attribute, expectancy, represents the expected frequency of threats or information appearing (information bandwidth). A high information bandwidth suggests frequent status changes that drivers must update. The traffic density and speed can induce interaction and/or potential conflicts among vehicles; thus, these aspects represent the expectancy of a glanced focal point (De Waard et al., 2008; Werneke and Vollrath, 2012). To make the results easily interpretable, the speed was mean centered, with a reference speed of 25 mph (the sample mean), thereby making alternative specific constants for the vision shifts related to the field driving conditions. The final attribute of the SEEV model, value, represents the importance and safety relevance of a focal point. Drivers focus primarily on future trajectories where they may encounter possible threats (Summala et al., 1996; Werneke and Vollrath, 2012, 2013). Areas containing threats involving high crash risk might attract greater attention from drivers than other less relevant focal points (Martens and Fox, 2007; Koustanaï et al., 2012). Thus, this model took the intersection into considerations to determine the possible effect of drivers approaching or passing an intersection. In addition, a variable named “distraction” was included to represent the relevance or importance of in-vehicle or off-road distractions (Blanco et al., 2006; Kiefer and Hankey, 2008). In addition to SEEV model attributes, renewal-cycle-related attributes were also included. A long off-road glance duration increases the probability of vision shifting back to the front (Brown et al., 2000). Previous studies have also shown that the duration of forward and non-forward glances influences focal point choices (Wong and Huang, 2013a). Therefore, the model included two renewal cycle related attributes: the consecutive off-road duration before the driver’s next glance and the forward-glance duration in the current renewal cycle. Both attributes were also mean centered

Table 1 Specification of the multinomial logit model. Variable

Description

Effort ASC

Alternative specific constants for MNL

Salience Rain

A dummy variable, equal to one when the driver drove in rainy days

Expectancy Speed LOS_B LOS_C LOS_DE

Driving speed recorded with glances. This study set 25 mph as the reference level of the speed attribute A dummy variable, equal to one under level of service (LOS) B, which is 12–20 pc/mi/ln A dummy variable, equal to one under LOS C, which is 20–30 pc/mi/ln A dummy variable, equal to one under LOS D–E, which is 30–67 pc/mi/ln

Value Intersection Distraction

A dummy variable, equal to one when driving through intersection A dummy variable, equals to one when a driver is distracted by cognitive, cell phone, in-vehicle devices, external clutter and activity, etc

Renewal cycle Off-road duration On-road duration

Consecutive duration that drivers glance off-road before choosing the next focal point to be glanced. This study set 1 s as the reference level of the off-road duration attribute Duration of forward glance in the current renewal cycle. This study set 4 s as the reference level of the on-road duration attribute

110

S.-H. Huang, J.-T. Wong / Accident Analysis and Prevention 74 (2014) 107–117

based on the sample means of on-road (4 s) and off-road (1 s) durations, respectively. 4. Data from naturalistic driving Data from the event database of the 100-car naturalistic driving study, including the 68 crash and 760 near-crash incidents, conducted by the Virginia Tech Transportation Institute (Neale et al., 2002; Dingus et al., 2006; Klauer et al., 2006; VTTI, 2012) were available for the analysis. Each recorded incident contained the visual glances and related attributes of drivers within 30 s before crashes or near crashes. The 30 s duration was divided into two parts, before and after the precipitating events, which were determined as the primary causes of incidents. According to the Fleury and Brenac (2001), there are five stages related to driving: (1) the situation prior to driving, (2) the driving situation, (3) the discontinuity situation (the precipitating events preceding a crash or near-crash incident), (4) the emergency situation for evasion, and (5) the collision situation if the emergency evasion is not successfully performed. Based on the classification, it was assumed that the data collected before the precipitating events, namely the stage of driving situation, were close to the sample drivers’ habitual driving behavior and could be used for model development. The data recorded in the 100-car event database involves three parts: glance data, time series data, and reduced data. The glance data recorded the location and the associated duration of sample drivers’ glances in every 0.1 s. The database contains one forward focal point, the glances straight out of the windshield, and 12 nonforward focal points: left forward, right forward, left mirror, left window, right mirror, right window, rearview mirror, instrument cluster, center stack, cell phone, interior object and passenger. Some of these focal points were rarely glanced at. To meet the sample size requirement for the model estimation, this study grouped the three focal points on the left (and right) side together to manage vision shifts from the left (and right) side of a vehicle. In addition, all glances inside the vehicles, including those directed at an instrument cluster, center stack, cell phone, interior object and passenger, are grouped together as in-vehicle distractions. Therefore, five types of focal point are used in this study: forward (F), left side (L), right side (R), rearview mirror (ReM), and invehicle distraction (InvD). The time series data describe driving behavior, including speed and distractions. A distraction refers to a non-driving-related action rather than a focal point. A driver can continue to look at the roadway while distracted. The other attributes of the reduced data include the relation to a junction, traffic density, pre-event maneuvers, and surfaces. Among the data attributes, the relation to junction and pre-event maneuvers are the two safety-relevant attributes that are used to describe the possible direction of oncoming threats, which are related to the value construct of focal points. Several combinations could be derived from maneuvers and relations to a junction. Considering the sample size, the model specification and the interpretation of the results, only data associated with traveling straight maneuver were used for the investigation. 5. Results In total, 1461 focal point choices while driving on road segments were collected. As shown in Fig. 3, most drivers drove during dry days (85.49%), on segments (68.65%) and without distraction (82.47%). As for the traffic density, LOS A, B, C and D/E share 27.53%, 34.77%, 21.14% and 16.56% of the focal point choices, respectively. Four sub-models representing the path from the associated four non-forward focal points were estimated by NLOGIT 3.0 (Greene,

