Visual search of driving situations: Danger and experience

Transcription

1 Perception, 1998, volume 27, pages Visual search of driving situations: Danger and experience Peter R Chapman, Geoffrey Underwood Department of Psychology, University of Nottingham, University Park, Nottingham NG7 2RD, UK; peter.chapman@nottingham.ac.uk Received 15th August 1997, in revised form 27 May 1998 Abstract. Previous research on visual search in driving suffers from a number of problems: small sample sizes, a concentration on mundane situations, and a failure to link results to more general psychological theory. The study reported in this paper addresses these issues by recording the eye movements of a large sample of drivers while they watched films of dangerous driving situations and comparing the findings with those from more general studies on scene perception. Stimuli were classified according to the types of road shown and the degree of danger present in the scenes. Two groups of subjects took part, fifty-one young novice drivers who had just gained a full driving licence and twenty-six older more experienced drivers. Dangerous situations were characterised by a narrowing of visual search, shown by an increase in fixation durations, a decrease in saccade angular distances, and a reduction in the variance of fixation locations. These effects are similar to the concept of 'attention focusing' in traumatic situations as it is described in the literature on eyewitness memory. When road types are compared, the least visually complex rural roads attracted the longest fixation durations and the shortest angular saccade distances, while the most visually complex urban roads attracted the greatest spread of search but the shortest fixation durations. Differences between the groups of subjects were also present. Novices had longer fixation durations than experienced drivers, particularly in dangerous situations. Experienced drivers also fixated lower down and had less vertical variance in fixation locations than novices. 1 Introduction It is widely accepted that deficiencies in visual attention are responsible for a large proportion of road traffic accidents (eg Sabey and Taylor 1980). There are, however, surprisingly few widely replicated findings from studies on the eye movements of drivers. Moreover it is difficult to generalise from existing research to the kinds of dangerous situations which actually cause road accidents. An additional problem is that much of the research on the eye movements of drivers has been purely atheoretical with little attempt to link results to the existing literature on eye-movement recording in other domains. Although the bulk of eye-movement analysis has been conducted in reading research, there is also a body of research on eye movements in the viewing of general scenes. Since the work of Buswell (1935) it has been accepted that viewers concentrate their fixations on the informative parts of visual scenes thus subjects viewing artwork tend to fixate on people rather than background regions of pictures. Viewers are able to identify the informative regions of a scene rapidly and without ever fixating uninformative regions (Antes 1974; Mackworth and Morandi 1967). Informative regions of a scene are fixated for significantly longer than less informative regions (Antes and Penland 1981; Friedman 1979; Loftus and Mackworth 1978) and it is clear that these times are largely dependent on the task the subject is performing (Yarbus 1967). In general, fixation durations provide a good measure of the information content of the region being fixated and the strategies being used by the viewer in processing information from the region (though cf Henderson et al 1997). 1.1 Eye-movement recording on the road At the simple level drivers' visual search can be described as concentrating on a point near to the focus of expansion (the point in the visual field in front of the driver where objects

2 952 P R Chapman, G Underwood appear stationary) with occasional excursions to items of road furniture and road edge markers (eg Helander and Soderberg 1972; Mourant and Rockwell 1970; Shinar et al 1977). This reliance on the focus of expansion in the scene has been assumed to be because it provides precise directional information to the driver and is the location near to which future traffic hazards are likely to be first visible. In the context of general scene viewing it can also be regarded as being the generally most informative part of the visual scene in terms of the concentration of objects. Increasing the complexity of the visual scene (by adding vehicles, road furniture, or irrelevant signing) increases the number of eye movements made and decreases the mean fixation durations on individual objects (Erikson and Horberg 1980; Luoma 1986; Miura 1990; Robinson et al 1972; Rutley and Mace 1968). This seems to be a natural response to having more objects available in the visual field to look at. However, it is not clear whether decreases in fixation durations mean that objects are processed incompletely or that redundant fixation time is simply reduced. Cohen (1981) found that subjects viewing slides in the laboratory adopt longer fixation durations than those actually driving a vehicle in the same situations. He concludes that this is because of the lack of time