Mesopic and glare driving performance in a driving simulator

Abstract

Objective:

Difficulties in night driving due to deficits in visual function is a common problem among older drivers. Signage, hazards and road markings can be more difficult to identify for those with presbyopia, especially in nighttime conditions that include glare from headlights. This study evaluated the visual function and driving abilities of participants in various lighting conditions to find efficient and effective testing procedures for predicting how older individuals will perform in night driving conditions.

Methods:

A driving simulator study was conducted to investigate the relationship between driving performance measures and visual function measures in different lighting conditions. Participants were examined in a laboratory setting to measure visual acuity and contrast sensitivity in bright light, dim light and dim light with glare conditions. The participants were asked to drive in a driving simulator in simulated day, night and night with glare conditions. Analysis using a linear mixed model was performed across the lighting conditions.

Results:

There are statistically significant relationships between the experimental variables, with age being an important factor. Age was found to have a statistically significant effect on the standard deviation of lateral position, which is a measure of vehicle control. Age was also found to have a statistically significant effect on reaction time and accuracy of a secondary identification task. Visual acuity and contrast sensitivity were found to have a statistically significant effect on mean velocity.

Conclusions:

Results suggest that visual acuity alone is insufficient to fully describe visual function for predicting performance in a driving task, but the ideal variables for prediction of performance in a night driving situation are not clear.

Introduction

Visual perception is a necessary ability to drive a vehicle safely, but many factors can degrade that ability. Distraction may cause a driver to not only lose concentration on the driving task but also to move their gaze away from the roadway. Lighting conditions change during the course of a day. Personal visual deficits can degrade visual acuity and other aspects of visual function. Presbyopia (age-related loss of near visual acuity) exacerbates the challenges of nighttime driving. Even correction of presbyopia can affect visual performance while driving (Chu et al. 2010). Some reduced visual acuity and contrast sensitivity is common as a person ages, and lighting conditions that challenge a person’s abilities to drive safely can be even more of a problem for older drivers.

Previous investigations into the relationship between mesopic visual function and driving performance revealed that photopic visual acuity measures alone are not sufficient to assess driving ability (Gruber et al. 2013). Jones et al. showed a moderate correlation between contrast sensitivity and the distance of hazard detection in a static base driving simulator (Jones et al. 2022; Ungewiss et al. 2022). In the work by Wood and Owens, contrast sensitivity, but not visual acuity, was found to have a significant effect on recognition of potential roadway hazards on a closed course driven at nighttime (Wood and Owens 2005). These study findings show the insufficiency of photopic visual acuity as a sole measure of night driving proficiency.

Nighttime mesopic driving conditions make it more difficult to identify potential hazards and to read and respond to signage (Wood 2020). However, low lighting conditions from nighttime and twilight driving are not the only source of visual difficulty. The glare from headlights in otherwise dim lighting conditions is an unpleasant reality common to the experience of those who are driving during nighttime hours. Kimlin et al. reported that older drivers perform significantly worse with hazard and pedestrian detection in mesopic conditions with glare than in conditions without glare (Kimlin et al. 2017). However, this finding was based on a targeted subset of older drivers that report night driving difficulty. Although some older drivers decide to reduce their driving in more difficult lighting conditions due to their perceived lowered visual abilities, many do not. An over-estimation of their personal ability to drive at night can lead to more crashes and crash-related injuries and fatalities (Okonkwo et al. 2008; Schlueter et al. 2023).

Glare conditions are especially disruptive to many older drivers, and a driving simulation environment is an ideal platform to investigate the relationship between functional vision measures and driving performance. The overall objective of this research is to identify a feasible suite of visual function tests that predict the ability of an older adult to drive in various lighting conditions. As a step towards that goal, we measured functional visual measures and driver performance in three different lighting conditions. We present statistical analyses of the effects of visual acuity and visual contrast sensitivity in older adults on performance in a driving simulation environment across three lighting conditions: daytime, nighttime, and nighttime with glare.

