Head scanning behavior predicts hazard detection safety before entering an intersection

Abstract

Objective:

We conducted a driving simulator study to investigate scanning and hazard detection before entering an intersection.

Background:

Insufficient scanning has been suggested as a factor contributing to intersection crashes. However, little is known about the relative importance of the head and eye movement components of that scanning in peripheral hazard detection.

Methods:

Eleven older (mean 67 years) and 18 younger (mean 27 years) current drivers drove in a simulator while their head and eye movements were tracked. They completed two city drives (42 intersections per drive) with motorcycle hazards appearing at 16 four-way intersections per drive.

Results:

Older subjects missed more hazards (10.2% vs 5.2%). Failing to make a scan with a substantial head movement was the primary reason for missed hazards. When hazards were detected, older drivers had longer RTs (2.6s vs. 2.3s), but drove more slowly, thus safe response rates did not differ between the two groups (older 83%; younger 82%). Safe responses were associated with larger (28.8° vs. 20.6°) and more numerous (9.4 vs. 6.6) gaze scans. Scans containing a head movement were stronger predictors of safe responses than scans containing only eye movements.

Conclusion:

Our results highlight the importance of making large scans with a substantial head movement before entering an intersection. Eye-only scans played little role in detection and safe responses to peripheral hazards.

Application:

Driver training programs should address the importance of making large scans with a substantial head movement before entering an intersection.

Précis:

Larger and more numerous gaze scans before entering four-way intersections were associated with safer detections of peripheral motorcycle hazards approaching along the cross street. Scans containing a significant head component (larger scans) were a stronger predictor of detection likelihood and response safety than scans containing only eye movements (smaller scans).

1. Introduction

Early detection of hazards is crucial for executing a safe response when driving. This is particularly important at intersections where a large field of view must be scanned with eye and head movements in order to check for peripheral hazards approaching from both sides of the driver’s vehicle (Garber & Srinivasan, 1991; Ryan, Legge, & Rosman, 1998). This gaze scanning typically includes two types of scans: head+eye scans (scans containing both a substantial head and an eye movement component) and eye-only scans (scans containing no meaningful head component) (Savage, Zhang, Swan, & Bowers, 2020).

Although insufficient scanning has often been suggested as a reason for increased crash risk (Bao & Boyle, 2009; Romoser & Fisher, 2009; Romoser, Pollatsek, Fisher, & Williams, 2013), relatively few studies have directly investigated the relationship between scanning and the detection of hazards at intersections (Bowers, Bronstad, Spano, Goldstein, & Peli, 2019; Yamani, Samuel, Roman Gerardino, & Fisher, 2016). Using the safe environment of a driving simulator, Yamani et al (2016) found that older drivers who made fewer scans toward potential hazards after entering the intersection were more likely to crash than older drivers who made more scans. More recently, Bowers et al (2019) examined the relationship between head scanning and the detection of pedestrians at intersections, again in a driving simulator. Head scan magnitudes of older drivers were 15° smaller when the target pedestrian was not detected as compared to when it was detected. Thus the results of these two prior driving simulator studies suggest that both scan numbers and scan magnitudes are important in safe detection of hazards. However, each study had limitations. Yamani et al (2016) did not analyze scan magnitudes and measured only eye position within the scene without recording head position. Bowers et al (2019) tracked only head movements (and not eye movements) and the target was a stationary pedestrian positioned 90° to the left or right of the intersection that presented no immediate danger. In the current study we addressed these limitations by recording head (head-in-world) and eye (eye-in-head) movements, characterizing scanning in terms of scan numbers and magnitudes, and using moving motorcycles as ecologically-valid hazards at intersections. Furthermore, the current study extended prior research by also evaluating the relative importance of head+eye scans and eye-only scans to hazard detection likelihood and response safety at intersections.

From the age of 70 crash rates at intersections begin to increase (Bryer, 2000). An analysis of police accident reports suggested that over 50% of older drivers’ crashes at intersections were related to inadequate scanning behavior (Hakamies-Blomqvist, 1993). Consistent with these data, we previously reported that older drivers scanned less extensively on approach to intersections in a driving simulator (Savage et al., 2020). We did not, however, address the relationship between scanning behavior and detection of hazards. Three questions remain unanswered: 1) Does scanning behavior at an intersection predict detection likelihood and response safety for hazards at that same intersection?; 2) Does the reduced scanning behavior of older subjects mean that they are more likely to fail to detect hazards or have unsafe responses?; and 3) Is the contribution of the head or the eye components of scanning more important to detection and response safety?

