After scanning for other targets, gaze always returns to the road ahead. From inspecting the video it is clear that in terms of dwell time the (far) road would be the predominant gaze category. Note that the gaze almost exclusively seeks out the “far” road region, as opposed to the road immediately in front of the car. This is consistent with the two-level/two point control models (Donges, 1978; Land, 1998; Salvucci and Gray, 2004; Boer, 2016) that are based on the assumption that experienced drivers use gaze to obtain visual preview of road geometry used for anticipatory control (matching the predictable road curvature), as opposed to near road information for compensatory control (maintaining lane position against unpredictable perturbations). That there are very few fixations to the near road does not mean by any means that the driver would not be using visual information for stabilizing control, though. Indeed, it has been shown an experienced driver can monitor near road information peripherally (Summala et al., 1996), freeing the experienced driver to allocate more attention and overt gaze to the far road region than a novice (cf. Mourant and Rockwell, 1972; Lehtonen et al., 2014).

Of course, this behavior is also consistent with driving instruction frequently exhorting beginning drivers to try to look far enough ahead. But how far is far? There is no specific distance or time distance in the literature that would define “far” vs. “near” road. Typically, time headway to the “far” region in bends assumed to be about 1–2 s (cf. Lehtonen et al., 2014), which fits well with modeling literature as well (Boer, 2016).

In bends, it is conventional to use the lane edge tangent point (TP) as the point for segregating “near” and “far” road space (Land, 1998; Salvucci and Gray, 2004; Lappi, 2014). This is the point on the inside of the curve at which the visual orientation of the lane edge reverses its direction in the driver's visual field (Figures 2B,C). Note that the TP is a travel point (i.e., a point that moves with the observer in the 3D scene frame of reference, even though it may sometimes remain stationary in the observer's egocentric frame). Actual distance and time distance to the TP in the 3D world is therefore variable, and how far is “far” thus depends on curve geometry (time distance also depends on driving speed).

Observing the video, we can see that in simple bends the TP and Occlusion Point (OP) together with the lane edge opposite to TP create a Far Road “triangle” (Figures 2B,C). For a good portion of the time, this gives as good qualitative characterization of where we look when “we look where we are going” on the road¹. The Occlusion Point (OP) is the point furthermost part of the road to which a continuous, unobstructed preview of a possible trajectory (future path) is visible (Lappi et al., 2013a), i.e., the point where “the road disappears from view.” Like the TP, OP also is a travel point; it moves ahead as the observer travels along the road and does not follow the local optic flow (Figure 3; clear illustrations are e.g., 17:28:16.79–17:20:28.15 and 17:30:40.29–17:30:46.02).

The driver's visual field does not only cover the external 3D roadspace, of course, but also relevant targets in the vehicle frame. These include the instrument panel (speedometer, tachometer) and mirrors (rear view and side mirrors). These are scanned predominantly (but not exclusively) on straights (e.g., 17:30:46.02–17:31:06.26; 17:27:57.13–17:28:02.61), presumably because there is less task load than in curves (Tsimhoni and Green, 2001), and less need for visually monitoring the road ahead. Stabilizing steering control requires little overt gaze, and the much longer time headways to visual occlusion means that there is much less time pressure for spotting hazards.

Oncoming vehicles, or pedestrians/bicyclists coming the other way and being overtaken are monitored by fixations. (e.g., 17:22.09–17:22:16.53; 17:29:06.83–17:29:10.96 [cars emerging from blind bends]; 17:32:57.75–17:33:05.49 [coming up on two bicyclists simultaneously]).

Whenever there is a side road or a road from a yard that intersects the road, the driver tends to scan it with a fixation or often multiple fixations (e.g., ts 17:20:34.56–17:20:40.79 [side roads on both sides of a straight]; ts 17:22:08.55–17:22:09.02 [sideways glance to a side road on the left side of a right hand bend]).

