Abstract
Objective
To assess the attentional demand of different contextual factors in driving.
Background
The attentional demand on the driver varies with the situation. One approach for estimating the attentional demand, via spare capacity, is to use visual occlusion.
Method
Using a 3 × 5 within-subjects design, 33 participants drove in a fixed-base simulator in three scenarios (i.e., urban, rural, and motorway), combined with five fixed occlusion durations (1.0, 1.4, 1.8, 2.2, and 2.6 s). By pressing a microswitch on a finger, the driver initiated each occlusion, which lasted for the same predetermined duration within each trial. Drivers were instructed to occlude their vision as often as possible while still driving safely.
Results
Stepwise logistic regression per scenario indicated that the occlusion predictors varied with scenario. In the urban environment, infrastructure-related variables had the biggest influence, whereas the distance to oncoming traffic played a major role on the rural road. On the motorway, occlusion duration and time since the last occlusion were the main determinants.
Conclusion
Spare capacity is dependent on the scenario, selected speed, and individual factors. This is important for developing workload managers, infrastructural design, and aspects related to transfer of control in automated driving.
Application
Better knowledge of the determinants of spare capacity in the road environment can help improve workload managers, thereby contributing to more efficient and safer interaction with additional tasks.
Introduction
The attentional demand on the driver varies with the driving context (Kujala, Mäkelä, Kotilainen, & Tokkonen, 2016; Patten, Kircher, Östlund, Nilsson, & Svenson, 2006; Tivesten & Dozza, 2015), with the demand increasing when more information has to be sampled and processed within a shorter timeframe (Barrouillet, Bernardin, & Camos, 2004). The objective of this paper is to investigate how contextual variables in the driving scenario affect attentional demand.
When a driver is attentive, his or her mental model of the current situation is good enough to predict the scenario development, including the knowledge of where and when new information sampling is necessary (Clark, 2015; Engström et al., 2018; Kircher & Ahlstrom, 2017). In manual driving, the visual channel is the main source of information (Sivak, 1996; Spence & Ho, 2008), and gaze direction has even been used as a proxy for attention (Doshi & Trivedi, 2009; Theeuwes, Kramer, Hahn, & Irwin, 1998; Yantis & Jonides, 1990). If the visual input is cut off for more than a short time, the mental representation of the current situation, including its likely future development, quickly becomes uncertain (Senders, Kristofferson, Levison, Dietrich, & Ward, 1967). The more difficult a situation is to predict, the more frequently the driver has to verify or correct and update the mental model with new information (Senders et al., 1967; Tsimhoni, 2003). Drivers likely have a personal limit to how much uncertainty they tolerate (Kujala et al., 2016), so while attention constitutes more than visual information sampling, the necessary sampling frequency can give an indication of the situational demands.
When situational demands are low, the driver can either spend more time sampling the required information in a redundant fashion or use the arising visual spare capacity to deal with additional tasks, which may or may not be driving related. With increasing situational demands, the spare capacity diminishes or disappears. If demands increase even more, drivers typically slow down to accommodate the necessary information sampling (Fuller, 2005; Horberry, Anderson, Regan, Triggs, & Brown, 2006; Metz, Schömig, & Krüger, 2011). This notion of visual spare capacity has been suggested and demonstrated in various ways (Birrell & Fowkes, 2014; Green & Shah, 2004; Hughes & Cole, 1986). For example, more experienced drivers tend to glance less often at traffic-relevant targets than less experienced drivers, indicating that the former are more effective at sampling the relevant information for a sufficiently accurate mental model (Clark, 2016; Underwood, 2007; Underwood, Chapman, Bowden, & Crundall, 2002). An additional sign for the existence of spare capacity is the fact that drivers mention, and therefore have seen and processed, objects not related to traffic when asked to think aloud while driving (Young, Salmon, & Cornelissen, 2013).
As mentioned earlier, we assume that environmental, situational, and internal features determine when information is needed and when it may be possible to cease sampling for a short time. Longer durations without information sampling require a more predictable environment, such that the tolerated uncertainty limit is not breached (Kujala et al., 2016). Longer durations might also require more time between nonsampling events, as it may take longer to reduce the accumulated uncertainties (see also Samuel & Fisher, 2015).