2. Methods

2.2. Subjects

We analyzed data for 29 subjects, including 11 older subjects (≥ 60 years) and 18 younger subjects (20 to 40 years). Characteristics of study subjects are shown in Table 1. They were the same subjects as those reported in Savage et al. (2020). They were all current drivers who met the vision requirements for driving in Massachusetts, had at least two years of driving experience and no adverse ocular history (self-reported) of eye disease that might affect visual acuity or visual fields. Both older and younger subjects had visual acuity (Test Chart 2000 PRO, Hertfordshire, UK) and contrast sensitivity (MARS chart, The Mars Perceptrix Corporation, Chappaqua, NY) within the typical normal range (Elliott, Yang, & Whitaker, 1995; Haymes, Roberts, Cruess, Nicolela, LeBlanc, Ramsey, Chauhan, & Artes, 2006). There was no statistical difference in the number of miles driven per year between the two age groups (W= 76.5; p= .12; Table 1). The study followed the tenets of the Declaration of Helsinki and was approved by the institutional review board at the Schepens Eye Research Institute. Informed consent was obtained from each subject.

Table 1.

Characteristics of study subjects

Factor	Older (N=11)	Younger (N=18)
Age [years], mean (SD)	67.5 (6.7)	26.5 (5.9)
Male [N] (%)	7 (64)	9 (50)
Binocular visual acuity [logMAR1], mean (SD)	0.00 (0.07) 20/20	−0.07 (0.05) 20/17
Binocular contrast sensitivity [log], mean (SD)	1.70 (0.09)	1.78 (0.07)
Annual mileage [miles], median	2860	1170
Driving experience [years], median	49	6

¹logMAR – Logarithm of the minimum angle of resolution

2.3. Materials

2.3.1. Driving simulator and head and eye tracker

The driving simulator (LE-1500, FAAC Corp, Ann Arbor, MI) consisted of five 42-inch LCD monitors yielding approximately 225° horizontal field of view. The central screen provided the view through the windshield, while the screens to the left and right provided the view through the side windows. The dashboard, including a speedometer, was displayed at the bottom of the central monitor. The controls and dashboard resembled a fully automatic transmission Ford Crown-Victoria along with a motion base seat with 3 degrees of freedom (Figure 1). Driving simulator data were collected at 30 Hz, including the position of the subject’s vehicle, its speed and heading, as well as information about other scripted vehicles in the virtual world.

Figure 1.

FAAC driving simulator with SmartEye 6-camera head and eye tracking system (yellow circles)

Head and eye movement data were recorded using a remote, digital 6-camera tracking system at 60 Hz (SmartEye Pro Version 6.1, Goteborg, Sweden, 2015 - Figure 1 yellow circles). The system tracked head and eye movements across approximately 180° (90° to the left and right), enabling us to capture the large scans that drivers made before entering an intersection. Matlab software (Mathworks, R2015a) was developed to post-process the merged simulator and head and eye data to generate plots and output variables used in the statistical analysis.

2.3.2. Driving scenarios

A route with 42 intersections was selected. Events along the route were scripted using Scenario Toolbox software (version 3.9.4. 25873, FAAC Incorporated). The route was set in a light industrial virtual world including both city and residential areas with the roads on a grid system. The posted speed limit was 35 mph. The route included oncoming and cross traffic. Subjects drove the route twice, once guided by GPS instructions (“turn right/left at next intersection”) and once following a lead car since the method used to guide drivers through the virtual world could affect scanning behaviors and also detection performance.

In this paper we report data for 16 of the 42 intersections. These 16 intersections had motorcycle hazards approaching along the cross-street from either the left or right side. They were all 4-way intersections with a range of traffic control devices, cross traffic, and intersection maneuvers (see Table 2 for details). They were selected to provide a variety of intersections as might be encountered in real world driving. In addition to the 16 motorcycles at intersections, there were seven motorcycle distractors, which did not appear at intersections, to avoid anticipation of motorcycles appearing only at intersections. Data for an additional 15 of the 42 intersections without any hazards were reported previously (Savage et al. 2020).

Table 2.