pressure in his laboratory task and argues that on the road subjects adopt more task-relevant strategies and pick up more information per unit time. Whether or not this is the case it suggests that fixation durations while driving may be determined both by the visual stimulus and by the task as perceived by the subject. The eye-movement patterns become slightly more complex when the driver is required to negotiate a curve. Drivers generally adjust their fixation locations to maximise their sight distance and provide information about the future curvature of the road (Helander and Soderberg 1972; Shinar et al 1977). In many cases this means fixating on the tangent point made by the driver's line of sight ahead to the inside of the curve (Land and Lee 1994), though information about lane position from closer to the current position of the vehicle also seems to be necessary for accurate curve following (Land and Horwood 1995; McLean and Hoffmann 1971, 1973). In such situations it appears that the most informative region of the visual scene has changed, possibly because of the need for visual input to allow steering control, but possibly just because of the occlusion of objects by the road edge. Although extended fixations near the tangent point are frequently observed, these may be largely a function of the lack of alternative sources of visual information, a possibility which is supported by large individual differences in the number of off-road features which drivers choose to fixate (Land and Lee 1994). The research of Mourant and Rockwell (1972) on the relationship between experience and eye movements in driving is widely cited as demonstrating differences between experienced and novice drivers in their visual search, specifically that novices concentrate their search in a smaller area, closer to the front of the vehicle than experienced drivers do (eg Evans 1991). Although this result is accepted wisdom for driving instructors (eg Miller and Stacey 1995), it is perhaps surprising from what we know about scene perception. Scene viewers fixate objects rather than regions, and do not fixate between objects even when it would be beneficial for them to do so (Henderson et al 1997). Differences in mean fixation locations may therefore correspond to differences in the objects that are fixated and the time at which they are fixated rather than representing a preference for one region of space over another. There are also important caveats to the Mourant and Rockwell (1972) results. The study involved only ten subjects, four experienced drivers and six novices, and the novices had no more than 15 min previous driving experience before taking part in the first phase of the study. Although the novices were found to scan less widely in the horizontal axis than the experienced drivers, training of these novice drivers appeared to make their horizontal scanning behaviour less rather than more like that of experienced drivers. Mourant and Rockwell (1970) also found that substantial reductions in the spread of visual search could be achieved simply by having drivers repeat the same route three times. The reported

3 Danger and experience 953 difference in vertical gaze location (novices looking closer to the front of the vehicle) was only actually reliable at the p < 0.05 level in one of the nine driving subtasks described in the study. Mourant and Rockwell (1972) did note a tendency for pursuit movements to decrease in frequency as a result of training. This result is one which would clearly be predicted from scene-perception research on the basis that information acquisition is speeded as a function of experience, meaning that pursuits need to be sustained for less time. While these reported differences in visual search between novice and experienced drivers remain plausible and interesting, the evidence for their ubiquity across varying levels of experience and traffic situations is still relatively slight. Some differences in gaze location may be caused directly by control-skill limitations during the earliest stages of learning to drive. Thus it is possible that until a driver has learned to control steering with information from peripheral vision frequent fixations are necessary on the edge of the road to determine road position. From a perspective of reducing accident statistics these early stages of driving are of relatively little interest. For example Forsyth et al (1995) found accident rates while learning to drive to be less than 0.6% compared with the 18% accident rate for drivers in the first year after passing the test. It can be argued that the current British practical driving task is able to screen out candidates with grossly inadequate control skills it may be much harder for examiners to judge the adequacy of a candidate's visual-search strategies. The types of differences that would be particularly interesting are those caused by the experienced driver's increased knowledge of the road environment. It is likely that substantial traffic experience would allow drivers to predict the locations of potential hazards and modify their search strategies accordingly. In support of this hypothesis Theeuwes (1996; Theeuwes and Haganzieker 1993) has demonstrated the importance of top - down processes in the perception of traffic scenes. In these experiments