Methods

Participants were from the group of individuals aged 60 years or older who drove at least once per week, had a valid drivers license and had no history of motion sickness. Additional eligibility included English speaking, without relying on visual aid devices beyond habitual correction (e.g., bioptic lens), and without conditions that impede safe driving, such as cognitive impairment. Participants were recruited via existing research databases (e.g., ResearchMatch, StudySearch) and social media (e.g., Twitter and Instagram). Interested individuals were directed via social media postings or email links to complete an initial eligibility screening survey. If eligible, signed consent was obtained and in-person visual function and simulated driving assessments were scheduled, respectively, with the two assessments one month apart. Protocol details and preliminary data from this study were reported previously (Yang et al. 2024). This article provides a more complete analysis of the modeled relationship between visual function and driving assessment. The study was approved by institutional review boards at authors’ institutions, protocol #STUDY00000461.

Vision assessment

Visual acuity (VA), contrast sensitivity function (CSF) and visual field map (VFM) measurements were obtained from the participants using the qVA, qCSF and qVFM procedures. The visual acuity assessment consisted of 20 trials. Each trial contained three letters with a high contrast (Figure 1 (a)). Subjects were instructed to report all three letters. The qVA algorithm (Lesmes and Dorr 2019) was used to select the letter size from trial to trial. Each CSF measurement comprised 30 trials. Stimuli were bandpass-filtered letters (Figure 1 (b)) with size and contrast varied across trials using the qCSF algorithm (Hou et al. 2015). The task was to identify the displayed letters. The area under log CSF (AULCSF) was employed as a summary measure of CSF. The qVFM test, containing 120 trials, measured a 48°×48° visual field (Xu et al. 2019). Each trial included a beep sound, and a potential target (a light disc) with a circle cue (Figure 1 (c)). Subjects were asked to report presence or absence of the target while fixating at the center of the display. Both target location and luminance were adaptively varied across trials. The volume under the surface of the VFM (VUSVFM) after normalization served as a summary metric for VFM.

Figure 1:

Example images used in visual assessments.

VA and CSF values were obtained under three different lighting conditions in a laboratory setting: photopic (bright light; background luminance $\ge 9cd/m^2$), mesopic (low-light; $0.1cd/m^2\le$ background luminance $\le 1cd/m^2$) and mesopic with glare (low-light with an additional light source directed at the face). VFM was measured only in the photopic and mesopic settings and not used in the current analysis. A summary metric, the Area Under the Log CSF (ALUCSF), was used to quantify the CSF. For the glare condition, the light source, a Fiilex V70 lamp with a dome diffuser, was positioned at eye level and provided an illuminance level of 305 lux at a 32-degree angle from the left. All subjects were required to undergo a dark adaptation period of at least 5 minutes before the photopic and mesopic tests could commence. Dark adaptation is a standard technique to let the sensitivity of the participant’s visual system adapt to the light levels in the testing environment (Lamb and Pugh 2004).

Driving assessment

Participants drove through three scenarios in a driving simulator. The simulator used is a six degree-of-freedom system with a 260° front-projection cylindrical forward screen, a rear screen, LCD side mirrors and an external sound system. The simulation software used was SimCreator by Realtime Technologies. An image of the simulator system can be seen in Figure 2. The driving scenarios consist of driving on a straight, suburban roadway, with half meter-sized boxes appearing above the roadway. Participants were directed to press a button on the steering wheel when they saw a box with a horizontal stripe but not when they saw a box with a vertical stripe. Before performing the three assessment scenarios, participants completed a 5-minute practice scenario to acclimate to driving on the simulator. In the assessment scenarios, 12 boxes were shown to the participants: two decoy (vertical) boxes and ten target (horizontal) boxes.

Figure 2:

The driving simulator.

The three scenarios driven correspond to the three lighting conditions described above: photopic, mesopic, and mesopic with glare. The photopic scenario was set during daytime, whereas the mesopic was set during nighttime, with lighting in the scene coming from virtual headlights from the participant vehicle. The glare condition in the simulator was achieved by a lamp fixed to the hood of the car, close to the driver’s side of the vehicle. This condition is intended to simulate the experience of bright headlights shining at the driver from cars in the opposing direction of travel. The lamp used was a 650 lumen LED bulb positioned approximately 2.1 meters away from the driver’s eye point, four degrees to the left of the driver. This lamp is different than the one used in the vision assessment environment and provided around 23 lux at the eye. The order that the three scenarios were provided to the participants was randomized using counterbalancing to reduce bias from order effects.