Inventory of all 4-Way Intersections with Motorcycle Hazards

INT #	Traffic control device on the subjects’ approach	Other Traffic	Intersection maneuver	MC Side
1	Stop	Car left	Straight	Right
2	Stop	Van left / Car ahead	Straight	Right
3	Stop	None	Straight	Right
4	Stop	Car right	Straight	Left
5	None	None	Straight	Left
6	None	Car right	Straight	Left
7	None	None	Straight	Left
8	Stop	None	Straight	Right
9	Yield	None	Right	Left
10	None	Bus left / Bus ahead	Straight	Right
11	Traffic Light1	None	Right	Left
12	Traffic Light1	Bus left	Right	Left
13	None	None	Straight	Right
14	None	None	left	Left
15	Yield	None	Left	Right
16	Stop	Van left	Straight	Right

^1.The traffic light turned to red when the subject’s vehicle was about 20 m from the stop line

All 16 intersection motorcycle events had the same design (Figure 2), which simulated a motorcyclist who was speeding through an intersection and failed to perceive the subject’s vehicle. The motorcyclist did not obey any control devices and did not follow normal priority rules. When the subject’s vehicle was at 30 meters from the intersection, the motorcycle was triggered to appear 60 meters away from the intersection along the cross street to the left (n = 8) or right (n = 8). The motorcycle then traveled at a constant speed of 45mph (or 20.1 m/s), well in excess of the 35 mph speed limit, until it entered the intersection, about 3 s after being triggered. These parameters were determined empirically based on measurements in a pilot study of subjects’ driving behaviors (speed, deceleration) when approaching intersections used for the experimental drives. They were selected to create a scenario in which the motorcycle represented a realistic hazard for a range of subjects’ intersection approach behaviors and deceleration profiles. However, to avoid any potential psychological stress caused by a collision between the motorcycle and the subject’s vehicle, the motorcycle was programmed to disappear just before entering the potential collision zone. Subjects were able to drive at whatever speed they were comfortable with up to a maximum of 35 mph (speed was capped at 35 mph or 15.7 m/s). If the subject’s vehicle was travelling at this maximum speed their vehicle would reach the intersection 1.8 s after triggering the motorcycle (or 1.2 s before the motorcycle). However, because subjects typically slowed down on approach to intersections, it took longer than this minimum time to reach the intersection and the motorcycle always represented a potential hazard.

Figure 2.

Schematic representation of motorcycle intersections (not drawn to scale). The red square represents the collision zone between the subject’s vehicle and the scripted motorcycle hazard.

2.4. Procedure

The driving simulator session began with two acclimation drives. The first took place on a rural highway without other vehicles. The second was in the same city as the experimental drives but without other vehicles. For these two acclimation drives, subjects were given as much time (10-20 minutes) as they needed in order to become comfortable maneuvering the simulated vehicle.

After the acclimation drives, we adjusted the six cameras of the SmartEye tracking system and calibrated the subject’s head position. Next, subjects completed a practice drive which included all the elements (motorcycle hazards, oncoming and cross traffic) of the experimental drives. During the practice drive the SmartEye system automatically built a profile of the subject and tracked their facial features. Eye position was calibrated after the practice drive by means of a five point calibration procedure on the center screen.

After calibration, subjects completed two experimental drives (GPS and Lead Car; see Figure 3 for example screenshots) in counterbalanced order. Eye and head data were collected and analyzed for these two experimental drives. Each drive typically lasted between 10 and 15 minutes depending on the subjects’ driving speed. Subjects were instructed to drive as they normally would, follow all rules of the road and to press the horn as soon as they saw a motorcycle. Occasionally subjects forgot to press the horn and responded verbally to indicate detection of a motorcycle. Any such events were noted by the experimenter. If a subject stepped out of the simulator for a break, re-calibration of the SmartEye tracker was performed before the second experimental drive.

Figure 3.

Examples of the scene on the center screen for each of the guidance methods within the light industrial city, taken from the point of view of the subject’s vehicle. Left panel shows the view when guided by GPS instructions. The right panel shows the view when following a lead car. Screenshots were taken from the center screen of the driving simulator

In the GPS drive, auditory navigation instructions were delivered when the subject’s vehicle was approximately 70 meters from an intersection. In the Lead Car drive, subjects followed a car that was scripted to drive at 35 mph. The Lead Car made periodic stops to ensure that subjects did not lose sight of it. Subjects were instructed to drive as they normally would and follow all normal traffic rules. In the Lead Car drive they were instructed to follow the Lead Car at a safe distance, as they would when following a friend’s car in an unfamiliar city. Subjects were not, however, given any specific instructions about how to scan.

2.5. Quantifying detection performance and response safety

Events where the subject responded verbally rather than by pressing the horn (1.9% of all motorcycle intersection events) were manually entered as seen and included in analyses of miss rates. However, they were not included in analyses of reaction times and response safety since there was no horn press response.