Visual occlusion has been used to study when and for how long visual information is taken in, and when it is not. Occlusion is usually assumed to occur for the period when no additional visual information is needed. Visual occlusion can be achieved either by wearing dedicated glasses (e.g., van der Horst, 2004) or, in a simulator, by blanking out the screen (e.g., Samuel & Fisher, 2015). Depending on the construction of the glasses or the part of the screen that is blanked out, occlusion is either complete or partial. Foveal occlusion with intact peripheral vision seems to enable successful lane tracking (Lamble, Laakso, & Summala, 1999; Summala, Nieminen, & Punto, 1996) and is more likely to resemble the visual intake of the road environment during an interaction with, for example, a mobile phone.
The default state of the occlusion can either be blanked out (“occluded”) or affording vision (“open”), and the timing and duration of the occluded or open period can be determined by either the driver or the experimenter. The first on-road occlusion study employed a predetermined opening time of 0.5 s and occluded vision as the default (Senders et al., 1967). The drivers could choose when they wanted to open the occluding visor, and it was found that higher speeds and more curved roads required more frequent visual intake. This setup is likely to come the closest to assessing the minimum visual demand, as drivers must actively request vision, which they are granted for short periods. It might, however, overestimate the time that drivers consider “safe” to look away, as vision might be blocked again earlier than the driver would have chosen.
Letting drivers choose both when and for how long they block vision of the traffic environment with a default open state may give a better indication of the willingness to focus elsewhere. This fully self-paced approach was used in a study in real traffic on the motorway (Kircher & Ahlstrom, 2018). This study revealed that drivers were willing to occlude themselves more frequently and for longer periods when driving in the slow lane, less frequently and for shorter periods when driving in the fast lane, and the least when changing lanes into the fast lane, illustrating the importance of the intended maneuver for situation prediction and information intake.
In naturalistic settings, it is possible, however, that the duration of a glance away from traffic is not completely determined by the driver, but by the target of the glance. Some additional tasks are not easily interruptible and may therefore force the driver to glance away for longer periods (Brumby, del Rosario, & Janssen, 2010; Brumby, Salvucci, & Howes, 2009; Iqbal & Bailey, 2005; Janssen, Brumby, & Garnett, 2012; Lee, Gibson, & Lee, 2015; Noy, Lemoine, Klachan, & Burns, 2004; Salvucci & Kujala, 2016). This condition can be emulated by self-paced occlusion with a predetermined duration, an approach employed here.
The aim of this study is to investigate which individual external factors independently influence occlusion duration, considering both situational factors and occlusion duration. Occlusion probability is used as a proxy for situational demand, which, in turn, is based on the predictability of the situational development. More concretely, we assume that occlusion probability decreases as predictability decreases. This can be due to visual obstruction or the difficulty of predicting the maneuvers of other road users.
Method
Participants and Apparatus
A convenience sample of 33 participants was recruited from a database of interested participants and through personal contacts. All participants had driven at least 3,000 km over the last year and had a valid driving license. They were each reimbursed with a cinema ticket. Three of the participants did not follow the experimental instructions: Two forgot to perform the occlusion task in several trials and one participant fell asleep at the wheel. Therefore, their data were excluded, leading to a final sample of 30 participants. Their age range was 22 to 75 years (M = 34.4; SD = 14.7 years) with a driving experience range of 2 to 57 years (M = 15.2; SD = 14.5 years).
This study complied with the American Psychological Association Code of Ethics and was approved by the Regional Ethics Review Board in Linköping, Sweden (DNR: 2014/0177-8.2). Informed consent was obtained from each participant.
A fixed-base driving simulator with an automatic gearbox was used. The visual system consisted of nine 21-in screens, providing a 180° × 30° field of view. Participants could blank out the central 110° of the forward scene by pressing a microswitch attached to their finger (Figure 1). The peripheral view remained intact, simulating visual focus away from traffic with preserved ambient vision. After a predetermined duration, the central visual scene reappeared.
Simulator setup (left), with the microswitch for occlusion initiation (right).