Fig. 3. Glance data distribution by attributes.

2002). The insignificant coefficients were removed based on a backward stepwise procedure. 5.1. Vision shift from the left side Table 2 shows the estimated sub-models for vision shifts from the left side of vehicles. The first layer results show that drivers, following a left glance, were more likely to shift their vision back to the front to begin another renewal cycle. Comparatively, they preferred not to shift vision from the left directly to another nonforward focal point, particularly under the LOS D and E. The probability of shifting vision directly to other non-forward focal points decreased with increased off-road glance durations. This result implies that the drivers were aware of the increased risk either from traffic situations or off-road long glances and showed cautious behavior. Moreover, the drivers travelling at high speeds, with distractions, or under LOS C conditions were more likely to repeatedly shift their vision to the identical focal point on the left side following a forward glance. The layer 2 model in Table 2 shows the choices of non-forward focal points for starting a new renewal cycle after a left glance. The most frequently observed vision shift was from the left side to the front and then to the rearview mirror or the right. When approaching intersections, the probability of vision shifting to the rearview mirror decreased; however, the probability of switching to the right increased, suggesting a vision behavior from the left to the right with a brief forward glance in between to check the traffic situations on the intersected roadway. This vision behavior may represent a strategy of preventing long off-road glances. Notably, it occurred less frequently under high speeds and heavy traffic (LOS C–E) conditions, in which drivers would prefer repeatedly shifting vision between the left and the front instead of observing both sides of the vehicle. As for the goodness of fit of the model, the r2 in layer 2 is relatively low. The vision shift after a left glance may be rather stochastic. The shift sequence may rely on the dynamic road environment and driver responses to instant situations. Unfortunately, the data needed for more detailed model development and calibration are unavailable. 5.2. Vision shift from the right side Table 3 shows the results of the estimated logit model for vision shifts after right glances. The layer 1 model shows that drivers were less likely to shift vision directly from the right to another non-forward focal point, particularly under high speeds, rain, LOS C, or long off-road glances. The probability of repeatedly shifting vision between the right and forward is lower than that of starting a new renewal cycle. However, under the traffic density for LOS D/ E, the probability of a repeated renewal cycle after a right glance increases substantially. This result suggests that drivers were alert to situations on the right side when traffic increased.

S.-H. Huang, J.-T. Wong / Accident Analysis and Prevention 74 (2014) 107–117

111

Table 2 Estimated logit model for the path from the left side.

ASC Off-road duration On-road duration Speed Intersection Distraction Rain LOS B LOS C LOS DE Sample size LL(b) LL(0)

r2

Layer 1 (choosing types of vision transition)

Layer 2 (choosing another non-forward focal point when beginning a new renewal cycle)

New renewal cyclea

Lb

R

ReM

0.11

0.41

0.44**

Repeated renewal cycle

Direct to other non-forward focal point

0.79**

1.03** 0.03** 0.01**

0.01** 0.01**

0.01*

0.01* 0.02* 0.60*

InvDa

0.01* 0.77*

0.62*

0.49** 232 432.44 492.17 0.10

1.54** 77

139

53 302.55 321.62 0.03

0.74* 1.09** 69

63

47

a

Set as the base alternative in model estimation. Represents a path from any one of the focal points on the left side to another one also on the left side. * Significant at the level of 0.1. ** Significant at the level of 0.05. b

When starting a new renewal cycle, the layer 2 model in Table 3 shows that a vision shift to the left after a right glance was highly preferred. However, the probability decreased when driving at high speeds and under LOS B–E, possibly because of the increased risk of task complexity under these conditions. Another notable path is the vision moving from the right, sequentially to the forward and then to the rearview mirror. Similar to vision shifts from the left, the probability of a rearview mirror glance decreased substantially when approaching intersections. Moreover, the result clearly demonstrates that the probability of vision transitions from the right to the forward side and then to an in-vehicle distraction increased during distractions. 5.3. Vision shift from the rearview mirror Table 4 shows the estimated logit model for vision shifts from the rearview mirror. The result shows that off-road duration had a positive effect on the probability of choosing repeated renewal

cycles, suggesting that the sample drivers relied heavily on the rearview mirror and checked it repeatedly. Compared with other types of vision transition, the probability of repeated rearview mirror glances increased with speed and also increased under rain conditions, which implies that high expectancy and low conspicuity stimulate drivers to enhance their situational awareness. Consequently, drivers put more visual resource on the forward and rearview mirror glances and less on the other non-forward focal points. The probability of shifting vision directly from the rearview mirror to other non-forward focal points is exceedingly low compared with the other two types of vision transitions. Looking at the rearview mirror inhibited drivers from perceiving the situation changes around the vehicle, resulting in a potential loss of awareness towards leading traffic. As a consequence, drivers were more likely to shift vision back to the front (for a new renewal cycle or a repeated one) than allocate attention to other areas (Brown et al., 2000).