2.5.1. Miss rates

Miss rates were the number of intersection motorcycles that were missed (i.e., events for which there was no horn-press response) expressed as a percentage of the total number of motorcycle intersection events.

2.5.2. Reaction times when a hazard was detected

Reaction time was computed as the time between when the motorcycle first became visible and the horn-press response.

2.5.3. Safety of responses when hazard was detected

Two metrics were used to quantify the safety of responses, each capturing a separate aspect of safety. For both metrics, the hypothetical collision point was used, given that the motorcycle disappeared prior to entering the collision zone. The first metric was the post-encroachment time (PET), which is defined as the time difference between the first road user leaving the collision zone and the second road user entering it (Allen, Shin, & Cooper, 1978). It represents the safety margin or the extent to which the two road users miss each other. The lower the PET, the more critical the situation. A PET of less than 1 second is considered a dangerous situation in real-world traffic conflicts (Hupfer, 1997; Kraay, van der Horst, & Oppe, 2013; Zangenehpour, Strauss, Miranda-Moreno, & Saunier, 2016; Johnson, Laureshyn & De Ceunynck, 2018), and was therefore classified as unsafe in the current study.

The second metric was the deceleration-to-safety time (DST; (Hupfer, 1997)). This metric describes the minimum deceleration rate (in m/s²) needed to avoid a collision; i.e., to stop the subject’s car from entering the collision zone before the motorcycle leaves the collision zone. The higher the DST, the more critical the event. In real-world situations, a deceleration rate greater than 4 m/s² is considered potentially dangerous (Hupfer, 1997), and was therefore classified as unsafe in the current study. Decelerating at a rate of 4 m/s² could result in a driver either becoming a hazard to themselves (if the road conditions were wet or slippery) or to the driver in the vehicle behind them. The required deceleration was calculated using the following equation:

$$\text{Deceleration}=\frac{2(S_{car}-V_{car}{t}_{mc})}{t_{car}^2}$$

S_car is the distance in meters between the location of the subject’s vehicle at the horn-press and the location where their vehicle will enter the collision zone. V_car is the velocity of the subject’s vehicle in meters per second at the time of the horn-press. t_car is the time in seconds it would take the subject’s vehicle to travel from its current location (given their speed at the time of the horn press) to the point at which it enters the collision zone. t_mc is the time in seconds it would take for the motorcycle hazard to exit the collision zone, thereby no longer being a hazard.

In summary, when hazards were detected, the response was classified as unsafe if: 1) the subject’s vehicle entered the intersection before the motorcycle (and the motorcycle might have crashed into them); 2) the PET was less than 1 second; or 3) the DST was greater than 4 m/s², or any combination of these situations.

2.6. Quantifying gaze scanning behavior

Gaze scans were analyzed from 100 m before the intersection until the point at which the subject’s car first entered the intersection (0 m), defined as crossing the white line at controlled intersections or where the white line would have been at intersections without a control device. Thus, we included all scans made on approach to the intersection as well as any scans that were made when the subject’s vehicle was stationary and the subject was scanning for hazards before entering the intersection. We did not evaluate scans made after entering the intersection.

We define a gaze scan as the whole series of lateral movements that take the eyes the furthest to the right or to the left from straight ahead (0°). Gaze scans could be composed of a single saccadic eye movement or multiple eye movements, either with or without an associated head movement (see Figure 4). Gaze scans were automatically detected using a custom algorithm; see Swan et al. (2021) for full details. In brief, the algorithm first marked saccades (velocity above 30 °/sec, greater than 1° eccentricity, and longer than 0.33s duration) and then merged sequential saccades that were headed in the same direction (i.e., towards the left or right), on the same side of the straight ahead gaze position (i.e., on the left or right side), and were within 400ms of each other. Merging saccades was necessary given that many gaze scans, especially large gaze scans, were composed of multiple saccades (see Figure 4 – right panel, G2; and Figure S1 in the supplement). Each gaze scan had a corresponding start and end eccentricity, with the difference being the scan magnitude. For each gaze scan, there was a corresponding head movement that was estimated by setting the start of the head movement as the start of the gaze scan and the end of the head movement as the local maximum of head eccentricity around the end of the marked gaze scan (Savage et al., 2020). Only gaze scans above 4° (more than four times the manufacturer’s accuracy under ideal conditions) headed away from straight ahead (0°) were used in analyses.

Figure 4.