Road signs are fixated—even from quite impressive distances. These include road signs proper (speed limit signs poor road surface and bends caution signs, stop sign), as well as street name and navigational instruction signs (e.g., 17:20:17.47–17:20:21.52 [speed limits, bends]; 17:20:25-48–17:20:28.40 [bumpy road]; 17:34:08.52–17:34:25-72 [multiple]).

Post boxes and other mid-sized objects near the road are occasionally “checked out” (e.g., ts 17:24:37.57–17:24:38.45). Here the driver is likely using high-spatial-resolution foveal vision for detailed analysis and object recognition of a target already localized and individuated as distinct from the ambient background using peripheral vision.

There are actually very few fixations at “scenery” not covered in the above categories. This is in itself important. The absence of any significant number of fixations on the general scenery, that is, the concentration of gaze on specific target and the apparent absence of any visual search in itself indicates that peripheral visual information is used in a very efficient way to guide the gaze at the relevant locations with high accuracy and reliability. Note that we therefore prefer to use the terms visual exploration or scanning rather than visual search, see further discussion below.

In sum, the general pattern of gaze coordination is the following: First, the default gaze mode is “eyes on the road,” or “looking where you are going.” Second, glances elsewhere are performed when a specific relevant target to look at has been identified (also, there must be “spare capacity” to allocate gaze time to non-steering related targets).

With a case study approach cannot tell from the data alone whether any stereotypy observed is idiosyncratic to the participant. But when we find repeatable patterns that are reported in the literature we can consider them general.

Scanning the Far Road in bends. Perhaps the most robust coordination pattern is the fairly systematic scanning of the Far road (Figures 2, 4) in bends. As the Occlusion Point travels up the road, revealing more of the scene behind (Figure 3), the gaze seeks out the road surface/inside road edge emerging into view.

This orientation toward the inside edge of an upcoming bend (the apex region) is sometimes called “tangent point orientation.” But when one observes the actual scan pattern, it should clear there is much more complexity involved than the driver aiming to stabilize gaze on a single point. What does the full picture of scanning the far road in bends look like, then? TP orientation occurs, especially in the approach and entering phases of the bend (Land and Lee, 1994), but there are also fixations to lane edges further ahead, beyond the tangent point, and to the road surface in the Far Zone region. These guiding fixations (GF) are interspersed by saccades to make look-ahead fixations (LAFs) further into the bend and beyond. When the view up the road is occluded, as is the case on this road which runs through woods, some of these LAFs reach the OP (see Figure 4), but not all. As the driver enters the bend, fixations of the TP and the inside lane edge beyond the tangent point, and fixations to the road surface in the Far Zone, mainly beyond the tangent point continue, interspersed by LAFs further up the road.

Because of this rich pattern and multiplicity of gaze targets, it is better to reserve the term tangent point orientation to fixations at or very near (within 3° of) the tangent point itself, performed at the very end of the approach and beginning of the entry (i.e., straddling the turn-in). This is according to the definition in the original Land and Lee (1994) study². As a methodological side comment, note that because of the visual projection of road geometry into the forward-looking visual field can bring these points very near to one another, making the definite determination of the actual gaze target of many individual fixations difficult, and traditional AOI methods unreliable (Lappi et al., 2013a; Lappi, 2014). But while in many individual cases the classification of a given single fixation could be ambiguous bases on instantaneous gaze position alone, the overall pattern of multiple gaze targets in the Far Zone is clear.

In the general case, due to the highly dynamic character of the way the road surface presents itself in the visual field (Figure 4), Far road fixations cannot be separated from LAFs by any hard-set distance or gaze angle criterion³. Instead, we propose that LAFs should be defined by the following return saccade. That is, in a LAF a fixation on the TP/far road/road edge is followed by a saccade further up the road (but not necessarily all the way to the OP), and a return saccade back to the road surface/lane edge closer to the vehicle (i.e., gaze polling, Wilkie et al., 2008). This “zig-zagging” pattern is very evident in most bends—for example the right hand bend in the beginning of Supplementary_Movie_2_17_21_31_rl.mp4 and the left hand bend in Supplementary_Movie_3_17_23_48_l.mp4—and we have seen it to occur more or less frequently in the raw data of every driver we have ever tested in our previous studies.

