In this article, we exhaustively review eye-tracking research on real-world night driving. Night driving was chosen for two reasons. First, driving a vehicle is among the most dangerous activities humans regularly perform. Globally, in 2015 alone, 1.3 million people died due to road traffic injuries [1]. Second, important in the context of the present review, driving at night is particularly dangerous. Despite similar total numbers of crashes during daytime and nighttime, fatalities per 100 million miles are disproportionately higher at night than during the day [2]. Although fatal accidents occur for a variety of different reasons, the low luminance at night impairs human vision, as we will also briefly explain below, and thus contributes to the danger of nocturnal traffic crashes [3]. Therefore, understanding the conditions for safe night driving behavior is very important. Here, we used a systematic and exhaustive review of all existing eye-tracking research on real-world night driving to understand how much can be concluded from this research about the conditions of safe night driving.

We focused on real-world night driving, since real-world night driving occurs under the most realistic light conditions, as opposed to driving in a simulator [4], and in a three-dimensional environment. Furthermore, as the risk of accidents is only present during real-world driving, driving behavior in the real world provides the yard-stick to which conclusions from simulator research have to be compared [5]. Therefore, we refer to simulator studies of night driving only occasionally, for example, to highlight gaps in the eye-tracking research on real-world night driving. In subsequent paragraphs, you should be building a case for your study. Explain what has been found in previous research on this topic, describe what gap exists in this literature, and explain how your study will fill the gap (i.e., provide a unique study that will contribute new knowledge in the area).

Although the ultimate goal is to identify and understand the conditions that facilitate or hinder safe, successful driving, this review is restricted to gaze behavior since eye movements provide a relatively universal window into many different cognitive processes underlying driving behavior (e.g., memory, decision making, action control, etc.). Driving depends on vision [6] and eye movements indicate from which locations in the visual environment information is picked up [7]. Admittedly, the connection between the selection of visual information from a particular location and the direction of gaze is not perfect. For example, human vision is not as sensitive during saccades (i.e., the jerky jumping movements of the eyes from one location to another) as during fixations (i.e., the phases when the eyes stand relatively still) and during smooth pursuit eye movements (i.e., the eye movements that keep track of a moving object, such as an oncoming car; cf. [8]. Thus, for meaningful interpretations of eye-tracking data, it is necessary to discriminate between different types of eye movements, such as fixations and saccades. In addition, visual information is sometimes selected from the periphery of the visual field, indicating that gaze direction and covert shifts of spatial attention are not always aligned [9-11]. However, in numerous instances, gaze direction and fixation duration are valid indicators of the kind of visual information that has been selected (e.g., 7,12). Therefore, at least gaze directions during fixations and smooth pursuit eye movements are good approximations to the visual locations from which humans select information during driving.

To answer the ultimate questions concerning helpful and harmful conditions during night driving using eye tracking, a model explaining the appropriate expedient gaze behavior is necessary. In the case of gaze directions, it is important to know which areas of the visual field contain relevant information for driving and which areas contain irrelevant or distracting information. This requires a deeper analysis of the task of driving and the subsequent localization of visual information, which is relevant to the task. Therefore, in the present review we first summarize the major characteristics of the visual situation during night driving, and secondly, we analyze the task of driving. These two introductory chapters set the proper stage for the subsequent research review.

As mentioned above, fatalities are higher during night driving than during daytime driving, and one contributing factor could be the low illumination at night. Low illumination is a challenge for humans, as human vision has primarily adapted to daylight illumination, and thus, the eye is more responsive to higher illumination. During the day, illumination ranges from 1,000 lux on an overcast day to about 100,000 lux at direct sunlight [13]. During the night, these values decrease considerably, ranging from 0.0001 lux at a moonless night to 1 lux at a full moon night, meaning that there is much less stimulation of the eyes at night than during the day. Luckily, the eyes can adapt to these illumination differences. About 95% of all retinal photoreceptors are rods (about 60,100,000). These rods are highly sensitive to light and respond even to single photons. This enables humans to report the presence of a single photon with above-chance accuracy [14]. Thus, rods can compensate for low illumination at night.