Examples of lateral gaze (blue) and head (red) movements on approach to an intersection. The left panel serves as a reference, with the car’s travel path being left to right (i.e. increasing in time). Movements towards the left have a negative eccentricity and movements towards the right have a positive eccentricity. The right panel includes the gaze scans (green) detected using our custom algorithm. Gaze scan G1 is an eye-only scan comprising 1 eye saccade, G2 and G3 are eye-only with 2 eye saccades, G4, G5, and G6 are eye-only with one eye saccade, G7 is eye-only with 2 eye saccades, G8 is eye-only with one eye saccade, G9 is a head+eye scan comprised of a large lateral head rotation and eye saccade, and G10 is eye-only with one eye saccade.

Gaze scans were classified into two major categories (Savage et al., 2020): 1) scans which comprised predominately eye movement only (“eye-only” scans); and 2) scans which contained both a substantial head and eye movement component (“head+eye” scans). The classification was based on the magnitude of the head scan component of each gaze scan (Table 4), as implemented in prior driving simulator studies (Bowers et al., 2014; Bowers et al., 2019). The term “all-gaze” scans is used when data are pooled across eye-only and head+eye scans.

Table 4.

Classification of gaze scans based on the size of the head movement component and the subject’s distance to the intersection.

Distance to Intersection [m]	Size of Head movement1 [°]	Classification of Scan
100 – 50	≥ 4	head+eye
	< 4	eye-only
50 – 20	≥ 6	head+eye
	< 6	eye-only
20 – 0	≥ 10	head+eye
	< 10	eye-only

¹Thresholds were the same as those used to define a head scan in prior driving simulator studies (Bowers et al., 2014; Bowers et al., 2019)

2.8. Statistical Analyses

Prior to conducting statistical analyses we excluded single intersections on a per-subject basis where there was excessive noise in the gaze data for that intersection (removed ~7% of all intersections). Ultimately 856 intersections and 7538 scans were included in analyses. For the analyses of continuous numerical data, Linear Mixed Models (LMMs) were constructed in RStudio of the R statistical programming environment (Version 3.6.2 – R CoreTeam, 2019). Categorical outcome variables were analyzed by means of Generalized Linear Mixed Models (GLMMs).

Our first set of analyses evaluated the effects of age (older vs. younger) on the three motorcycle detection measures (miss rates, reaction times and safe response rates). We created models in which we entered age (younger vs. older) as a fixed factor. Guidance type (GPS vs. Lead Car) was not included as a factor because preliminary analyses found no significant main effects of guidance type and no significant interactions with age (see supplementary materials). For all of our models we entered the unique event (intersection) number as a random factor to account for any variance contributed by the individual intersections as well as a random effects structure for subject to account for the variability contributed by individual differences.

The second set of analyses evaluated whether scanning behavior was predictive of the safety of responses when hazards were detected using the following metrics: the magnitude and number of all-gaze scans, head+eye scans, and eye-only scans. We created a GLMM in which we entered the safety of responses (safe or unsafe) as the outcome variable and the magnitude or number of scans as the predictor variable. Age group (older vs. younger) was also added as a fixed factor to these models.

Magnitudes of eye-only scans and all-gaze scans were normalized with a log2 transform prior to entering them into our models. All other continuous numerical outcome variables were roughly normally distributed. Outliers greater than 3 standard deviations from the mean (in transformed units) were removed. When reporting the data for normalized variables, we transformed them back to their raw unit format for ease of understanding.

P-values for main effects were estimated by means of the lmerTest package (Kuznetsova, Brockhoff, & Christensen, 2017). P-values for any interactions between age and safety were calculated by model comparisons. We compared the simplest form of each model (with all interactions removed) with the same model plus the interaction of interest. The interaction model and baseline model were then compared using an analysis of variance (ANOVA), with the resulting p-value derived from our χ² statistic representing the significance of the interactions of interest.

3. Results

The main results are reported below. Additional analyses addressing detection types (fixational or peripheral), effects of guidance (lead car vs. GPS) and control device (stop sign vs. no sign) are provided in the supplementary materials.

3.1. Detection performance

3.1.1. Missed hazards

Older subjects had significantly higher miss rates than younger subjects (10.2% vs. 5.2%; χ²(1)= 6.84; p= .009).

3.1.2. Detected hazards

When hazards were detected, RTs were slower for older than younger subjects (means 2.63 s vs. 2.25 s; β= .−.33, SE= .14; t= −2.37; p= .027), and for unsafe than safe responses (2.7 s vs. 2.3 s; β= .18, SE= .08; t= 2.28; p= .023). There was no interaction between age and safety for the RT data, χ²(1,6)= 2.49; p= .11. In contrast, rates of safe responses did not differ between the two age groups (older 83% vs. younger 82%; β= .−.09, SE= .63; z= .15; p= .88).