Although, the scanning patterns in driving are sometimes called “search” (primarily when complex situations call for identification and interpretation of multiple potential hazards e.g., Underwood et al., 2003; Wolfe and Horowitz, 2004), for the analysis of the core cognitive requirements of driving (high-speed vehicle control) we prefer the term scanning or visual exploration, as there is hardly any evidence of visual search proper—at least not in the way the term is used in experimental psychology. In visual search paradigms eye movements are used to look for a target among distractors, where the target is masked by the clutter and there is hence substantial uncertainty over target location (as in a typical Feature Integration Theory paradigm search matrix, or a Where's Waldo? image). The way the fixations appear to immediately find specific targets even at quite impressive distances (the road, other road users, traffic signs, see e.g., fixations to the three traffic signs in Supplementary_Movie_2_17_21_31_rl.mp4) suggests this is not the case in driving, as in fact we see very few fixations to “scenery” (general observation G7).

Intermittency combined with a high concentration of gaze on relevant targets (lack of search) implies efficient peripheral vision processes for target identification and saccade planning. Here we should bear in mind, though, that the participant is an expert driving school instructor, for whom traffic signs, for example, are highly relevant in terms of carrying out in-car instruction. Thus, he might have superior parallel covert “search” strategies compared to more typical drivers. Whether traffic signs are as “salient” to everyday drivers in (terms of being able spot them from distance with peripheral vision and “attracting” gaze) is not clear.

Let us return to the far road fixation targets in curve driving. How can the individual fixations in the scanning pattern be interpreted? This is an important question, because the reason most eye movement research focuses on fixation behavior is that fixation is considered functionally as the “window” when new visual information is available to the brain, punctuated by saccades during which relatively little information is received, and analysing where and when fixations are made is taken as a road to inferring underlying cognitive processes. We would like to point out here that for a moving observer, a fixation tracking a fixed target in the scene—and most travel points as well—is, from an oculomotor point of view, a pursuit movement (see discussion in Lappi, 2016).

In very broad terms, far road fixations can be considered simply “looking where you are going.” But interpreting this strategy—why we should look where we are going in the first place—needs to take place in terms of how the brain processes the information gleaned from the fixation(s), and uses it real-time control of gaze and locomotion. Overt behavior (gaze position) does not uniquely specify the information that might be gleaned, and many interpretations for the (guiding) far road fixations have been put forward (for review see Wann and Land, 2000; Land and Tatler, 2009; Lappi, 2014). A detailed discussion of all the different interpretations, and their underlying theoretical motivations and commitments, are beyond scope of this paper. References to key papers are given in Table 3. The fixation classification scheme (Table 3, Figure 4) is intended to be compatible with any and all theoretical interpretations, i.e., not committed to any particular theoretical viewpoint or interpretation.

Far Path targets (Table 3, class 2) are postulated in several steering models to be steering points, as is the tangent point (class 5). (While the travel points and waypoints on the path would instantaneously occupy the same location, they will generally move in different directions in the visual field, cf. Figure 3, they will be useful for quite different steering strategies, see (Lappi, 2014) for detailed discussion). Look-ahead fixations (classes 3,4), LAFs on the other hand are considered to be different from such guiding fixations, because the tight gaze-steering coordination needs to be uncoupled during a LAF. They may nevertheless support higher-level trajectory planning (see L6, below). Note that we have used the term Path Edge (classes 5 and 6), which we define as those parts of the lane edges, in the Far region, which at any particular moment in time constrain available paths. (We reserve the term lane edge for the entire edges of the driver's own lane, extending beyond these “path edge” regions—even beyond the current field of view ahead and behind the vehicle). Road edge (classes 5 & 7) refer to the edge of the opposing lane.