experienced drivers demonstrated considerable impairment in performance when stimuli appeared in unpredictable locations. These results may explain differences in the on-road eye movements of novice and experienced drivers reported by Underwood et al (1997). In this study experienced drivers were found to have increased horizontal search and decreased fixation durations relative to novices in demanding dual-carriageway driving. These results can be interpreted in terms of experienced drivers having greater knowledge about the potential locations for threat-related information. However, in on-road driving there is always the danger that what is being measured is some form of interaction with the driver's level of control skills (eg novices finding lane maintenance difficult at high speeds) or the type of situation that the driver actually gets into (novice drivers creating dangerous situations by virtue of poor driving). These problems can only be resolved by laboratory studies of driver behaviour. Miltenburg and Kuiken (1990) avoided many of the problems with on-road studies by having drivers watch video recordings of six common traffic situations and recording their eye movements. They tested forty-seven subjects split into four groups on the basis of their driving experience, ranging from novice drivers with less than 1 year of driving experience, to very experienced drivers with more than 5 years of experience and more than km driven in the previous year. They found that for one of their scenes, crossing an intersection, experienced and very experienced drivers fixated more briefly than inexperienced drivers who in turn fixated more briefly than novice drivers. They found no evidence supporting their other two experimental hypotheses, that novice drivers might fixate closer to the front of the car, or that experienced drivers might fixate relevant objects sooner. They did find a number of differences between their groups in an a posteriori analysis, particularly noting that novice and inexperienced drivers spent longer fixating near to the vehicle when the film showed the car negotiating a bend. Other differences were present but did not change across the four groups in a way that suggested a relationship with traffic experience.

4 954 P R Chapman, G Underwood 1.2 Visual search in dangerous situations The importance of danger in the judgment of traffic scenes is emphasised by a number of studies in which a relationship between people's abilities to detect hazards in driving scenes and their accident involvement is sought (Pelz and Krupat 1974; Quimby and Watts 1981; Quimby et al 1986). In such research it is assumed that the time taken to detect hazards in driving scenes and make a manual response represents the kind of safety margin that would be available to the driver on the road. Two recent reviews of the literature on individual accident liability have both concluded that such hazardperception abilities represent the most promising perceptual or cognitive predictors of road-traffic-accident involvement (Elander et al 1993; Lester 1991). McKenna and Crick (1994) additionally report that hazard-perception ability improves with the transition from novice to experienced-driver status and further improves with the transition to expert driver as a result of specific training techniques. There are thus good reasons to anticipate that novice and experienced drivers should differ considerably in their visualsearch strategies in dangerous situations, and that these differences may be important in predicting and possibly reducing accident involvement. There is also good evidence that visual attention and eye movements may be systematically altered in dangerous situations. Since the work of Easterbrook (1959) it has often been suggested that stress or arousal may cause a narrowing in the range of cues attended to by an organism. This work has recently become important to researchers in the area of eyewitness testimony interested in the phenomenon of 'weapon focus' (Kramer et al 1990; Christianson 1992; Loftus 1979). Here it is suggested that the emotional stress experienced by a crime victim may cause a narrowing of attention and subsequent failure to remember peripheral details about the crime situation. Thus Loftus et al (1987) found that when subjects view stories in which a customer in a shop holds either a gun or a cheque subjects fixate significantly longer and more often on the gun than they do on the cheque. Christianson et al (1991) have reported similar results for eye fixations on a slide showing a woman involved in an accident while cycling, with more and longer fixations on central information in the emotionally arousing conditions and impaired memory for peripheral information from the slide. These results cause us to predict that there may be attention focusing in dangerous driving situations. This could be characterised as longer or more frequent fixations on central information, and fewer or shorter fixations on peripheral information. Studying the eye movements of drivers in dangerous road situations raises a number of practical and ethical difficulties. Even when everyday road situations are used there are significant problems attempting to research eye movements during actual driving. Real traffic situations will always differ from one subject