Analysis of speed at the time of horn press response revealed that both older and younger subjects drove slower when responses were considered safe (mean= 13.93 mph) as opposed to unsafe (mean= 25.47 mph), β= 12.6, SE= 1.6; t= 8.03; p< .001, and older subjects drove significantly slower (mean= 13.7 mph) than younger subjects (mean= 17.1 mph), β= 4.4, SE= 1.66; t= 2.65; p= .013. We also found a significant interaction between age and safety χ²(1,6)= 5.46; p= .02. This interaction came about as the difference between safe and unsafe responses was larger for younger (difference: 12 mph) than older drivers (10.5 mph).

3.2. Scanning behavior and missed hazards

When the hazard was missed, subjects failed to make a head+eye scan toward the motorcycle in the majority (93%) of missed events but failed to make an eye-only scan in that direction for only a minority of missed events (15%). In other words, when a hazard was missed subjects frequently made eye-only scans toward the hazard but rarely made a head+eye scan. When expressed as a proportion of all events, rates of failing to make a head+eye scan and failing to detect the hazard were higher for older than younger subjects (9% vs. 5%; χ²(1) = 5.78; p= .016).

3.3. Scanning behavior and safety of responses for detected hazards

3.3.1. Number of scans

Subjects made significantly more head+eye scans when the response was safe (3.8) as compared to unsafe (1.8), β= −.84, SE= .14; z= −6.07; p< .0001 (Figure 5, left panel), significantly more eye-only scans when the response was safe (5.6) as compared to unsafe (4.8), β= −.17, SE= .08; z= −2.25; p= .024 (Figure 5, middle panel), and significantly more all-gaze scans when their responses were safe (9.4) as compared to unsafe (6.6), β= −.36, SE= .062; z= −5.91; p< .0001 (Figure 5, right panel).

Figure 5:

Average number of scans per intersection when motorcycles were detected for head+eye scans (left), eye-only scans (middle) and all-gaze scans (right). Data are split by safe and unsafe responses for older and younger subjects. Error bars represent the SEM

We found no main effects of age for numbers of head+eye scans, β= −.016, SE= .15; z= .11; p= .91, eye-only scans, β= −.2, SE= .14; z= −1.43; p= .15, and all-gaze scans β= −.09, SE= .08; z= −1.5; p= .25. However, there was a significant interaction between safety and age for head+eye scans, χ²(1,6)= 4.08; p= .043. The difference in the number of scans between older and younger subjects was slightly greater when detections were unsafe (difference= .59 scans fewer for older subjects) than when they were safe (difference= .37 scans). For eye-only and all-gaze scans numbers, there were no significant interactions between age and safety (χ²(1,6)= .02; p= .89 , and χ²(1,6)= .71; p= .4, respectively).

3.3.2. Scan magnitudes

The magnitude of head+eye scans was significantly larger when subjects’ responses were safe (53.7°) as compared to unsafe (46.7°), β= −10.03, SE= 2.91; t= −3.45 p< .001 (Figure 6, left panel). However, eye-only scan magnitudes did not differ between safe (10.7°) and unsafe responses (10.2°), β= .01, SE= .078 t= .13; p= .89 (Figure 6, middle panel). The magnitude of all-gaze scans was also larger when the response was safe (28.8°) than when it was unsafe (20.6°), β= −.28, SE= .06; t= −4.57; p< .001 (Figure 6, right panel).

Figure 6:

Average scan magnitudes when motorcycles were detected for head+eye scans (left), eye-only scans (middle) and all-gaze scans (right). Data are split by safe and unsafe responses for older and younger subjects. Error bars represent the SEM

Older subjects made significantly smaller eye-only scans than younger subjects (9.6° vs. 11.3°); β= .023, SE= .08; t= 2.93; p= .007, and significantly smaller all-gaze scans (23.8° vs. 30°), β= −.15, SE= .07; t= −2.12; p= .04. For head+eye scan magnitudes, the main effect of age did not reach statistical significance β= 4.16, SE= 2.18; t= 1.91 p= .07. We found no interactions between age and safety for head+eye scan magnitudes, χ²(1,6)= 2.54; p= .11, eye-only scan magnitudes χ²(1,6)= .88; p= .35, and all-gaze scan magnitudes, χ²(1,6)= .49; p= .48.