Yet, rods are found in the periphery of the retina, and they are largely absent at the retinal fovea. Importantly, the periphery of the retina has a low spatial acuity because retinal ganglion cells pool the visual signal across large areas [15]. In addition, rods respond indiscriminately to light of different wavelengths, and thus, rods do not allow color discrimination. In conclusion, visual adaptation to lower illumination at night comes at the cost of lower spatial acuity (or resolution) and less color vision.

In contrast, only at the fovea, single photoreceptors connect to single retinal ganglion cells, thereby enabling a high spatial resolution of vision. Therefore, high spatial resolution depends on cones (about 3,200,000), which are concentrated at the fovea [16,17]. In addition, cones are differentially sensitive to different light wave-lengths, thereby enabling color vision. However, the absolute threshold of the cone response is about 200 photons [18], which means that cones contribute to vision only if the luminance exceeds natural twilight level (about 10 lux). At twilight level, rods already get saturated. As a consequence, in well-lit environments (i.e., during photopic vision) human vision is cone-mediated and in dark environments, it is rod-mediated (i.e., scotopic vision).

There are further differences between scotopic and photopic vision that can impact night driving performance. Cones respond very fast to altering illumination [19], enabling the visual system to detect flicker exceeding 100 Hz in peripheral vision during bright light [20]. Furthermore, the highest contrast sensitivity of the cones at the fovea (i.e., a detection of 0.5% spatial or temporal contrast) is found under photopic light conditions. Under scotopic light conditions, contrast sensitivity of the cones is much lower (more than 5% contrast is needed). The actual sensitivity of the human visual system also depends on the light adaptation of the retina. Here, cones recover their photocurrent much quicker (within 20 ms after full bleaching) than rods (which take 20 min to recover their full contrast sensitivity; ,21). Therefore, in addition to the lower color sensitivity and the decreased spatial acuity, the lower sensitivity to temporal flicker and spatial contrast and the longer duration to recover full spatial contrast sensitivity are further disadvantages of scotopic vision that potentially contribute to decreased uptake of visual information at night, and hence to impaired driving performance during night driving compared to daytime driving.

To compensate for the lack of natural light at night, car headlamps and road lighting are used to illuminate the driving scene. The headlamps are especially important for drivers to detect and recognize unlit objects and hazards at night, for example, animals or pedestrians ([22-24]. Nevertheless, daylight illumination is never achieved. The illumination provided by headlamps falls mainly into the range of twilight illumination, meaning that there is enough light for cones and not too much light to saturate the rods so that both photoreceptors contribute to what is called mesopic vision [25]. During mesopic vision, complex cone and rod interactions may occur [26], for a review, see Stockman and Sharpe [27].

In addition to the low illumination, visual glare challenges drivers’ vision during night driving, and these factors interact with non-visual factors, such as drivers’ age or (age-dependent) visual impairments. For example, glare and light reflections impair object and hazard detection and recognition particularly in older drivers [28-30] However, also non-visual factors like alcohol consumption and fatigue that often coincide with night driving contribute to the risk of fatal crashes [2].

Although lower illumination at night provides challenging visual conditions to human drivers, humans are usually capable of relatively safe night driving. One trivial reason is that ambient light, headlamps, and lighted or reflective objects ensure that relevant visual information can easily be seen even during night driving. However, more interestingly, not all task-relevant visual information is equally affected by nocturnal light conditions. Object detection and recognition that depend on visual acuity at the center of the visual field are more degraded by the low illumination during night driving than lane and distance keeping, which do not depend on high visual acuity [31-33]. Thus, in many but not all respects driving performance is relatively robust against the visual challenges of night driving [34] and only an in-depth analysis of the task of driving will allow us to understand which visual information is task-relevant and needs to be recognized by the driver.