Experiment Design and Driving Task
The experiment followed a 5 (occlusion duration) × 3 (scenario) within-subjects design. The occlusion durations were 1.0, 1.4, 1.8, 2.2, and 2.6 s, defined based on pilot experiments with five drivers and self-selected occlusion durations. The drivers were asked to occlude the front view as long and as often as possible while driving safely. Based on this distribution, the 5th and 95th percentiles were chosen as endpoints. The scenarios are described in Table 1. The order of the 5 × 3 combinations, henceforth called trials, was randomized between participants, with the restriction that the different trials were approximately equally distributed among the different chronological positions. When the participant maintained the speed limit, each trial took 3 min. All trials were connected by a neutral road stretch that took 12 s to drive (Figure 2). A lead vehicle, driving at the speed limit posted in each trial, was present throughout the drive. The participants were instructed to follow the lead vehicle, but the following distance was not explicitly specified.
Descriptions and Screenshots of the Three Scenarios Used
Driving route layout with subtrial and neutral phase.
The scenario environments resembled their naturalistic counterparts, such that the factors varying along the road differed between scenarios. The positioning of curves and parked cars are indicated in Figure 4 in the “Results” section. The investigated factors belonged to infrastructural features and other traffic; weather, lighting, and visibility were kept constant.
Procedure
Each participant received information about the study and signed an informed consent form. After that, the participant entered the simulator, and the microswitch for occlusion was attached to the participant’s middle finger. Before the actual data collection started, the participant drove a 5-min training session to get used to all occlusion durations and the three scenarios. The participant was instructed to occlude the forward view as often as possible, while still driving safely along the route. Within each trial, the occlusion duration was determined by the experimental setting, but the participant initiated the occlusion based on his or her own judgment of the scene. While on the connecting road between trials, the participant was informed about the occlusion duration and speed limit for the next trial.
Analysis
Simulator and occlusion data were recorded at 20 Hz, including speed, distance to surrounding vehicles, road curvature, and slope. The dataset was enriched by manually adding sight distance, distance to parked cars, and distance to intersections. For sight distance, a software developed in-house was used, where a lead vehicle could be moved manually, such that the distance when it still was just in sight could be logged. The parked cars and location of intersections were annotated manually, such that the distance toward them could be computed based on the logged data.
The relationships between occlusion probability and several explanatory variables were analyzed using binary logistic regression (e.g., Hosmer, Lemeshow, & Sturdivant, 2013). Separate models were fitted to each scenario, and the scenario with the urban road was further divided into two parts: near an intersection (less than 200 m) and no intersection (excluding 200 m before and 150 m after an intersection). The reason for this was that the intersections were assumed to have a high impact on attentional demand. The dataset was divided into time intervals of the same duration as the occlusion, and each time interval was treated as one observation in the regression analysis. The time intervals were created such that a new interval always started when an occlusion started. If the time between the end of one interval and the start of the next occlusion was less than the occlusion duration, that part was excluded from the regression analysis to avoid intervals of varying durations. The dependent variable took the values “occluded” and “not occluded” depending on whether or not the participant was occluded during the specific time interval.
The aim of the statistical analyses was to explore which explanatory variables were covarying with the outcome, occlusion probability. This was investigated by using three different methods of Lasso, group Lasso, and stepwise logistic regression, respectively. The results showed that the Lasso method was not good at handling categorical variables with more than two levels with current dataset here. And the group Lasso method seemed to penalize the categorical variables too much, as the variable participant did not enter any of the Lasso models, something that strikes us as highly unlikely based on other analyses of the data and previous experience from other experiments (e.g., Kujala et al., 2016). So, this study chose to using stepwise logistic regression in which both main effects and all two-way interaction terms could enter the model (the exception was participant, which could only enter as a main effect). The stepwise procedure started with fitting a model only including an intercept. Next, chi-square statistics, showing the added contribution for each eligible explanatory variable, was calculated. The variable with the lowest corresponding p value was selected for entry if the p value was below a predefined level. A model including this variable was then fitted and the procedure was repeated until no more variables met the entry-level criteria. In each step, the p values of the included variables were calculated and compared with an exit level. If the p value was higher than the exit level, the variable was again removed from the model. However, if a two-way interaction term was included in the model, both corresponding main effects were also included, regardless of their p values. The significance level for both entry and exit was set to .05.