Driving is a self-paced task in that the driver has a choice in the speed s/he wishes to travel at. However, once a speed is chosen the targets emerge at a given pace and obsolescence rate (cf. Senders et al., 1967; Kujala et al., 2015). This places a high importance on the accurate timing of fixations and saccades.

At an aggregate level this temporal coordination is reflected in a robust ca. 1 s gaze-action delay. Chattington et al. (2007) report gaze lead time (peak of gaze-steering cross correlation) of 0.98 s for 60 s epochs. We have observed similar cross correlations on the same road as used in this study for seven participants (unpublished data from Lappi et al., 2013a).

At an individual fixation level, judgments of gaze-action delay depend on an interpretation of which action(s) each individual fixation actually supports. For steering related guiding and look-ahead fixations this question is still unresolved (cf. previous point). The most frequently referred to phenomenon remains the final fixation at or near the tangent point before turning into a bend (Land and Lee, 1994). This is clear for example in the blind right hand bend in Supplementary_Movie_4_17_30_39_r.mp4. However, given that the TP region is frequently fixated several times in anticipation (not just “just in time”), and that other locations in the Far Road Triangle are fixated in curve negotiation, the full picture of how fixation timing and locomotor action timing are related remains unresolved (cf. the next point).

The just-in-time strategy also implies the visual system needs to be able to recover either from peripheral visual information or from memory where the need-to-know information is at any moment in time. Cf. discussion of lack of search above (L2) and the role of memory below (L6).

We remind the reader of the general observation that our driver never “stares” at any particular location or object for any extended period of time. Scanning the scene with the rapid saccade–fixate–saccade pattern happens all the time. That is, the overall pattern gives the intermittent “feel” that is characteristic of visual exploration (and exploratory behavior in other modalities, e.g., manual exploration).

Intermittency is clear for example in Supplementary_Movie_5_17_30_55_lr.mp4 in the way looking “where you are going” is interspersed with fixations to a traffic sign and other road furniture, an intersection on the right, and the side mirror. Also in the right hand bend in the beginning of Supplementary_Movie_2_17_21_31_r.mp4 where the fixations in the Far Road region are interspersed by look-ahead fixations and sideways glances at a traffic sign. What are the implications of such a sampling pattern for control and cognitive processing?

Visual steering control models in psychology (for a review see Lappi, 2014) and driver models in vehicle dynamics engineering (for a review see Macadam, 2003) generally do not address this intermittency in visual input (but see Johns and Cole, 2015 for discussion and one of few experimental studies investigating the effects of intermittency of visual input to steering control; cf. “active gaze” approach in artificial intelligence and mobile robotics, Ballard, 1991; Fermüller and Aloimonos, 1995). Rather, the current state-of-art approaches in sensing and control are typically “reactive” systems, i.e., the system passively receives continuous input and produces output control signals in response. In contrast, in psychology and cognitive science it's well established—and apparent also in the present data - that in active dynamic tasks humans “proactively” sample visual information as needed, leading to input that is intermittent, and determined by the active observer (e.g., via eye movements) rather than imposed by the environment as a “forcing function.” This allows humans to transcend their relatively slow information processing and limited sensory resolution to achieve impressively high aptitude in high-speed steering control (in driving and other domains).

On the other hand given that some relevant visual information may be available not continuously but as discrete fixations, critical action decisions may only be doable at certain points in time, or at least there are likely to be limited optimal “windows” for timing locomotor action initiation relative to oculomotor actions. This issue has been studied in e.g., sports psychology in the literature on the quiet eye phenomenon (Vickers, 2016); but so far it has not been studied in the driving domain, nor modeled in driving models, beyond the above mentioned general 1 s lead time (cross correlation) between “apex orientation” and steering.