A historically early, straightforward formal description of the task of driving was provided by Gibson and Crooks [35]. These authors based their description of the driving task on a perceptually defined field of safe travel. In this task description, driving is conceptualized as locomotion through a field of space, and the field of safe travel corresponds to the part of the space through which locomotion is safe. The main purpose of locomotion is to move safely from one place to another. This is achieved by avoiding obstacles and other hazards based on visual information in the environment. The necessary adjustments of locomotion are considered to correspond to a visually guided path selection through the field of safe travel [35]. During driving, the field of safe travel would consist “of the field of possible paths which the car may take unimpeded” [35]. Therefore, its boundaries are obstacles (e.g., the shoulder of the road) which (potentially) impede safe travel. Such obstacles have a negative perceptual valence in a driving task, whereas the field of safe travel has a positive valence, especially its center (viewed from the perspective of the driver) where the boundaries are the farthest away. Furthermore, the model incorporates a minimal stopping zone, which is the area in front of the car through which the car would have to pass before it could finally be brought to a halt in case of an emergency stop. Figure 1 depicts an exemplary field of safe travel including the minimal stopping zone.

For safe driving, the minimal stopping zone should be within the larger field of safe travel, and a larger field-to-zone ratio should be preferred over a lower one. Importantly, the description defines the boundaries of the field of safe travel not only by physical obstacles but also by other situational circumstances, such as intersections, speed limits, traffic rules, including areas where obstacles could potentially appear. For example, the edge of the headlamp beam during night driving would be such an area. During night driving, new obstacles would become visible first when crossing the edge of the headlamp beam, and therefore, drivers should carefully monitor this area.

However, it is questionable whether the task description of Gibson and Crooks [35] is general enough to do justice to all situations that come up during driving. For example, at times, active search for particular information, such as for traffic signs indicating allowed maximal speed, could supersede the search for the boundaries of the field of safe travel. Such driving tasks be better conceptualized within general decision models, like the situation awareness model from Endley [36]. Situation awareness is based on a complete assessment and accurate comprehension of the situation, but this requires more than the processing of incoming visual information. Situation awareness is an understanding of the meaning of the visual information and how it relates to the goals of the person in a specific situation or task. Thus, it can be adjusted to cover a wider range of driving tasks.

The situation awareness model is more sophisticated than the model of Gibson and Crooks [35], as it includes not only environmental or visual but also organismic factors, such as the driver’s attention, goals, working memory, cognitive load, perceived complexity of the situation, automatization of actions, and mental models. The latter denote flexible semantic structures of drivers that would allow them to represent even theories, for example, about future events. This description of the task of driving highlights some factors that are also important for safe driving but were not reflected upon in the earlier model of Gibson and Crooks [35].

Endsley [36] defined situation awareness as a state of knowledge, dependent on but at the same time separate from the processes needed to acquire this state, which she labeled situation assessment. Additionally, Endsley [36] distinguished situation awareness from decision making and performance itself. Even with perfect situation awareness, drivers might make wrong decisions or show poor performance, but with inaccurate situation awareness, even perfect decision makers and performers would go wrong. Other important factors affecting situation awareness are attention and working memory, which in turn are influenced by the current workload, stress, and the perceived complexity of the situation.

Situation awareness consists of three hierarchical levels. At the first level, the person perceives elements of the environment. For example, a driver perceives obstacles, her own speed, etc. At the second level, a driver comprehends the current situation by understanding the significance of the perceived elements. For example, she understands whether other road users constitute a hazard or not. The second level is situation comprehension. It integrates the relation between the elements as well as the elements’ locations in the environment. Due to this difference between first and second level processing, inexperienced drivers might perceive the same elements as experienced ones but might fail to comprehend the situation appropriately if they do not understand the significance of the relations or the locations of the perceived elements for their goals. The first and second levels together enable the third and highest level of situation awareness. Understanding the significance of the perceived elements allows the person to anticipate the future of the situation and, therefore, to act more effectively and to make better decisions.