It should be noted that stepwise logistic regression that automatically selects variables based on data has some drawbacks. Random variation in the data set might lead to overfitting, such that the final model reflects only this particular data set and may not be generalizable to new data if the experiment is repeated. Also, the p values in the final model are underestimated and the sizes of the regression coefficients are likely to be too large (e.g., Hosmer et al., 2013; Thompson, 1995). To minimize the problem of overfitting, we only chose variables that we hypothesized were relevant for the decision to occlude or not. Cross validation with 1,000 repeated random subsamplings (90% training, 10% validation) was used to investigate whether the same variables were selected in each iteration and to verify that the prediction performance was similar between the training and validation datasets. Accuracy, sensitivity, and specificity were chosen as measures of prediction performance using a cutoff value of 0.5. That is, an observation is predicted as an occlusion if the predicted occlusion probability equals or exceeds 0.5.
In total, 11 explanatory variables could enter the model (see Table 2). The variable participant was entered to account for the within-subjects design of the data, and categorical variables were recoded as dummy variables. However, not every variable was relevant to all four models. For example, parked cars were only present in the urban scenario. A description of the relevant variables for each model is found in the “Results” section.
Descriptions of the 11 Explanatory Variables
Factors of interest, according to the regression results, were visualized to explore how they influenced occlusion probability. The occlusion probability curves were calculated by dividing the number of occlusion cases by the number of all cases. This was done for each distance point using a resolution 2 m. For visualization purposes, the resulting occlusion probability curves were smoothed using a moving average filter with a window length of 10 samples (20 m).
Data extraction and preprocessing were conducted in MATLAB version R2013 (MathWorks, Natick, MA, USA), the analyses of variance were performed in SPSS 19.0 (IBM, Armonk, NY, USA), and the stepwise logistic regression was implemented using proc logistic in SAS 9.4 (SAS Institute, Cary, NC, USA).
Results
Speed and lane exceedance were analyzed to investigate possible compensatory behavior or loss of control. For all scenarios and occlusion durations, speeds did not differ significantly between occluded and unoccluded driving. Lane exceedance occurred in 1.3% of all occlusion cases. As a baseline condition without occlusion was unavailable, comparison was made with the percentage of lane exceedance occurrences in the same location in other occlusion duration conditions, given that no occlusion occurred there. The percentage of lane exceedance when not occluded was 3.5%.
Individual variations in occlusion frequency were larger on the motorway and for shorter occlusion durations (Figure 3 and Table 3). Two-way repeated-measures analyses of variance identified significant main effects for scenario, F(2, 58) = 24.7, p < .05,
Boxplot of the number of occlusions per scenario and occlusion duration. Note that longer occlusion durations cover a greater distance each.
Mean Number of Occlusions and Mean Percentage of Occlusion Over Distance per Scenario and Occlusion Duration, Including the 95% Confidence Interval
Along the urban and rural roads, the occlusion probability (the ratio of passages with occlusions to all passages of each single location, across participants) varied depending on the contextual factors (Figure 4), with lower probabilities around intersections, when passing parked cars, and when approaching oncoming traffic. In the urban environment, the longest occlusions typically occurred when a block of parked cars had just been passed and no intersection was nearby, whereas shorter occlusions followed the same pattern, but less distinctively. On the motorway, the occlusion probability was more stable.
Occlusion probability in three scenarios. Curve is annotated on the top of each graph, with red indicating right curves and blue left curves; a higher color density indicates a sharper curve. For the urban road, the positions of the intersections (black blocks) and the parked cars (gray blocks) are annotated. For the rural road, the locations where oncoming vehicles were met are indicated with green dots. Note that the x-axes of the three graphs have different scales.
Cross validation showed that the selected variables were stable across different random subsamplings of the dataset (Table 5). Some of the interactions, such as the interaction between distance headway to oncoming vehicle and curvature, were only selected in a fraction of the simulations and should be interpreted with caution.