Table 3 Estimated logit model for the path from the right side. Layer 1 (choosing types of vision transition)

Layer 2 (choosing another non-forward focal point when beginning a new renewal cycle)

New renewal cyclea Repeated Direct to other non-forward focal point L renewal cycle ASC Off-road duration On-road duration Speed Intersection Distraction Rain LOS B LOS C LOS DE Sample size 155 LL(b) 238.15 LL(0) 304.31 0.19 r2 a

1.14** 0.01

0.80** 0.11**

Ra,b

2.13**

*

0.59**

InvD 0.05

*

0.05**

0.04** 1.53** 1.17**

1.30** 1.24** 0.94** 58

64

1.37** 1.90** 1.80** 68 176.64 214.87 0.12

Set as the base alternative in model estimation. Represents a path from any one of the focal points on the left side to another one also on the right side. Significant at the level of 0.1. ** Significant at the level of 0.05. b

ReM

2.30** 29

34

24

112

S.-H. Huang, J.-T. Wong / Accident Analysis and Prevention 74 (2014) 107–117

Table 4 Estimated logit models for the path from the rearview mirror.

ASC Off-road duration On-road duration Speed Intersection Distraction Rain LOS B LOS C LOS DE Sample size LL(b) LL(0)

r2

Layer 1 (choosing types of vision transition)

Layer 2 (choosing another non-forward focal point when beginning a new renewal cycle)

New renewal cyclea

L

Repeated renewal cycle

Direct to other non-forward focal point 26.21** 2.86**

0.14 0.06**

Ra

InvD

0.35*

0.09

0.02**

0.57*

151 208.36 339.47 0.36

137

21

61 164.15 165.89 0.01

43

47

a

Set as the base alternative in model estimation. Significant at the level of 0.1. ** Significant at the level of 0.05. *

The layer 2 model shows that to start a new renewal cycle after a rearview mirror glance, the sample drivers preferred shifting their attention to left focal points. However, excluding the ASC for the left, no variables showed significant effects. The r2 of the layer 2 model is extremely low, which may suggest that after rearview mirror and forward glances, the driver’s vision transitions were somewhat random, without specific sequences. This result could also be attributable to the small sample size and insufficient explanatory variables for dynamic road/traffic situations. 5.4. Vision shift from in-vehicle distractions Table 5 shows the estimated logit model for vision shifts from in-vehicle distractions. The unique characteristic of the result is the high probability of repeated renewal cycles, particularly when a driver was distracted at high speeds. The repetition becomes increasingly evident as the off-road glance duration increases and the on-road glance duration decreases. This result clearly indicates the visual behavior of distracted drivers; the more distracting activities there are, the smaller the duration of forward glances and the Table 5

more repeated in-vehicle glances. After the interaction with in-vehicle distractions, the layer 2 result shows that drivers frequently shifted their vision to the left or the rearview mirror instead of the right. Glancing at these two focal points helped the drivers understand the traffic situations in surrounding areas, which they may not have been able to update during distractions. The result also shows the positive effect of offroad glance duration on the probability of rearview mirror glances. This finding emphasizes the importance of the situation on the rear side. Drivers who spent a long time glancing in-vehicle distractions were more likely to start a new renewal cycle to the rearview mirror. Interestingly, the drivers who drove under rainy conditions tended to shift attention to the left, possibly to check the lateral position of the vehicle. 5.5. Off-road glances involving multiple focal points Based on the estimation results, direct vision transition between two non-forward focal points is rare. However, it may be the primary cause of crashes due to long off-road glances and is

Estimated logit models for the path from in-vehicle distraction.

ASC Off-road duration On-road duration Speed Intersection Distraction Rain LOS B LOS C LOS DE Sample size LL(b) LL(0)

r2 a

Layer 1 (choosing types of vision transition)

Layer 2 (choosing another non-forward focal point when beginning a new renewal cycle)

New renewal cyclea

L

Repeated renewal cycle 0.53** 0.03** 0.01** 0.01*

3.33** 0.35**

Ra

0.57**

ReM 0.82** 0.07**

0.05**

0.94** 0.81* 0.53**

126 289.92 469.11 0.36

279

Set as the base alternative in model estimation. Significant at the level of 0.1. ** Significant at the level of 0.05. *