4. Discussion

4.1. The effects of age on detection performance

Both older and younger subjects detected the majority of motorcycles. However, older subjects missed a significantly higher percentage of motorcycles than younger subjects (10% vs. 5%). In the real world, any failure to detect a motorcyclist speeding through an intersection could result in the most serious of traffic conflicts, a collision. Thus the finding of a higher rate of missed hazards in the older group is concerning and is consistent with reports of higher rates of at-fault collisions in older drivers at intersections (Garber & Srinivasan, 1991; Pai, 2011; Preusser, Williams, Ferguson, Ulmer, & Weinstein, 1998).

When motorcycles were detected, older subjects had longer reaction times than their younger counterparts, but the rate of safe responses did not differ between the two age groups because older subjects drove more slowly. Our calculations of response safety took account of the speed and distance of both road users at the time of the horn press. As such, driving more slowly increased the time window in which subjects were able to make a safe response. Older drivers have previously been reported to have overall slower driving speeds in simulated (Doroudgar, Chuang, Perry, Thomas, Bohnert, & Canedo, 2017; Zhang, Bowers, & Savage, 2020) and on-road driving (Porter, & Whitton, 2002; Horberry et al., 2004), as well as slower speeds when approaching intersections in simulated (Caird, Chisholm, Edwards, & Creaser, 2007) and on-road driving (Liu, 2007). However, even though safe responses rates of older drivers did not differ from those of younger drivers when motorcycles were detected, they still missed more motorcycles, which is the most critical kind of unsafe event with the greatest potential for adverse consequences in on-road driving.

4.2. The relationship between scanning behavior and missed hazards

In the majority of events when a motorcycle was missed, there was no head+eye scan toward the motorcycle (on average 93% of missed detections), but there was usually at least one eye-only scan in that direction. These results clearly highlight the importance of making at least one large scan with a substantial head movement component in each direction before entering the intersection. Rates of failing to make a head+eye scan before entering the intersection and failing to detect the hazard were higher for older than younger subjects, especially at intersections without signage (see supplement S4.1), suggesting one reason for higher collision rates of older drivers in real-world driving.

4.3. The relationship between scanning behavior and response safety for detected hazards

The number and magnitude of all-gaze scans were significant predictors of response safety when hazards were detected. Subjects made significantly more all-gaze scans (9.4 vs. 6.6) and significantly larger all-gaze scans (28.8° vs 20.6°) when responses were safe as compared to unsafe. Consistent with our prior study of scanning behaviors at intersections without motorcycle hazards (Savage et al., 2020), we found that older subjects made significantly smaller all-gaze scans than younger subjects but did not differ from younger subjects for the number of all-gaze scans.

As a next step we wanted to determine whether head+eye or eye-only scans were more important for safe responses when the hazard was detected. Both older and younger subjects made more head+eye scans when responses were safe (3.8) as compared to unsafe (1.8), and the size of head+eye scans was greater when responses were safe (53.7°) as compared to unsafe (46.7°). Subjects also made more eye-only scans when responses were safe (5.6) as compared to unsafe (4.8). However, the magnitude of eye-only scans did not differ between safe (10.7°) and unsafe (10.2°) responses. The difference in eye-only scan numbers between safe and unsafe responses was less than 1 scan per intersection. Conversely, the differences between safe and unsafe responses for head+eye scan magnitudes (9.1°) and numbers (2.3 scans) were much greater and have both statistical and ecological significance. Thus our results strongly suggest that head+eye scans were more crucial than eye-only scans to response safety when motorcycles were detected.

4.4. Limitations

Previous research has demonstrated that the risk of being involved in collisions at intersections starts to increase from the age of 70 years (Bryer, 2000). However, the average age of our older subjects was only 67 years. Thus we might have found greater age-related differences in safe response rates if we had included more subjects over 70 years of age. Nevertheless, the findings of higher miss rates and smaller all-gaze scans in older subjects are in the expected direction.

In the current study we instructed our subjects to press the horn as soon as they detected a motorcycle. This task is conceptually quite different from a typical hazard perception task (Savage, Potter, & Tatler, 2013). Because the target in our current study was always the same, subjects most likely would have simply pressed the horn as soon as they saw a motorcycle, without needing to make a judgment about the level of hazard posed by the approaching motorcycle. In contrast to this, in a hazard perception task, detecting a hazard relies on knowledge of the rules of the road as well as contextual information which needs to be recalled and processed before a hazard judgment can be made. Therefore, the processing of potential hazards is most likely more cognitively demanding than was the case for our motorcycle detection task.