With the help of the situation awareness model of Endsley [36], night driving can be defined as a special case of driving in which it is more difficult for the driver to accomplish a high level of accurate situation awareness because of the decreased visibility of the environment at night. Furthermore, some persons are more affected than others by nocturnal light conditions. Thus, organismic factors might contribute to impaired situation awareness during night driving.

Related to the idea of sorting visual information into task-relevant and irrelevant information on the basis of task definitions, it is possible to discriminate between controlled and automatic forms of visual attention and eye movements [37]. In Trick et al.’s [37] framework, visual attention corresponds to the selection of visual information from the environment, a process tightly linked to eye movements (7, see also above). In this conception, the controlled forms of visual attention (being slow, effortful, intended, and more flexible) would correspond to the selection of task-relevant visual information. In contrast, automatic forms of visual attention (being fast, effortless, and unintended) could correspond to the distraction caused by task-irrelevant information, too. By adding a second, orthogonal dimension (origin of the selection principle: exogenous or stimulus-driven vs. endogenous or driven by prior experience and mental representations) to their attention framework, Trick et al. [37] discerned altogether four different types of visual attention: (a) reflexes (automatic-exogenous), (b) habits (automatic-endogenous), (c) exploration (controlled-exogenous), and (d) deliberation (controlled-endogenous).

After having described the visual situation at night and the task of driving, we next review existing eye-tracking research on real-world night driving. However, the studies reviewed used different theoretical rationales and sometimes lacked a clear theoretical rationale altogether. Therefore, it was not possible to summarize these studies with direct reference to any of the task-related factors discussed above. Instead, we categorized the reviewed studies by two orthogonal dimensions that most closely resembled important characteristics of the above models. One dimension that we used to categorize the reviewed studies was the environmental versus organismic origin of the factors modulating behavior. We asked if a factor that modulated eye movements during real-world night driving corresponded to a factor located in the environment (e.g., the headlamp condition used, the road geometry, the presence vs. absence of traffic signs, etc.) or whether it corresponded to a factor located within the driver’s organism (e.g., the driver’s age, her experience, etc.). This dimension connects to the exogenous-endogenous distinction of Trick et al. [37] and also allowed categorizing many of the influences listed in the situation awareness model of Endsley [36]. For example, we would have categorized Endsley’s elements of a situation, such as traffic signs or obstacles, as environmental factors, whereas we would have categorized task-relevant factors, such as cognitive load, as organismic.

The second dimension that we used to categorize the reviewed studies was the temporal inertia with which a factor exerted its influence. Some factors would have exerted a uniform influence for more extended durations – that is; the influences would have typically remained unchanged for minutes, hours, days or sometimes even years. Examples for such more stable or inert influences would be the weather conditions, the drivers driving experience or the age of the driver. Some automatic influences à la Trick et al. [37], such as a learned habitual tendency to look for traffic signs at roadside locations, would correspond to these stable factors. The influences of other factors could have varied more drastically in a relatively short time – that is, from moment to moment. For example, cognitive load by a demanding secondary task could have increased momentarily when a situation would have required driving and at the same time scanning the roadside for potential hazards. Table 1 provides an overview of different factors and how they roughly corresponded to the poles of the two dimensions of our categories.

Focusing only on research that measured eye movements during real-world night driving enabled us to review view all published literature exhaustively. We used a combination of the keywords “night” or “dark” or “darkness”, and “driving”, and “eye movements” or “eye tracking” to search in the databases Web of Science, PsycARTICLES, PsycINFO, TRID (includes Transportation Research Information Services and International Transport Research Documentation), and the Transportation Research Board Publication Index. Additionally, we extended our search to Deep Blue, the institutional repository of the University of Michigan (http://deepblue.lib.umich.edu), and the search engines Google Scholar, Google, and Bing.