Direct to other non-forward focal point

22

49 126.28 138.42 0.06

23

54

S.-H. Huang, J.-T. Wong / Accident Analysis and Prevention 74 (2014) 107–117

113

Fig. 4. Probability of directly transiting vision to one of the other non-forward focal points.

worthy of further investigation. Fig. 4 shows the probability of drivers consecutively shifting vision from one non-forward focal point to another under driving speeds of 25 and 50 mph. Fig. 4 (a) and (b) shows that the probability of shifting vision directly to another non-forward focal point after a brief right or left glance is relatively high and decreases with an increased duration of glancing off-road. According to Wong and Huang, (2013a), the sample drivers sometimes shift vision along the path from the left mirror to the left window and the path from the right forward to the right window. These paths could represent their visual behavior of checking the road side area or the adjacent lanes to create a mental picture of how the traffic is likely to develop. This result also supports the statement that drivers are more likely to shift their vision to a closer focal point than to a distant one (Wickens et al., 2003, 2007Horrey et al., 2006; Werneke and Vollrath, 2012). By contrast, Fig. 4 (c) and (d) strongly suggests that the sample drivers did not glance at another non-forward focal point immediately after looking at in-vehicle distractions or the rearview mirror. The probabilities decreased sharply with the off-road glance duration, particularly when traveling at a high speed, indicating that the drivers realized the high risk of long glances directed at either in-vehicle distractions or the rearview mirror. 6. Discussion Based on the model estimation results, Table 6 summarizes the characteristics and possible implications of driver visual attentionallocation patterns under varying conditions. Although the results may not represent typical visual behavior, the vision transition

patterns of the sample drivers derived from the proposed model are interesting and merit our attention for further study. 6.1. Pattern of vision transitions Modeling the process of vision transition provides valuable clues for understanding drivers’ visual attention allocation. The paths observed in this study mostly involve vision shift to the front, the left, and to the rearview mirror. Although these patterns were derived from the limited sample drivers, the common paths could still illustrate an insightful message for understanding drivers’ visual behavior. Notably, there were few right glances found in the study. The possibility of ignoring safety-related information on the right roadside means more attention should be paid to sustainable roadside design to meet drivers’ general visual behaviors, such as locating roadside signs in areas where drivers frequently glance (Crundall et al., 2006). Repeated vision transitions between the front and an in-vehicle distraction or the rearview mirror were frequently found. Whether the repeated renewal cycles were for specific purposes or simply a typical behavior of vision transition was not clear. Nevertheless, the probability of repeated glances at the rearview mirror increased with off-road duration, which might suggest that drivers took time to check and reconfirm the rear-side status. This type of visual behavior may be associated with a way of defensive driving and maintaining situation awareness for road traffic. Meanwhile, the probability of repeated glances at an in-vehicle distraction increased with off-road duration but decreased with on-road duration, suggesting that the drivers were inclined to have brief forward glances to complete the intended in-vehicle activities.

114

S.-H. Huang, J.-T. Wong / Accident Analysis and Prevention 74 (2014) 107–117

Table 6 Patterns of vision transition under varying conditions. Vision transition patterns Conditions

Starting a new renewal cycle

General conditiona

 Rearview mirror is glanced at following a  Repeatedly glancing at ReM, particularly under  Forward, left (mirror) and rearview mirror are

High speed

 Probability of cross-vehicle transition (L–F–R  Probability of repeated all renewal cycles  Pay close attention to the forward direction

Repeated renewal cycle

Implications

glance towards the left or right, or towards the conditions of long off-road durations. the three most attended focal points. in-vehicle distractions. (L–F–ReM, R–F–ReM,  Repeatedly glancing at InvD, particularly under  Drivers take time to repeat ReM glances for InvD–F–ReM) the conditions of long off-road and short onreconfirming traffic conditions. road durations.  Left side is glanced at following a glance  Drivers glance at InvD repeatedly to complete towards the right, at the rearview mirror, or the intended activities. towards in-vehicle distraction. (R–F–L, ReM–F– L, InvD–F–L)

and R–F–L) decreases significantly.

increases.

when driving at high speeds.

 Instead of a long glance, repetitious glances are adopted to compensate for the potential loss of awareness. When a driver is distracted

 Probability of R–F–InvD increases.

 Probability of repeating glances at InvD and L  Drivers’ attention is frequently distracted by increases.

InvD.

 InvD glances are often done repetitiously.  Left glances are more likely to be repeated while a driver is distracted.

Intersection

 Probability of L–F–R increases.  Probability of L–F–ReM and

 None

 Drivers shift their vision from left to right frequently when approaching intersections.

R–F–ReM

 Rearview mirror is less attended when

decreases.

approaching intersections. a

Represents the condition for which all variables (except the alternative constants) are set at a value of zero.