The instructions to press the horn as soon as a motorcycle was detected may have resulted in more active visual scanning behaviors than would be the case in real-world driving. Moreover, subjects experienced a relatively high prevalence of motorcycles in each drive. Both these factors may have increased the likelihood and speed of detection (Beanland, Lenné & Underwood, 2014). Thus scanning deficits and detection deficits might be greater in on-road driving than in our driving simulator study. On the other hand, feelings of simulator discomfort (which were higher in the older than the younger group (Savage et al., 2020)) may have resulted in less active scanning than in real-world driving.

Intersections in the real world contain many different types of hazards, which we did not test in our current experiment. Our paradigm focused on detection of a peripheral hazard which appeared at a relatively large eccentricity (about 60°). The results suggest that large head+eye scans are crucial for detection of this kind of hazard. However, other hazards may appear at smaller eccentricities e.g., an oncoming car which suddenly makes a turn without signaling. Eye-only scans may be important for detection of such hazards. The next step will be to apply our methodology to investigating the relationship between scanning and responses to hazards for different types of ecologically-valid hazards at intersections.

4.4. Conclusions and practical implications

Older subjects failed to detect significantly more motorcycles than younger subjects, primarily because they did not make a head+eye scan in the direction of the hazard, which may place them at increased risk for collisions in real-world driving. When a motorcycle was detected, older subjects were slower to respond to the motorcycle but were not less safe as they drove slower than their younger counterparts.

Our approach of analyzing head+eye and eye-only scans separately provides new insights into the relative importance of head and eye scanning in detection of peripheral hazards before entering four-way intersections as well as response safety when detected. In particular, failure to make a large head+eye scan in the direction of the motorcycle was the primary reason for missed detections, while making fewer head+eye scans and smaller head+eye scans were strongly associated with unsafe responses when motorcycles were detected. In contrast, eye-only scans played little role in the detection of peripheral hazards, though they may be important for detection of other kinds of hazards.

It is well established that older persons have age-related reductions in neck flexibility that reduce maximum head rotation extent (Isler et al. 1997, Dukic and Broberg 2012, Chen et al. 2015). Reduced neck flexibility likely contributed to the finding that older subjects were more likely to fail to make a head+eye scan and made smaller head+eye scans than younger subjects. We previously reported that older subjects made more eye-only scans than younger subjects (Savage et al., 2020), which might be an attempt to compensate for making fewer head+eye scans. However, even if older drivers make more eye-only scans, the results of the current study suggest that eye-only scans are not sufficient for detection of peripheral hazards.

Prior studies have demonstrated the effectiveness of training older drivers to make a scan to check for hazards after entering an intersection (Romoser & Fisher, 2009; Romoser, 2013). Our findings suggest that training programs for older drivers should also address the need to make at least one large head+eye scan in each direction before entering four-way intersections. Given limitations in neck flexibility, shoulder movements may be needed. The results also highlight that younger drivers may need training in the importance of head+eye scanning before entering an intersection. Designing and evaluating such training programs are topics for future studies.

5. Key Points.

• Older drivers missed more hazards than younger drivers

• Failing to make a large scan was the main reason for failing to detect hazards

• When motorcycles were detected, older drivers’ responses were not less safe because they drove more slowly

• Safe responses were associated with larger and more numerous gaze scans in both younger and older drivers

• Gaze scans with substantial head component were stronger predictor of response safety than eye only scans

Biographies

Steven W. Savage is a postdoctoral fellow at Schepens Eye Research Institute of Massachusetts Eye and Ear, Department of Ophthalmology, Harvard Medical School in Boston, MA, USA. He received his PhD in psychology from the University of Dundee, Scotland in 2015.

Lily Zhang is a research assistant at Schepens Eye Research Institute of Massachusetts Eye and Ear in Boston, MA, USA. She received her MSc in mechanical engineering from the University of Minnesota in 2009.

Garrett Swan is a postdoctoral fellow at Schepens Eye Research Institute of Massachusetts Eye and Ear, Department of Ophthalmology, Harvard Medical School in Boston, MA, USA. He received his PhD in cognitive psychology from the Pennsylvania State University in 2017.

Alex R. Bowers, PhD is an associate scientist at Schepens Eye Research Institute of Massachusetts Eye and Ear and an associate professor in the Department of Ophthalmology, Harvard Medical School in Boston, MA, USA. She received her PhD in vision rehabilitation, from Glasgow Caledonian University, Scotland in 1998.