Table 4 presents a summary of the stepwise variable selection procedure, and Table 5 provides the results of the final logistic regression models. In both tables, the variables are listed in order of entrance. The models predict the probability to occlude. Participant entered first in all models, but otherwise the models were quite different. For example, time since the last occlusion only entered in the motorway model, in which task demands were homogeneous. On the urban road, with a much more complex driving situation, other variables were important, such as distance to intersection and presence of parked cars. Occlusion duration entered the models on the rural road and motorway, but was not shown to affect the decision to occlude on urban roads. The models of urban roads without intersections and rural roads also included interaction effects, meaning, for example, that the effect of occlusion duration differed depending on the level of curvature on rural roads. Cross validation showed that the selected variables were stable across different random subsamplings of the dataset (Table 5). Some of the interactions, such as the interaction between distance headway to oncoming vehicle and curvature, were only selected in a fraction of the simulations and should be interpreted with caution.
Summary of the Stepwise Logistic Regression Procedure, Applied to Four Scenarios/Designs
Note. The models predict the probability to occlude. The variables are listed in order of entrance in the model. The grayed variables are variables included in the stepwise procedure which did not enter the model (n.e.). All variable names are explained in Table 2.
Only includes left and right curves for this scenario; the straight part of the road had no parked cars and was therefore omitted.
Results of the Final Logistic Regression Models for the Four Scenarios/Designs
Note. The models predict the probability of occlusion based on the full dataset. All variable names are explained in Table 2.
The percentage of times that a variable was selected in the 1,000 cross-validation iterations.
Only the overall significance of the variable is provided for participant and not results for each category (i.e., participant).
Accuracy and specificity of the logistic regression models were generally high, whereas the sensitivity was rather low (Table 6). Low sensitivity was expected due to the self-paced design and default unoccluded state and the imbalanced class distributions (88% unoccluded observations in the Urban near intersection dataset, 86% in Urban no intersection, 83% in Rural, and 73% in Motorway). The mean differences in percentage between the test and validation datasets were less than one percentage point for accuracy, sensitivity, and specificity, indicating that the stepwise regression models were robust to random variations in the data.
Mean and Standard Deviation of Accuracy, Sensitivity, and Specificity Across the 1,000 Randomly Resampled Cross Validation Iterations
Urban Road
The odds ratios for all variables entering the logistic regression model describing urban roads near intersections are shown in Figure 5. The odds ratio between sharp curve and straight road was less than one, meaning that the participants were less likely to occlude in a sharp curve than on a straight road. For left and right curves, the odds ratio was not significantly different from one, meaning that no effect on occlusion probability can be demonstrated. For continuous variables such as distance headway to oncoming vehicle (hwonc.), the odds ratio is interpreted as the relative change in odds when increasing the variable by one unit, here 100 m (note that this scale is used only when calculating the odds ratio; in the model, the original distance is used). Therefore, the greater the distance headway to the nearest oncoming vehicle, the more likely the participants are to occlude. In this study, there was a strong relationship between the distance to the next intersection and occlusion probability (Figure 5). A more detailed analysis determined that the distance-based occlusion probability decreased from 150 to 100 m before the intersection to approximately 25 m before the intersection, where it started to increase again (Figure 6). This result was stable across all occlusion durations.
Results of logistic regression on urban road, 0 to 200 m before an intersection; odds ratios. An odds ratio greater than one indicates a higher occlusion probability when comparing different levels of a class variable. The odds ratio for a continuous variable is the relative change in odds when increasing the variable by one unit (one unit equals 100 m for the continuous variables). The error bars represent confidence intervals (95% level). All variable names are explained in Table 2.
Occlusion probability as a function of distance to the intersection per occlusion duration. Negative values indicate the distance before the intersection; positive values indicate the distance after entering the intersection.
On urban road stretches without intersections, the odds ratios indicate that the participants were more likely to occlude when there were no parked cars along the road (Figure 7). Moreover, the participants were less likely to occlude the greater the distance headway to the lead car when driving on road stretches with parked cars. No significant effect of distance headway was found where there were no parked cars.
Results of logistic regression on urban road, excluding 200 m before and 150 m after the intersection; odds ratios. An odds ratio greater than one indicates a higher occlusion probability when comparing different levels of a class variable. The odds ratio for a continuous variable is the relative change in odds when increasing the variable by one unit (one unit equals 100 m for the continuous variables). The error bars represent confidence intervals (95% level). All variable names are explained in Table 2.
Rural Road
On the rural road, curves to either right or left led to a lower occlusion likelihood than on straight road segments. An increased distance to the nearest oncoming car increased the occlusion probability regardless of occlusion duration and curvature (Figure 8). The likelihood of occluding was greater for longer occlusion durations.