Therefore, designing in-vehicle devices requiring less drivers’ effort might be a potential countermeasure to improve traffic safety. 6.2. Speed High speeds result in a long stopping distance and a short time to collision. They also narrow drivers’ peripheral vision and blur safety-critical information from the roadside. Therefore, the potential risk of long off-road glances increases. Driver’s attention concentrating on the leading area is required. Consequently, drivers should refrain from troublesome vision transitions, such as shifting vision from one side of a vehicle to another side. To ensure safety against leading traffic and to compensate for the degraded peripheral vision, drivers are advised to conduct repeated renewal cycles to allocate visual attention in collecting surrounding information. In addition to the drivers’ visual behavior, unmanageable driver vision transitions could be induced by poorly designed road environments. Therefore, to provide a safe driving environment, a road safety audit for removing onerous distractions or designing roads with speeds agreeable to drivers’ expectation could be effective. Furthermore, safety information should be provided based on drivers’ general visual behavior. For example, the sample drivers were found to shift vision from one side of the vehicle to the other side through a forward glance less frequently in high speed situations. Training driver good visual behavior when driving at high speeds would be beneficial. 6.3. Distraction A brief and repeated glance technique was found to be a crucial compensatory strategy for drivers interacting with in-vehicle distractions, which constitute the primary source of distraction.

Interactions with in-vehicle distractions were generally conducted repetitiously. Therefore, it is advised that in-vehicle facilities be designed dividable, enabling drivers to complete an entire activity in separate glances. The content format, size, and complexity of invehicle devices must be designed to ensure that drivers can operate safely, correctly and efficiently. NHTSA (2012) has proposed voluntary guidelines for vehicle manufacturers to discourage the introduction of excessively distracting devices. In addition, thresholds for the time required for drivers to processing information, such as 2 s in each glance and 12 s in total (Fitch et al., 2013), should be set as a guideline for law enforcement and for design. Distractions have an impact on the allocation of visual resources. This study found the drivers repeatedly glanced towards the left while distracted. Drivers glancing at only limited focal points may not be able to observe their surroundings effectively. In such cases, a collision avoidance system could potentially assist drivers in preventing possible conflicts (Shaheen and Niemeier, 2001; Tan and Huang, 2006; Maltz and Shinar, 2007). In addition, by measuring off-road durations and the number of repeated glances, in-vehicle monitoring devices may be applied to monitor driver’s distractions and to offer alerts to driver. 6.4. Intersection Passing through or approaching an intersection creates distinct attention demands and visual attention allocation patterns. To ensure intersections are safely passed, drivers must invest their attention resources on the intersected roadway and glance less frequently at the rearview mirror. As a result, vision shifts from the left to the forward side and finally to the right appeared frequently, suggesting that drivers paid more attention to the possible threats from the intersected road. On the contrary, attention paid to other focal points was reduced accordingly. If a troublesome vision

S.-H. Huang, J.-T. Wong / Accident Analysis and Prevention 74 (2014) 107–117

115

Table 7 Potential pitfalls and possible safety countermeasures. Conditions

Potential pitfalls

General condition

 Potential ignorance of vision transition among focal points other than left  If needed, drivers should glance at those focal points as briefly as possible. side and rearview mirror.

Possible countermeasures

 Long glances at InvD.

 Necessary information is presented or alerted shortly, verbally and actively.

 Locate in-vehicle devices in the area where drivers can comfortably glance.

 Missing critical safety information caused by infrequent right glances.

 A blind zone detection and warning system for possible threats from the right side.

 In-vehicle intelligent safety devices for collecting critical roadside information. High speed

 Limited peripheral vision and shorter time to collision.

 Avoiding long glances at unrelated focal points.  Road audit to eliminate inadequate roadside distractions.

 Ignoring necessary roadside information caused by restricted peripheral  Sustainable road design with enhanced hazard information systems. vision.  Defensive driving training for attention allocation in high speed driving. When a driver is distracted

 Information contents are not adequate for driving tasks or environment.  Dividable information enabling drivers to comprehend with repeated glances.

 Accounting for characteristics of drivers’ vision transition patterns when designing driving information.

 Driving distractions are not monitored properly.

 Criteria of distraction level including the times of repetition, duration of each repetition, and total duration set for alerts.

 Collision warning system for crash prevention against threats from surrounding areas. Intersection

 Too many focal points require drivers’ attention.

 Intersection decision support to help drivers identify threats.  Intersection area signs and geometry design in considering drivers’ patterns of vision transition.

transition is required due to a complicated road traffic environment around an intersection, a significant driving risk may occur. To ease the effort required, identifying the vision shift patterns in different stages of approaching intersections is vital to resolving the problem. Signs or necessary information systems, such as intersection decision-support signs, could then be installed in proper locations to assist and ease drivers’ viewing (Laberge et al., 2006; Creaser et al., 2007; Neale et al., 2007). Finally, infrequent rearview mirror glances may imply that drivers have limited and insufficient situational awareness towards rear traffic situations. Following the above discussions, potential pitfalls that drivers might encounter and the possible countermeasures are illustrated as in Table 7. 7. Conclusion Driver distraction has become an increasingly critical issue in traffic safety, attracting great attention from researchers and government agencies. Salvucci (2009) proposed a Distract-R system to evaluate driver performance under secondary tasks, specifically the use of in-vehicle devices. In the system, allocation of visual attention resource is one of the primary issues to be further investigated. To establish better scientific understanding of the visual attention allocation under varying conditions, this study developed models capturing vision transition among various focal points. Based on the renewal cycle concept and the four constructs proposed in the SEEV model, four two-layer multinomial logit (MNL) type models for capturing patterns of vision shifts from the right, the left, the rearview mirror and the in-vehicle distractions, respectively, have been successfully formulated and estimated. The results of the first layer model are used to calculate the probabilities of new renewal cycles, repeated renewal cycles and