For the driving scenarios, participants traveled on a two-way undivided roadway with one lane in each direction of traffic. There was light traffic in the opposing lane and participants were told to drive at 45 MPH (72.4 km/h) and follow a car in front of them. The lead car was set to match the participant’s speed and stay within 30 meters of the vehicle, if the participant drove between 40 and 50 MPH (64.3 and 80.5 km/h). If the participant drove slower than 40 MPH (64.3 km/h), the lead car would drive ahead, and if the participant drove faster than 50 MPH (80.5 km/h), the lead car would not accelerate further. In this way, the participant’s speed was constrained to a 10 MPH (16.1 km/h) band, but variation in that range would feel natural.

In order to make the driving task more difficult overall, participants were asked to complete a cognitive load task. This task consisted of casual conversation between the moderator and the participant about general interest topics such as travel and food. The moderators were located in a control room and spoke to the participants through an intercom. This task was performed on all three drives.

Statistical analysis

Strong correlations between both visual acuity and contrast sensitivity with the lighting condition were expected. This correlation would lead to strong multicollinearity in the linear models if not addressed. In order to de-correlate the VA and AULCSF scores from lighting condition, z-scores of VA and AULCSF were computed for each lighting condition. These standardized values were used in linear mixed models to evaluate the relationship between visual function, age and lighting conditions on both box identification task performance and vehicle control measures. Participant identity was modeled as a random effect, while lighting condition, age, standardized VA and standardized AULCSF were modeled as fixed effects. The Kenward-Roger approximation was used to calculate effective degrees of freedom and F-tests were applied to evaluate statistically significant differences. The linear mixed models were computed in R (R version 4.4.2) (R Core Team 2024) using the afex package (Henrik Singmann et al. 2024).

To evaluate the participants’ driving control, two measures were analyzed: standard deviation of lateral position (sdLP) and mean velocity. These quantities were measured for each box instance that the participant encountered. The boxes appear floating above the roadway 40 meters ahead of the vehicle at predetermined points in the scenario. The measurement window for the driving control measure extend from when the box appears to three seconds afterwards. Three outliers were detected using the fence method (Jones 2019) for these two quantities and removed. Accuracy, which is the percentage of correct answers, was used to assess the ability of the participants to identify the correct boxes. Reaction time from the appearance of the box until the button was hit was also used to measure performance on the box task. Measurements for reaction time were limited to correctly identified boxes; response times to decoy boxes were ignored. Reaction time was log transformed before inclusion in the model to convert from a ex-Gaussian distribution to a more Gaussian one (Whelan 2008).

Demographics

Twenty participants went through the two assessments. However, the data from two participants were excluded from the analysis: one was the first participant and procedures were adjusted after their participation; one was excluded due to technical difficulties during the driving assessment. Therefore, the analysis was performed on the data from 18 participants: 8 women (Age M=67.88, SD=4.39) and 10 men (Age M=72.6, SD=6.75).

Results

Visual function and lighting condition

As expected, values for VA and AULCSF are highly correlated with each other and with lighting conditions. With lighting condition represented as a difficulty level from one to three (1: photopic, 2: mesopic, 3: glare), Kendall rank correlation coefficients between lighting condition and VA ($\tau =.73$, $z=6.86$, $p<.001$), between lighting condition and AULCSF ($\tau =-.78$, $z=-7.32$, $p<.001$) and between VA and AULCSF ($\tau =-.80$, $z=-8.57$, $p<.001$) were all greater than 0.7 and statistically significant.

Box identification task

For accuracy, the effects of lighting condition ($F\left(2,32.54\right)=3.86$, $p=.031$) and age ($F\left(1,15.38\right)=8.71$, $p=.010$) were found to be statistically significant, while the other variables were not. Accuracy was highly skewed to the right (Skewness: −1.83), having a maximum value of 1 for 31% of the values recorded.

The log transformation of reaction time did result in a more Gaussian distribution, reducing skewness from 2.51 (right skewed) to 0.94. This was further confirmed by visual comparison of q-q plots. For reaction time, the effects of the lighting condition ($F\left(2,403.92\right)=24.17$, $p<.001$) and age ($F\left(1,18.88\right)=10.99$, $p=.004$) were found to be significant while other variables were not. See Table 1 for a summary of results.

Table 1:

Summary table for linear mixed models of box-related dependent variables showing the F statistic, degrees of freedom, residual degrees of freedom using the Kenward-Roger approximation and the p-value for each predictor variable.