Results of logistic regression on rural road; odds ratios. An odds ratio greater than one indicates a higher occlusion probability when comparing different levels of a class variable. The odds ratio for a continuous variable is the relative change in odds when increasing the variable by one unit (one unit equals 100 m for the continuous variables). Because of significant interaction effects, separate odds ratios are presented for different levels of the interacting variables. The error bars represent confidence intervals (95% level). All variable names are explained in Table 2.
According to the stepwise regression procedure, the first variable to enter the model on rural roads, apart from participant, was distance to the nearest oncoming car. Additional analyses found that like distance to urban intersections, occlusion probability decreased from approximately 150 m to approximately 25 m before an oncoming car was met, after which the occlusion likelihood rapidly increased again, reaching a local maximum approximately 25 m behind the meeting point (Figure 9).
Occlusion probability varies with distance to the nearest oncoming car per occlusion duration. Negative values indicate distance before meeting point; positive values indicate distance after meeting point.
Motorway
On the motorway, participants were more likely to occlude the longer the time since the end of the last occlusion and with increased slope size (Figure 10). Furthermore, the participants were less likely to occlude during trials with the shortest occlusion duration compared with the longest duration and less likely in a right curve than on a straight road segment.
Results of logistic regression on the motorway; odds ratios. An odds ratio greater than one indicates a higher occlusion probability when comparing different levels of a class variable. The odds ratio for a continuous variable is the relative change in odds when increasing the variable by one unit. The error bars represent confidence intervals (95% level). All variable names are explained in Table 2.
The interval between shorter occlusions was shorter than that between longer occlusions. For occlusions of 1 s, the next occlusion was frequently already initiated after 1 to 2 s, and in 85% of the cases within 7 s, occlusion durations over 2 s were followed by longer unoccluded intervals (Figure 11).
Probability density functions of time since last occlusion for different occlusion durations. Vertical lines indicate the 85th percentile of time since last occlusion.
Discussion
In all four stepwise regression analyses, participant was always the first variable to enter the model. This indicates that the individual is a strong predictor of occlusion behavior, with some people being consistently more likely to occlude their vision than others. This was also found by Kujala et al. (2016). This may have to do with personality factors or with experience, but was not further explored here, as our main interest was to identify person-independent factors. The remainder of the discussion will therefore only treat those factors. Speed and lane exceedance analyses did not provide any indications that drivers might have compensated with speed reductions or might have been overconfident when occluding. As lane exceedance was so rare, no inferential testing was conducted, but the descriptive results indicate that, if anything, drivers exceeded the lane less often during an occlusion than when not occluding.
The factors determining where and when drivers sample information varies with the environment. Based on Clark (2015), we hypothesized that occlusions are avoided when the situation is experienced as less predictable, because more information needs to be sampled within a limited timeframe. As the speed limit and the lead car largely dictated the driving speed, speed adaptation was not used as a compensatory strategy to enable more occlusions. It is therefore assumed that more occlusions indicate a more predictive environment.
The combination of several potential occlusion predictors found in naturalistic scenarios shows that demands vary both qualitatively and quantitatively across and within scenarios, providing insight into the relative importance of the factors. Overall, the motorway situation appeared to be easier to predict and less variable in its predictability over distance than the other two scenarios. Even though the speed was higher on the motorway, theoretically enabling faster situational changes and reduced predictability, this was countered by infrastructural design, with physical separation from oncoming traffic, homogeneous behavior of other road users, no sharp curves, and no cross traffic. In comparison, the other two scenarios were less predictable, with all the above-mentioned factors being present. In the tested scenarios, the lower speed limits did not completely compensate for the decreased predictability, necessitating comparatively more information sampling.