to the next, and the subjects have control over aspects of the situation likely themselves to interact with eye-movement measures (ie control use, speed, lane position, following distances). These problems can be reduced by exploring behaviour in a driving simulator but there are important methodological problems even here. A number of studies have suggested that, while eyemovement patterns in the laboratory may be a reasonable reflection of behaviour on the road (eg Hughes and Cole 1986a, 1986b), dynamic rather than static scenes are required (Cohen 1981), and the fidelity of reproduction of the environment is extremely important (Staplin 1995; see also Henderson and Hollingworth 1998). This latter point is of particular importance when hazardous scenes are to be used. The anticipation of danger in real driving situations may depend on quite subtle cues which are difficult to identify and produce accurately in a driving simulator. There are thus considerable advantages to using filmed driving scenes in the manner of Miltenburg and Kuiken (1990) even though vehicle control is not required in such circumstances. Cohen's (1981) study and our own earlier discussion does nonetheless suggest that a degree of time pressure is necessary if subjects are to adopt realistic visual-search strategies when watching

5 Danger and experience 955 driving scenes in the laboratory. In the study we describe below this is created by requiring novice and experienced drivers to identify hazards in a film as quickly as possible while their eye movements are recorded. Two different aspects of viewing driving scenes are investigated in the study in addition to looking for group differences between novice and experienced subjects. One of these is the effect of different types of road on visual search. The anticipated effects of visual complexity on eye movements (Henderson 1992; Henderson and Hollingworth 1998) suggest that this should be systematically varied; we do this by dividing stimuli into rural roads (with relatively few stationary or moving objects) and urban roads (high in visual complexity). This comparison emerged independently as the most important dimension in the classification of road scenes in a study by Riemersma (1988). An additional category, suburban roads, is included as representing situations of moderate visual complexity. The second effect of interest is the possibility of observing attention focusing in dangerous situations. We have thus divided our stimuli into portions containing dangerous events and portions which are relatively safe. We would predict that these dangerous regions might involve longer or more frequent fixations on central information and fewer eye movements to items in peripheral areas of the scene. 2 Methods 2.1 Subjects The subjects were seventy-seven drivers in one of two categories, young novices, who were all tested within three months of gaining a full driving licence, and older experienced drivers who had held a full driving licence for between 5 and 10 years at the time of testing. Novice drivers were recruited with the aid of the Driving Standards Agency who distributed questionnaires to candidates who were successful in the practical driving examination. The older experienced drivers were recruited from responses to an advertisement in a local paper. These seventy-seven subjects were part of a sample of one hundred and twenty-seven drivers taking part in experiments in our laboratories. Some of the larger sample wore glasses or contact lenses, or proved difficult to calibrate on the eyetracker. The seventy-seven described here are those for whom good eye-tracking data were obtained. These consisted of fifty-one young novices, twenty-one females and thirty males, and twenty-six older experienced drivers, thirteen males and thirteen females. Subjects ranged in age from 17 to 30 years, with the mean age of the young novices being 18 years and the mean age of the older experienced drivers being 27 years. Young novices reported having driven a mean of 1047 miles in their driving careers (range 70 to 6000) while the older experienced drivers reported having driven a mean of miles (range to ). Subjects were paid a total of 20 for their participation in this experiment along with a number of other experimental tasks being conducted during the same session. 2.2 Stimuli and procedure The stimuli were a set of thirty-nine short film clips lasting between 18 and 73 s each. These films were divided into three sets containing thirteen films each and approximately matched for content. The films showed driving situations recorded from the driver's point of view and were designed to contain a number of potentially dangerous events such as bicycles pulling out suddenly, pedestrians emerging from behind parked cars, or cars ahead braking suddenly. Typical examples of such films were as follows, (i) Rural scene: view proceeds down narrowish empty country road with trees and fields on either side. After 20 s a horse and rider become visible around a corner and the driver passes the horse in a situation with restricted visibility. (ii) Suburban scene: driver's view follows another vehicle down a relatively empty suburban road with occasional oncoming vehicles and parked cars. The car ahead brakes suddenly and turns unexpectedly down a side street.