Effect	$\mathit{F}$	${\mathit{d}\mathit{f}}^{\mathbf{K}\mathbf{R}}$	${\mathit{d}\mathit{f}}_{\mathbf{r}\mathbf{e}\mathbf{s}}^{\mathbf{K}\mathbf{R}}$	$\mathit{p}$
Accuracy
Lighting Condition	3.86	2	32.54	.031
VA	0.03	1	47.65	.870
AULCSF	0.43	1	46.55	.518
Age	8.71	1	15.38	.010
Reaction time (log transformed)
Lighting Condition	24.17	2	403.92	< .001
VA	0.02	1	399.95	.895
AULCSF	3.14	1	275.36	.078
Age	10.99	1	18.88	.004

A plot of standardized estimate sizes is presented in Figure 3. Increased age had a positive effect on reaction time (resulting in longer reaction times) and a negative effect on accuracy.

Figure 3:

Standardized estimates for box identification-related dependent variables. Box reaction time is log-transformed. Lines represent extent of 95% confidence intervals.

Figure 4 shows estimated margin means for box-related measured variables for all three light level scenarios. Estimated marginal means (EMM) are a way to show the expected value of a variable, adjusted for the model (Lenth 2023). Tests for significant differences were made using Tukey’s HSD test. Differences were significant only between the photopic and glare conditions for accuracy ($t\left(32.54\right)=2.63$, $p_{\text{Tukey}\left(3\right)}=.033$) and reaction time ($t\left(403.71\right)=-6.37$, $p_{\text{Tukey}\left(3\right)}<.001$). All other differences were non-significant.

Figure 4:

Estimated marginal means of box identification measures for the three light level scenarios. The dots represent the point estimates, gray bars show the 95% confidence interval of the estimates and the arrows are used for comparing differences. If the arrows overlap between two scenarios, the two means do not differ significantly.

Vehicle control

For sdLP, the effects of lighting condition ($F\left(2,538.73\right)=21.96$, $p<.001$), and age ($F\left(1,16.88\right)=10.58$, $p=.005$) were statistically significant while the effects of VA ($F\left(1,441.59\right)=0.65$, $p=.419$), AULCSF ($F\left(1,272.76\right)=1.12$, $p=.291$) were not. For mean velocity, the effects of VA ($F\left(1,545.54\right)=4.31$, $p=.038$), and AULCSF ($F\left(1,472.86\right)=4.87$, $p=.028$) were statistically significant while the effects of lighting condition ($F\left(2,535.99\right)=1.83$, $p=.162$) was not. For SD of velocity, the effects of VA ($F\left(1,422.48\right)=6.19$, $p=.013$) was stastically significant, while the effects of AULCSF ($F\left(1,256.03\right)=0.06$, $p=.811$) and lighting condition ($F\left(2,539.06\right)=1.29$, $p=.275$) was not. See Table 2 for full summary results. A plot of standardized estimates can be seen in Figure 5.

Table 2:

Summary table for linear mixed models of vehicle control-related dependent variables showing the F statistic, degrees of freedom, residual degrees of freedom using the Kenward-Roger approximation and the p-value for each predictor variable.

Effect	$\mathit{F}$	${\mathit{d}\mathit{f}}^{\mathbf{K}\mathbf{R}}$	${\mathit{d}\mathit{f}}_{\mathbf{r}\mathbf{e}\mathbf{s}}^{\mathbf{K}\mathbf{R}}$	$\mathit{p}$
SD of lateral position
Lighting Condition	21.96	2	538.73	< .001
VA	0.65	1	441.59	.419
AULCSF	1.12	1	272.76	.291
Age	10.58	1	16.88	.005
Mean Velocity
Lighting Condition	1.83	2	535.99	.162
VA	4.31	1	545.54	.038
AULCSF	4.87	1	472.86	.028
Age	0.06	1	16.45	.816
SD of Velocity
Lighting Condition	1.29	2	539.06	.275
VA	6.19	1	422.48	.013
AULCSF	0.06	1	256.03	.811
Age	0.49	1	16.93	.495

Figure 5:

Standardized estimates for driving control-related dependent variables. Lines represent extent of 95% confidence intervals.