multiple-glance renewal cycles. Results from the second layer model reveal the probabilities of driver vision transitions in connecting various non-forward focal points indirectly. The advantage of using the renewal cycle to represent the vision shifts among non-forward focal points has been shown to provide better understanding of the drivers’ visual attention allocation. The study results reveal that the most frequently observed paths, in addition to the forward side, were related to the rearview mirror and the left side, suggesting that these two focal points are the most critical areas that drivers must constantly check. Meanwhile, instead of shifting vision directly between two non-forward focal points, drivers typically have an intermediate forward glance. Thus, connecting two distinct renewal cycles offsets the troublesome vision transition across a vehicle. The renewal cycle concept allows us to group these seemingly unrelated paths for better interpreting visual attention allocation patterns. The estimated model for the vision shift from in-vehicle distractions showed a high probability of repeated renewal cycles. The repetition becomes increasingly evident as the off-road glance duration increases and the on-road glance duration decreases. This result clearly indicates the visual behaviors of distracted drivers; the more distracting activities there are, the smaller the duration of forward glances and the more repeated in-vehicle glances. This finding also implied that drivers would be more cautious when interacting with the in-vehicle distractions, dividing long glances on in-vehicle distraction into several short glances with intermediated forward glances. Direct vision transition between two non-forward focal points is rare. The sample drivers did not glance at another non-forward focal point immediately after looking at in-vehicle distractions or the rearview mirror. The probabilities decreased sharply with the off-road glance duration, particularly when traveling at a high speed. This indicates that the drivers realized the high risk of long

116

S.-H. Huang, J.-T. Wong / Accident Analysis and Prevention 74 (2014) 107–117

glances at either in-vehicle distractions or the rearview mirror. Additionally, drivers passing through or approaching an intersection create distinct attention demand and visual attention allocation patterns. Vision shifts from the left to the forward side and finally to the right appeared frequently, suggesting that drivers paid more attention to the possible threats from the intersected road. On the contrary, attention paid to other focal points was reduced accordingly. The model results could offer insights into driver visual behavior and patterns of visual attention allocation. This visual attention modeling may contribute to the designing of traffic safety information. Nevertheless, there still are limitations that should be emphasized. The r2 in the second layer models were very low, possibly due to the data limitations–the small sample size and the incomplete affecting factors used in the models. Dynamic changes in the road environment and the traffic situations affect vision shifts substantially among non-forward focal points. Particularly, the salience of a target and the environment changes were not available in the dataset. Lacks of the data may cause biased estimate of the models. Moreover, the event database includes only drivers who ultimately experienced crashes or near-crashes, which may not be able to represent general driver visual behavior. The acquiring baseline data, and compared with the crash data, may provide more clues to distinguishing risky patterns from less risky ones. In addition, drivers with various physical and psychological characteristics could lead to different strategy of vision transition. The definition and determination of a glance still is a questionable issue requiring further efforts. The shortest glance observed in this study is 0.1 s, owing to the sampling rate of eyetracker. Although some studies, such as Tijerina et al. (2004), have considered 0.1 s glances as a complete one, such a glance could be a part of the rapid transiting process between two non-forward glances. It is fortunate that the forward glance with durations at 0.1 s, 0.2 s, and 0.3 s share only 1.5%, 2.1% and 3.7% of the focal point choices, respectively. This study did not distinguish the rapid transiting from the regular ones. Moreover, glances are not necessarily made with consciousness. Drivers may look at the target unconsciously without being engaged. In such conditions, the glance may not be related to the driving environment, and do not reflect drivers’ patterns of vision transitions, either. Distinction between glance with and without consciousness deserve more research attention. In sum, the findings of this study constitute the primary step toward more thorough analyses. Further research is necessary to assist drivers in operating motor vehicles on a safer road. Hopefully, in the long run, the derived patterns of vision transition could be inputs to a system like the Distract-R (Salvucci, 2009) for exploring detailed driver visual behavior under secondary tasks. In such a way, the vision transition model proposed in this study can be fruitful for preventing traffic accidents. Acknowledgments The authors would like to thank the Ministry of Science and Technology, Taiwan, Republic of China, for financially supporting this research (NSC 100-2221-E-009-120-MY3) and the anonymous reviewers who have provided very insightful comments. References Blanco, M., Biever, W.J., Gallagher, J.P., Dingus, T.A., 2006. The impact of secondary task cognitive processing demand on driving performance. Accid. Anal. Prev. 38 (5), 895–906. Borowsky, A., Shinar, D., Oron-Gilad, T., 2010. Age skill, and hazard perception in driving. Accid. Anal. Prev. 42 (4), 1240–1249. Brown, T., Lee, J., Mcgehee, D., 2000. Attention-based model of driver performance in rear-end collisions. Transport. Res. Rec.: J. Transport. Res. Board 14–20.