6 956 P R Chapman, G Underwood (iii) Urban scene: view proceeds down busy shopping street with parked cars and pedestrians on either side. A parked vehicle on the left suddenly reverses out into the road. The films were presented as MPEG video on a computer monitor 1 m from the subject, subtending 15.4 deg by 11.6 deg of visual angle. This represents a compression to approximately 50% of the angle the images would have subtended in actual driving. Eye movements were monitored by using an SRI Dual Purkinje Generation 5.5 eye tracker produced by Fourward Technologies. Calibration was monitored during the presentation of films and an opportunity for recalibration was available before the presentation of each new film. Head movements were reduced by using a head restraint and chin cup. Subjects were tested individually in a session lasting approximately 20 min. Each subject watched one of the three sets of thirteen films, selected at random. They were informed that they should watch the films as if they were the driver of the car and press a response button as soon as possible if they saw hazardous events coming up. Hazardous events were defined as the kind of event that would require them to brake or take some form of evasive action. Subjects were informed that each film would contain at least one hazardous event but some might contain more than one such event so they should keep watching until the end of the film in case further hazardous events occurred. The order of presentation within each set of films was randomly chosen for each subject. Eye-movement data were recorded directly onto the computer used to present the video stimuli. This computer also recorded the timing of response-button pressing by the subject. 2.3 Design Each film contained at least one dangerous event, some films contained up to four such events. The detectability of these events was confirmed by a pilot study in which fifty-four drivers watched the films while simply making hazard responses. Temporal windows were defined in each film around these dangerous events starting 1 s before the event and lasting until the objects involved were no longer visible. Such windows accounted for an average of 6.2 s in each clip, ranging from 2.5 to 16.0 s. Data were analysed separately within these 'danger' windows, and over the relatively 'safe' remaining portions of the films. The films included a wide variety of different road types and traffic participants. Stimuli were divided into three different sets of thirteen films, balanced so that each set contained the same number of films of each type of road environment. Four films of urban and suburban environments, and five of rural environments appeared in each set of thirteen. The experiment therefore includes one between-subjects factor, driver group (old experienced versus young novice), and two within-subjects factors, road environment (urban, suburban, rural) and danger window (recordings within the 'danger' window versus those from 'safe' portions of each film). 3 Results 3.1 Button responses Two general measures of button responses were calculated for all subjects. The first measure was the latency between the hazard window beginning (1 s before the hazard actually appeared) and the next button response from the subject. This was calculated separately for the three different types of road environment. This measure is only applicable within danger windows so only one within-subjects factor is considered. There was a significant effect of road environment on latency (i^iso = 25.3, p < 0.01), with rural roads eliciting significantly longer latencies than the other two types (p < 0.01; all significant a posteriori pairwise comparisons are reported by using Newman-Keuls tests). There were no significant differences between the groups (F x 75 = 0.06), nor was there any significant interaction between groups and road environment (F 2150 = 1.47). Means and standard deviations for all button-press and eye-movement measures are