In contrast to the motorway, the urban and rural roads displayed clear variations in occlusion frequency along their lengths, attributable to both infrastructural and traffic-related features. In the urban environment, intersections were experienced as difficult to predict, likely because traffic on the side roads had to be checked. On the rural road, the oncoming vehicles led to uncertainty. These findings are linked to previous research demonstrating that increasing road curvature leads to decreased occlusion likelihood (Kujala et al., 2016; Senders et al., 1967). The importance of other traffic for the willingness to occlude has been demonstrated indirectly in relation to maneuvers (Kircher & Ahlstrom, 2018). The rural road used in this study was rather narrow and winding, making both oncoming traffic and the participant’s own trajectory less predictable, increasing the necessity of sampling visual information in situations in which oncoming traffic is nearby. With no traffic nearby, it is still necessary to ensure one is staying on the road, which may explain why curvature was found to be a predictor.
On the urban link road, parked cars reduced the likelihood of occlusion, not only because the parked cars narrowed the available road space, necessitating more accurate lateral placement, but also because they might pull out into the road or hide pedestrians, making the environment less predictable. Somewhat counterintuitively, on road stretches with parked cars, the results also indicated that participants were less likely to occlude the greater the distance headway to the lead car. It should be remembered, however, that not everything can be accounted for in the model.
While the lower occlusion frequency for longer durations could be due to fewer opportunities to fit in long occlusions on urban and rural roads because of the heterogeneous environment, this was not true for the motorway. Rather, the data indicate that there is a form of upper percentage-of-occlusion limit per location, which can be reached through either many short or a few long occlusions. The internal representation used to predict the development of the situation does not deteriorate as much under short occlusion durations as under long ones, as drivers’ uncertainty about the environment has been found to increase with x1.5 where x is off-road glance duration (Senders et al., 1967). Thus, less unoccluded time is needed to “repair” the mental representation, allowing for the next occlusion to occur sooner. This complements the laboratory findings of Lu, Coster, and de Winter (2017), who demonstrated that longer viewing times and less complex situations yielded more accurate results in reproducing a traffic situation. It also complements the findings of Samuel and Fisher (2015), who investigated different unoccluded times between consecutive fixed 2-s occlusions and found that hazard detection worsened when the driver looked ahead on the roadway for less than 7 s between occlusions.
The stepwise logistic regression, as applied here, should be seen as a way of exploring the most important explanatory variables and of understanding possible interactions between them. Additional analyses are needed to capture detailed information on what happens, for example, in the transition between parked cars and no parked cars, immediately after an intersection, and when passing an oncoming car. Visual inspection of Figure 4 shows that the occlusion probability varies more with location for longer occlusion durations than for the shorter durations, especially on the urban and rural roads where longer occlusion durations necessitate more precise targeting of a suitable position along the road where information does not have to be sampled. Also, visual inspection of Figures 6 and 9 shows that the occlusion probability starts to increase again approximately 20 m before the entrance point of an intersection or the meeting point with an oncoming car, respectively. This supports the notion that drivers predict their environment (Chen & Milgram, 2013; Endsley, 1995; Engström et al., 2017; Gibson & Crooks, 1938; Johnson-Laird, 1983; Kircher & Ahlstrom, 2017; Kircher, Ahlstrom, Nylin, & Mengist, 2018). Upon passing the vehicle or the intersection, the occlusion probability reaches a local maximum, indicating that the self-selected, environmentally dictated longer viewing opportunity while approaching the oncoming car leads to an ensuing concentration of occlusion readiness. This is in line with findings by Chen and Milgram (2011), who showed that longer occlusion periods also required longer glance times and vice versa.
There are some limitations to this study. First, the study was conducted in a fixed-base simulator without haptic feedback, which may affect external validity. Based on the results of Greenberg, Artz, and Cathey (2003), keeping to one’s lane is more difficult without lateral motion and haptic feedback, which may tend to reduce the absolute number of occlusions. However, knowing that no real consequences would result from crashing might have increased the occlusion probability. Either way, it is unlikely that different occlusion duration conditions would have been affected systematically differently, so relative differences are likely to be valid.
Second, occlusion was not coupled to an additional task. The occlusion time could thus be used to continue mentally processing the situation, possibly allowing a more accurate prediction of the scenario development while occluded. Further research comparing occlusion behavior, driving behavior, and ability to predict depending on the presence and difficulty of an additional task would be necessary to explore this matter. A working hypothesis could be that more automated, low-level predictions such as lane-keeping would not interfere as much with an additional task as would predictions involving higher level processes, such as lane choice in a complex intersection or hazard detection. While Samuel and Fisher (2015) did not find that an additional search task worsened hazard detection in comparison with blank occlusion, more cognitively demanding tasks still might do so.