Creaser, J.I., Rakauskas, M.E., Ward, N.J., Laberge, J.C., Donath, M., 2007. Concept evaluation of intersection decision support (ids) system interfaces to support drivers’ gap acceptance decisions at rural stop-controlled intersections. Transport. Res. Part F 10 (3), 208–228. Crundall, D., Van Loon, E., Underwood, G., 2006. Attraction and distraction of attention with roadside advertisements. Accid. Anal. Prev. 38 (4), 671–677. De Waard, D., Dijksterhuis, C., Brookhuis, K.A., 2009. Merging into heavy motorway traffic by young and elderly drivers. Accid. Anal. Prev. 41 (3), 588–597. De Waard, D., Kruizinga, A., Brookhuis, K.A., 2008. The consequences of an increase in heavy goods vehicles for passenger car drivers’ mental workload and behaviour: a simulator study. Accid. Anal. Prev. 40 (2), 818–828. Dingus, T.A., Klauer, S.G., Neale, V.L., Petersen, A., Lee, S.E., Sudweeks, J., Perez, M.A., Hankey, J., Ramsey, D., Gupta, S., Bucher, C., Doerzaph, Z.R., Jermeland, J., Knipling, R.R., 2006. The 100-Car Naturalistic Driving Study, Phase II – Results of the 100-Car Field Experiment. National Highway Traffic Safety Administration, Washington, D.C. Dukic, T., Ahlstrom, C., Patten, C., Kettwich, C., Kircher, K., 2012. Effects of electronic billboards on driver distraction. Traffic Inj. Prev. 14 (5), 469–476. Endsley, M.R., 1995. Toward a theory of situation awareness in dynamic-systems. Hum. Factors 37 (1), 32–64. Fitch, G.M., Soccolich, S.A., Guo, F., Mcclafferty, J., Fang, Y., Olson, R.L., Perez, M.A., Hanowski, R.J., Hankey, J.M., Dingus, T.A., 2013. The Impact of Hand-Held and Hands-Free Cell Phone Use on Driving Performance and Safety-Critical Event Risk. National Highway Traffic Safety Administration, US Department of Transportation, Washington, D.C. Fleury, D., Brenac, T., 2001. Accident prototypical scenarios, a tool for road safety research and diagnostic studies. Accid. Anal. Prev. 33 (2), 267–276. Gershon, P., Ben-Asher, N., Shinar, D., 2012. Attention and search conspicuity of motorcycles as a function of their visual context. Accid. Anal. Prev. 44 (1), 97– 103. Greene, W.H., 2002. NLOGIT: Version 3.0, Econometric Software. Ho, C., Spence, C., 2008. The Multisensory Driver: Implications for Ergonomic Car Interface Design. Ashgate, Hampshire. Horrey, W.J., Wickens, C.D., Consalus, K.P., 2006. Modeling drivers’ visual attention allocation while interacting with in-vehicle technoligies. J. Exp. Psychol.: Appl. 12 (2), 67–78. Kiefer, R.J., Hankey, J.M., 2008. Lane change behavior with a side blind zone alert system. Accid. Anal. Prev. 40 (2), 683–690. Klauer, S.G., Dingus, T.A., Neale, V.L., Sudweeks, J.D., Ramsey, D.J., 2006. The Impact of Driver Inattention on Near-Crash/Crash Risk: An Analysis Using the 100-Car Naturalistic Driving Study Data. National Highway Traffic Safety Administration, Washington, D.C. Konstantopoulos, P., Chapman, P., Crundall, D., 2010. Driver’s visual attention as a function of driving experience and visibility. Using a driving simulator to explore drivers’ eye movements in day, night and rain driving. Accid. Anal. Prev. 42 (3), 827–834. Koustanaï, A., Van Elslande, P., Bastien, C., 2012. Use of change blindness to measure different abilities to detect relevant changes in natural driving scenes. Transport. Res. Part F: Traffic Psychol. Behav. 15 (3), 233–242. Laberge, J.C., Creaser, J.I., Rakauskas, M.E., Ward, N.J., 2006. Design of an intersection decision support (ids) interface to reduce crashes at rural stop-controlled intersections. Transport. Res. Part C: Emerg. Technol. 14 (1), 39–56. Levin, L.D., Henriksson, T., Mårdh, P., Sagberg, S., 2009. Older Car Drivers in Norway and Sweden – Studies of Accident Involvement, Visual Search Behaviour, Attention and Hazard Perception. VTI, Norway. Maltz, M., Shinar, D., 2007. Imperfect in-vehicle collision avoidance warning systems can aid distracted drivers. Transport. Res. Part F: Traffic Psychol. Behav. 10 (4), 345–357. Marmeleira, J.F., Godinho, M.B., Fernandes, O.M., 2009. The effects of an exercise program on several abilities associated with driving performance in older adults. Accid. Anal. Prev. 41 (1), 90–97. Martens, M.H., Fox, M.R.J., 2007. Do familiarity and expectations change perception? Drivers’ glances and response to changes. Transport. Res. Part F 10 (6), 476–492. Mcintyre, S., Gugerty, L., Duchowski, A., 2012. Brake lamp detection in complex and dynamic environments: recognizing limitations of visual attention and perception. Accid. Anal. Prev. 45 (0), 588–599. Metz, B., Schömig, N., Krüger, H.-P., 2011. Attention during visual secondary tasks in driving: adaptation to the demands of the driving task. Transport. Res. Part F: Traffic Psychol. Behav. 14 (5), 369–380. Neale, V.L., Klauer, S.G., Knipling, R.R., Dingus, T.A., Holbrook, G.T., Petersen, A., 2002. The 100 Car Naturalistic Driving Study, Phase I – Experimental Design. National Highway Traffic Safety Administration, Washington, D.C. Neale, V.L., Perez, M.A., Lee, S.E., Doerzaph, Z.R., 2007. Investigation of driver– infrastructure and driver–vehicle interfaces for an intersection violation warning system. J. Intell. Transport. Syst. 11 (3), 133–142. NHTSA, 2008. National Motor Vehicle Crash Causation Survey: Report to Congress. National Highway Traffic Safety Administration, US Department of Transportation, Virginia. NHTSA, 2012. Blueprint for Ending Distracted Driving. National Highway Traffic Safety Administration, US Department of Transportation, Virginia. Salvucci, D.D., 2009. Rapid prototyping and evaluation of in-vehicle interfaces. ACM Trans. Comput.–Hum. Interact. 16 (2), 9.1–9.33. Shaheen, S.A., Niemeier, D.A., 2001. Integrating vehicle design and human factors: minimizing elderly driving constraints. Transport. Res. Part C 9 (3), 155–174. Summala, H., Pasanen, E., Rasanen, M., Sievanen, J., 1996. Bicycle accidents and drivers’ visual search at left and right turns. Accid. Anal. Prev. 28 (2), 147–153.