Figure 6 shows estimated margin means for driving-related measured variables for all three lighting conditions. Tests for significant differences were made using Tukey’s HSD test. Differences were significant only for the sdLP measure between the photopic and glare conditions ($t\left(536.96\right)=-6.55$, $p_{\text{Tukey}\left(3\right)}<.001$) and the mesopic and glare conditions ($t\left(538.76\right)=-4.12$, $p_{\text{Tukey}\left(3\right)}<.001$). No significant differences were found between any lighting conditions for the mean velocity or SD of velocity measures.

Figure 6:

Estimated marginal means of driving control measures for the three lighting conditions. The dot represents the point estimation, gray bars show the 95% confidence interval of the estimation and the arrows are used for comparing differences. If the arrows overlap between two conditions, the two means do not differ significantly.

Discussion

This study examined the effects of visual acuity and contrast sensitivity on differences in driving performance measures in older adults in three different lighting conditions. It looked at both a driving scene visual identification task and vehicle control measures. The results indicate that age is a statistically significant factor in object identification reaction times and lane position variability while the visual function measures VA and CSF are significant factors in driving speed.

The overall high accuracy scores from the box task may indicate that the task was too easy overall. Tuning the target size or placement to make the boxes more difficult to notice and identify could reduce ceiling effects and show more differences between the different conditions. More complex identification tasks that rely on a wider variety of objects could be attempted as well as choosing a more demanding cognitive load task.

In the current study, age was the primary contributing factor to inaccuracy and longer reaction time in determining box identity in the go-no-go identification task. The effects of the visual function, as measured in our study, were not observed to have a significant influence on the box identification task. However, it is well known that age does have a negative effect on reaction time tasks (Woods et al. 2015). Isolating the effects of age vs vision for reaction time may require a different study design. There could be several solutions to this issue in future work. If the number of participants increases such that there were multiple levels of visual performance for a given narrow age range, this might mitigate the problem. Also, comparison with a separate, auditory reaction time test could establish a non-vision related baseline of reaction time performance for each individual.

Visual performance was a significant factor in the average speed that participants drove during the test. It is possible that individuals compensate for their lowered vision performance by driving more slowly. However, there is some contradiction in the direction of the effects. For VA higher values indicate lower acuity. The CSF values are the reverse: they increase with better visual performance. Because of this, it is unexpected that the estimates for both VA and CSF are negative for sdLP and mean velocity. Perhaps individuals are more aware of visual acuity deficits than contrast sensitivity deficits and that affects their compensation strategies while driving. More investigation will be needed to determine the nature of this relationship.

Limitations

The conversational task to increase the cognitive load of the participants was not necessarily an equal load between the participants, and we do not know if participants restricted their speech when a box was on screen. However, casual conversation is a realistic task that many people perform while driving, and similar attention shifting might occur on a real roadway test. Heavily cognitively loading tasks like OSPAN or n-back tasks may be employed in future research and would be expected to lower overall performance on the box task and the driving control measures.

There is a potential limitation in the fact that the absolute light levels in the visual laboratory setting, the simulated driving setting and the expected real-world setting are not the same. This is a potential drawback, but the expected real-world setting is highly variable. Headlights in real vehicles are often in the 2000 lumen range, but the lux falling on the eye changes based on the distances of the approaching car. One advantage of using different glare sources and lux levels in this study is that even without matching the glare sources, a relationship between visual function and driving performance was still evident. Regardless, for future studies, lower lux levels might be chosen as the glare source in the vision testing environment for a closer match with the simulated driving condition.

Conclusion

It is important to note that visual function is inherently more influential on some driving tasks than others. For example, reaction time for a go/no-go task is a single measure for an action that demands visual, cognitive and motor function. Cognitive decline associated with age will affect performance negatively even in an individual with no visual deficits. Since most older adults will have a combination of these deficits, the degree to which any one of them affect driving-related task performance is difficult to ascertain (Andersson and Peters 2020). Even more difficult is if there is an additive or super-additive effect for these functional deficits.

To disentangle these two domains in driving performance for older drivers, future work must include two elements: 1. greater variation in subjects representing a wider array of the combination of visual function and age and 2. greater variability in task difficulty. To avoid ceiling effects, adjustments in the visual appearance of the stimuli (contrast level, size, time in visual field) can be made to increase overall difficulty, even for those with normal vision. In addition, multiple tasks with multiple difficulties must be given to the participants to ensure comparison between those with different visual capabilities. Work can also be done to establish estimates of accuracy rates for similar tasks in a driving simulator to find a task (fit for Goldilocks) that is not too easy and yet not too hard.