In the next step, we scanned the results for studies that used a real-world setting and measured drivers' eye movements. In the papers that met this requirement, we looked for new references that we missed earlier. The final list contained 29 publications (see Table 2). We excluded studies that used eye tracking to measure the effectiveness of an algorithm [38] or very specific traffic interventions [39], such as reflective pavement markers [40] or ground mounted diagrammatic entrance ramp approach signs [41,42]. Furthermore, we excluded [43], because they investigated the gaze distribution related to the windshield without considering the traffic situation, and [44], who did not report the analysis of his real-world eye tracking data.

Our review is organized along the dimensions of our categories, starting with the stable environmental factors. In each section, we first quickly review the major findings concerning different factors in turn and then evaluate them against the background of driving models that we have presented in the Introduction. Only this evaluation allowed us to interpret what the findings from the eye-tracking studies might have meant regarding the cognitive processes involved and their potential benefits and harms for successful night driving.

Our review has shown that despite a lack of reference to a coherent description of the driving task in the studies reviewed, the findings can be interpreted in light of such descriptions. This was true for both driving models that we discussed in the Introduction: Gibson and Crooks’ [35] description of the driver’s task and the situation awareness model of Endsley [36]. For example, in line with Gibson and Crooks’ assumption of the importance of the boundaries of the field of safe travel, participants indeed spent much of their time looking at locations ahead of their cars in the direction of movement (e.g., [51]). During intersections, the drivers also spent more time looking at locations lying in the direction of their planned movements [67], a behavior that reflected the increased importance of the future boundaries of the field of safe travel. A shortcoming of the model of Gibson and Crooks [35] is its focus on visual signals alone. Even if a visual signal is seen, drivers must correctly interpret it for successful task performance. This was only acknowledged as important in Endsley’s situation awareness model. With its explicit reference to organismic factors of individuals (drivers), such as cognitive load, attention, and working memory, the situation awareness model does account for interindividual variability in task performance. This model was supported both by the finding that the performance on two simultaneous tasks – and thus, under overall increased cognitive load – diminished eye movements functional for safe driving [45] and by the general influence of organismic factors, such as drivers’ experience or drivers’ age, on gaze behavior (e.g., [96]).

However, sometimes a lack of proper control conditions left us uncertain about how to interpret the eye movements of the drivers. This was particularly true for the greater dispersion of eye movements during daytime compared to night driving [45-47] and during driving down roads in residential areas compared to driving down main traffic roads [52]. As long as the corresponding eye movements are not related to other task performance measures – but see Graf und Krebs [45] for an exception –, such as lane keeping accuracy, speed, or distance keeping to preceding traffic, it is difficult to assess whether the greater dispersion of fixations during the day was beneficial or detrimental to driving performance. Therefore, future eye-tracking studies of real-world night driving should incorporate task performance measures additionally to eye movements. At least, questionnaires to identify the intentions of the drivers should be included. This would also help to understand which gaze behavior was driven by goals and intentions and which was automatic and stimulus-driven alone. This distinction has been highlighted in the attention framework of Trick et al. [37] in which controlled (intended) and automatic causes of attention shifts and eye movements were distinguished. For example, in many studies it was not only unclear whether a factor’s influence on eye movements reflected beneficial or harmful effects on overall driving performance; sometimes it was also unclear whether an eye movement was due to an automatic salience effect or due to a controlled intentional strategy. As an example, think of the higher fixation rate on oncoming traffic [49,50] that could have reflected the attraction of the eyes by the salient movements of the oncoming cars or the controlled monitoring of an important characteristic of the current driving situation.

When we reviewed the eye movement findings of prior real-world night driving studies, we also noted that different subtasks seemed to be involved in driving: subtasks that could sensibly be distinguished from one another because performance on some of them was, for example, seemingly spared more by age than performance on other tasks. Therefore, we next propose a new description of the driving task in which we discriminate between three different groups of subtasks.