Memory processes cannot be readily “read off” from gaze behavior in our locomotor task—especially given that the route was only driven once (observing change in gaze behavior over multiple runs would be more informative of memory processes, as would analysis of landmark use in familiar surroundings, cf. Spiers and Maguire, 2007, 2008). Also note that the fact that in the present data even first fixations on a target are achieved without search, based on peripheral information (cf. discussion on L2, e.g., traffic signs) means that fixation-without-search in highly familiar surroundings (such as one's kitchen; Tatler and Land, 2011) cannot be interpreted as proof of memory use. Also the just-in-time fixation strategy (L4) emphasizes the online nature of eye-hand-body coordination and “letting the world be its own model” (cf. Brooks, 1991)—as opposed to maintaining information in memory (which requires cognitive resources and faces the problem of that information becoming obsolete).

On the other hand memory and intermittency (L5) are deeply connected at a theoretical level, because it is to a large extent intermittency that makes memory (as opposed to pure online control) powerful. Anticipation allows humans to transcend their relatively slow information processing and limited sensory resolution. Modern cognitive theories of skilled action are predicated on the hypothesis that humans make predictions of the immediate future, choose actions on the basis of these predictions (for reviews of this predictive approach to anticipation and control, see e.g., Bubic et al., 2010; Henderson, 2017; for a critique of predictive control and defense of anticipation from merely prospective control see Zhao and Warren, 2015).

The relevant memory processes would be navigational long-term memory—an area of intense active research in the cognitive and computational neurosciences (Spiers and Barry, 2015). Integrating this literature to the theory of skilled driving would significantly advance out understanding of the driving task, and the role of these representations in real-world tasks generally.

Here we suggest one of the roles of look-ahead fixations—justifying the strategy of taking gaze time away from imminent needs of the primary control task—is to maintain and update this trans-saccadic memory of scene layout. That is, LAFS are relevant for steering (with a substantially higher delay than the 1 s for guiding fixations)—both for selection of, or parameter setting for, motor plans (updating “inverse models” in control theoretical terms), and creating a richer internal (forward) model of the state of the environment and the prediction of likely effects of action.

Precisely what kind of trans-saccadic memory underlies spatial orientation, and maintains our coherent experience of space (cf. Land and Furneaux, 1997; Tatler and Land, 2011; Burr and Morrone, 2012; Spiers and Barry, 2015) is an important but underappreciated issue in understanding (expert) driver behavior

One of the first remarks a number of people have made upon viewing the video (Supplementary_Movie_1_full_video.mp4) is expressing surprise at how much the participant's head moves. We experience the world as stable, even when the platform from which we observe it is not static. To achieve this visual stability in mobile contexts the brain must be able to take into account, in very sophisticated ways, both controlled (active, predictable) head movement and (passive, unpredictable) perturbations (Angelaki and Hess, 2005; Tatler and Land, 2011; Lappi, 2016).

Compensatory eye movements (vestibulo-ocular and optokinetic responses) and compensatory head movement stabilize gaze stable against unpredictable perturbances. Eye- head coordination is, on the other hand, guided by top-down attentional processes when target motion is predictable. A target moving in the visual field is tracked (pursuit) and large gaze shifts (saccade) achieved in part by synergistically turning the head, not just rotating the eyes in their sockets.

Synergistic eye/head gaze shifts are most apparent in fixations to the side mirrors. Pursuit of roadside objects, on the other hand, is usually done with eyes only, even for high eccentricities. In contrast, pursuit of OP is accompanied by head rotation anticipating vehicle rotation—even though the eccentricity is small, the gaze rotation anticipating or guiding locomotor rotation seems to recruit also head rotation, suggesting it is not only stimulus eccentricity but also stimulus relevance to ongoing motor action that is important in eye–head coordination. Even saccades and eccentric pursuit of non-locomotor targets can be done without a substantial head component (i.e., the head is kept aligned with locomotor path, side roads are checked with a sideways glance, for example), whereas orienting to the locomotor path (presumably for preview guidance information) involves a substantial head component—even when the eccentricity of the locomotor target is small. That is to say: the head is tightly coupled to steering-related “guidance” but less tightly coupled from non-steering related “scanning”—the head/gaze coupling is also intermittent.