S.-H. Huang, J.-T. Wong / Accident Analysis and Prevention 74 (2014) 107–117 Tan, H.-S., Huang, J., 2006. Dgps-based vehicle-to-vehicle cooperative collision warning: engineering feasibility viewpoints. IEEE Trans. Intell. Transport. Syst. 7 (4), 415–428. Tijerina, L., Barickman, F.S., Mazzae, E.N., 2004. Driver Eye Glance Behavior During Car Following. National Highway Traffic Safety Administration, US Department of Transportation, Washington, D.C. Underwood, G., Chapman, P., Bowden, K., Crundall, D., 2002a. Visual search while driving: skill and awareness during inspection of the scene. Transport. Res. Part F: Traffic Psychol. Behav. 5 (2), 87–97. Underwood, G., Crundall, D., Chapman, P., 2002b. Selective searching while driving: the role of experience in hazard detection and general surveillance. Ergonomics 45 (1), 1–12. Underwood, G., Chapman, P., Brocklehurst, N., Underwood, J., Crundall, D., 2003. Visual attention while driving: sequences of eye fixations made by experienced and novice drivers. Ergonomics 46 (6), 629–646. 100-Car Data. Virginia Tech Transportation Institute.

117

Werneke, J., Vollrath, M., 2012. What does the driver look at? The influence of intersection characteristics on attention allocation and driving behavior. Accid. Anal. Prev. 45 (0), 610–619. Werneke, J., Vollrath, M., 2013. How to present collision warnings at intersections? –– a comparison of different approaches. Accid. Anal. Prev. 52 (0), 91–99. Wickens, C.D., Goh, J., Helleberg, J., Horrey, W.J., Talleur, D.A., 2003. Attentional models of multitask pilot performance using advanced display technology. Human factors. J. Hum. Factors Ergon. Soc. 45 (3), 360–380. Wickens, C.D., McCarley, J.S., Alexander, A.L., Thomas, L.C., Ambinder, M., Zheng, S., 2007. Attention–situation awareness (A-SA) model of pilot error. Human Performance Modeling in Aviation. CRC Press, London, pp. 213–239. Wong, J.-T., Huang, S.-H., 2011. A microscopic driver attention allocation model. Adv. Transport. Stud. Int. J. 53–64 (special issue). Wong, Huang, S.-H., 2013a. Attention allocation patterns in naturalistic driving. Acid. Anal. Prev. 58 (0), 140–147. Wong, J.-T., Huang, S.-H., 2013. Road Safety from the Perspective of Driver Attention Allocation. Annual Meeting of Transportation Research Board, Washington, D.C.