The first group of subtasks comprises the vehicle-related tasks. Their common denominator is that they all are primarily action-oriented and typically highly automatized [110]. These subtasks cover the operation of the gas and brake pedals, and the changing of gears. Experienced drivers almost completely automatize performance in vehicle-related tasks, like changing of gears [111]. It is sensible to distinguish these vehicle-related subtasks from other driving-associated tasks because visual perception is relatively indirectly related to the performance on the vehicle-related tasks. Therefore, the connection between eye movements and the performance on vehicle-related tasks could take different forms. First and foremost, the automatization of vehicle-related task performance could deliberate the eyes from monitoring performance in these tasks (i.e., of the feet or hand) and allow the usage of the eyes for the monitoring of the equally or more important road conditions.

In contrast to the vehicle-related tasks, road-related subtasks are guided by visual input in a relatively direct way. An example of a road-related sub-task is the distinct pattern of eye movements that is observed during the negotiation of right curves. To keep the lane, drivers fixate the right side of the road during right curves, which probably provides the necessary information to drive through the curve successfully [112]. Vehicle-related tasks, like steering, can show degrees of spatial compatibility between the motor control of effectors and the sensors, for example, when drivers gaze and steer in the same direction. Of further note, some road-related tasks, such as lane keeping, also seem to be spared by the loss of visual acuity in aged participants, possibly because performance in these subtasks does not require as much foveal vision.

Finally, during situation-related subtasks, the relevance of an object or a location for driving is determined by a more flexible interpretation of the situation. In general, experienced and inexperienced drivers show very different gaze patterns during driving, which probably reflects their different understanding of driving situations [86,91,113,114]. Another example is oncoming traffic at night. Oncoming traffic at night is not only a highly salient event (due to the headlamps of the oncoming traffic standing out against the dark surroundings), but it also triggers actions fitting to the situation, such as avoidance of looking into the headlights of the oncoming traffic or prompting the driver to manually switch from high beam to low beam to reduce glare-elicited drop in driving performance for drivers of the oncoming traffic (cf. [77]). Like in vehicle-related tasks, in situation-related tasks, the connection between particular eye movements and driving performance could be weak (e.g., to look away rather than at the headlights of oncoming traffic) but for a different reason than in vehicle-related tasks. Whereas in vehicle-related task performance automatization is the major reason why these tasks are not directly connected to particular eye movement patterns, in situation-related tasks the particular visual information would only be the basis for further, independent cognitive interpretations, akin to situation awareness à la Endsley [36].

Above, we have highlighted that future real-world night driving studies should best incorporate driving performance measures additionally to measuring eye movements. We believe that our description of the driving task through three different subtasks could be helpful in the context of such studies, as our task description would already provide theoretical reasons for expecting stronger links (e.g., between lane keeping and eye movements) or weaker links (e.g., between operating of the brake pedal and eye movements) between different variables. Thus, our task description could be helpful for formulating research hypotheses in the context of eye movements during driving. In particular, we believe that concepts such as automatization or compatibility, being deeply rooted in cognitive psychological research, could be helpful for integrating cognitive parameters into driving performance measurements.

In this review, we summarized, interpreted, and evaluated the meaning of eye movement patterns found in real-world night driving studies against the background of existing descriptions of the task of driving. This was done to decide which factors are potentially harmful to driving performance at night. We concluded that much of the drivers' eye movements could have reflected sensible scanning strategies (i.e., for the boundaries of the field of safe travel). However, their eye movements also indicated that, compared to daytime driving, night driving was associated with a more restricted distribution of fixations. Whether this reflected a harmful effect on night driving performance is not entirely certain, but it remains a likely possibility. Finally, we concluded with a proposal for a more fine-grained description of the driving task that could help to bridge the gap between eye movements and other dependent variables in driving research. However, its application has to await future research, as most studies reviewed did not relate eye movement to other task performance measurements for the evaluation of real-world night driving performance.

The authors declare that there is no conflict of interest regarding the publication of this paper.

ZKW Group GmbH funded this research during their cooperation with the University of Vienna for the project “Nachtfahrpsychologische Aspekte der Wirkung von Hyperscheinwerfern”. We thank Peter Hartmann, Christian Büsel and Elena Throm for helpful discussions on this review, as well as two anonymous reviewers whose constructive feedback helped to improve this paper vastly.