Third, the occluded area was fixed and consisted of the complete forward view, such that only limited peripheral information was available. There are indications that peripheral vision is underestimated and may contribute more to the ability to drive than previously assumed (Rosenholtz, 2016; Wolfe, Dobres, Rosenholtz, & Reimer, 2017). This can be seen, for example, during interactions with mobile phones, when drivers place their phone strategically to capture as much information as possible with the peripheral vision, which was impossible in the current study. Enabling maximum use of peripheral vision was already recommended by the European statement of principles on the design of human–machine interaction (Stevens et al., 2008), which applies both to standard and head-up displays.
Fourth, to avoid tiring the participants, no baseline was run. This made it more difficult to make comparisons with unoccluded behavior, though a workable solution was found. Still, a dedicated baseline should be considered in future studies. Also, a sophisticated analysis of gaze behavior to determine whether and how participants adapt their glance strategies to the reduced available glance time, and whether that has any effects on cognitive load, would be of interest given the frequent execution of additional tasks while driving.
Finally, stepwise logistic regressions, as well as other methods for variable selection, have limitations and drawbacks. Here, we regard the procedure as an explorative method that helped us understand the relationships between occlusion probability and variables describing the road and traffic environment. We believe that the results in general are in line with the hypothesis that occlusions are avoided when the situation is experienced as less predictable (based on Clark, 2015) and thus credible. Also, the similarities and inequalities between the four models fitted to different scenarios are realistic. However, we do not claim to have found the best model explaining occlusion probability, and the present findings should be confirmed in future studies. This especially applies to the interaction terms that were only selected in a fraction of the cross-validation iterations.
Conclusion
The results clearly indicate that drivers consider their surroundings when deciding to occlude their vision. The results also indicate that different types of environments generate different amounts of visual spare capacity, both globally, with the motorway providing more occlusion opportunities than the other scenarios, and locally, depending on the situation. From an applied viewpoint, the results indicate that the increasing knowledge that modern cars have about the present environment can be used to provide drivers with additional information at times of low attentional demand.
Key Points
Drivers plan where and when to sample information, and the factors determining the information collection vary depending on the environment.
This simulator study, based on the visual occlusion paradigm, provides insight into the importance of such factors in three environments (i.e., urban, rural, and motorway).
Contextual variables related to infrastructure (i.e., intersections), interacting traffic (i.e., oncoming vehicles), and occlusion (i.e., duration since the previous occlusion) were the main determinants of attentional demand on the urban road, rural road, and motorway, respectively.
Better knowledge of attentional demand in different driving scenarios can advance the development of situation-sensitive distraction warning systems and contribute to the design of in-vehicle user interfaces.
Footnotes
Acknowledgements
We gratefully acknowledge Tuomo Kujala at the University of Jyväskylä, Finland, for valuable comments on an early draft of this manuscript. This work was carried out at the Swedish driving simulation center Virtual Prototyping and Assessment by Simulation (ViP; 
), which is financed by the Swedish Governmental Agency for Innovation Systems (VINNOVA), grant no. 2011-03994, and the center’s partners. In addition, the Key Laboratory for Automotive Transportation Safety Enhancement Technology of the Ministry of Communication, PRC (grant no. 300102229508), supported the research.
Zhuofan Liu is a lecturer and researcher in Xi’an University of Posts & Telecommunications. He received his PhD in Vehicle Engineering from Chang’an University in 2018.
Christer Ahlström is a senior researcher in the Driver State Group at the Swedish National Road and Transport Research Institute (VTI). He received his PhD in biomedical signal processing from Linköping University in 2008, where he has been an associate professor in the Department of Biomedical Engineering since 2018.
Åsa Forsman is a researcher in Traffic Safety at the Swedish National Road and Transport Research Institute (VTI). She received her PhD in statistics from Linköping University in 2002.
Katja Kircher is a research leader and belongs to the Department of Human Factors in the Transport System at the Swedish National Road and Transport Research Institute (VTI). She received her PhD in industrial ergonomics from Linköping University in 2002, where she has been an associate professor in the Department of Behavioral Sciences and Learning since 2015.