7 Danger and experience 957 shown in table 1. In this table data are aggregated across subject groups on the grounds that differences on this factor were generally not significant; where significant main effects or interactions involving the between subjects factor are found they are plotted separately. The second measure, the mean number of button responses per second, is available both within and outside danger windows and provides a test of the effectiveness of this manipulation. There was a significant main effect of road environment on this measure (7^ 150 = 14.28, p < 0.01), one of danger window (F hl5 = , p < 0.01), and an interaction between the two CF 2,150 = 6.43, p < 0.01). Here danger windows elicited more button presses than safe windows in all cases, and rural roads elicited fewer button presses than the other two road types in danger windows only (all comparisons/? < 0.01). There were no significant main effects of group, nor interactions involving this factor. Table 1. Summary data averaged across subject groups for three road types and in 'danger' and 'safe' windows. Standard deviations are given in parentheses below each value. Urban Suburban Rural danger safe danger safe danger safe Mean response latency/ms Mean number of responses per second Mean fixation duration/ms Mean saccade length/deg Horizontal gaze angle/ 0 Vertical gaze angle/ 0 Horizontal variance Vertical variance na = not applicable (447) 0.28 (0.09) 491 (107) 2.01 (0.45) 0.49 (0.89) (0.41) 5.30 (2.05) 0.25 (0.21) na 0.06 (O.H) 389 (70) 2.16 (0.31) 0.48 (0.56) (0.37) 6.96 (1.50) 0.48 (0.31) 1529 (633) 0.30 (0.08) 586 (178) 1.74 (0.48) (0.82) (0.33) 3.05 (1.73) 0.22 (0.30) na 0.05 (0.05) 420 (68) 1.99 (0.39) (0.69) (0.33) 5.64 (1.42) 0.33 (0.25) 2005 (969) 0.23 (0.11) 623 (211) 1.41 (0.32) 0.22 (0.80) (0.43) 3.13 (1.16) 0.16 (0.42) na 0.03 (0.04) 437 (79) 1.71 (0.32) 0.55 (0.50) (0.32) 5.18 (1.18) 0.44 (0.40) 3.2 Eye-movement measures Eye-movement data from dynamic scenes require special criteria to be used in the determination of fixations since virtually all objects on the screen are in motion for much of the time. For our analyses fixations were declared to be in progress when the point of gaze remained within an area of 0.25 deg x 0.25 deg for a period of at least 50 ms. This area was updated with each new data point so that eye movements with an average speed of less than 5 deg s _1 were aggregated into fixations. No distinction was made between fixations and pursuit tracking movements; all data sequences that fulfilled the above criteria were defined as fixations while data at all remaining times were assumed to be saccadic movements and rejected from further analysis. Six overall measures of eye-fixation patterns were then calculated, the mean fixation duration, the mean angular distance of saccades, the mean horizontal and vertical fixation locations, and the variance in horizontal and vertical fixation locations over the individual film. Mean fixation durations are plotted in figure 1. Analysis of mean fixation durations revealed a main effect of both road environment (F 2l50 = 26.16, p < 0.01) and danger window (F l75 = , p < 0.01) and an interaction between the two CF = 5.92, p < 0.01). Here danger windows attracted longer fixation durations than safe windows for all road types (p < 0.01). Urban roads produced shorter fixation durations than both

8 958 P R Chapman, G Underwood E 550 danger (YN) - -m- - danger (OE) D safe (YN) - -a- - safe (OE) S * urban.--d""" suburban Road environment rural Figure 1. Mean fixation durations for young novice (YN) and older experienced (OE) drivers in 'danger' and 'safe' windows during films showing three different road environments. other road types in danger windows (p < 0.01). Fixation durations were significantly shorter for suburban films than rural ones in danger windows only (p < 0.05), and in safe windows urban films elicited significantly shorter fixation durations than rural ones (p < 0.05). On this measure there was also a marginal difference between young novices and older experienced drivers (7^75 = 3.38, p = 0.07), with young novices generally having longer fixation durations. There was also an interaction between subject group and danger window (F ll5 = 4.27, p < 0.05). Here danger windows elicited longer fixation durations than safe windows for both subject groups (p < 0.01), but fixation durations were only significantly longer for young novices in danger windows (p < 0.05). No other interactions were significant. Mean saccade lengths were calculated as the angular distance between the start of each fixation and the end of the previous one. There were significant main effects of road environment (F 2m =99.65, p < 0.01) and of danger window (F l75 =68.89, p < 0.01), and a marginal interaction between the two (F 2150 = 3.02, p = 0.05). Here saccade lengths were shorter in danger windows than safe windows for all road types (p < 0.01), and for both types of window saccade lengths were longer for urban clips than suburban clips, where saccade lengths were in turn longer than those in rural clips (p < 0.01). There were no significant main effects or interactions involving the between subjects factor. Mean gaze locations and variances were calculated for the horizontal and vertical axes separately. However, to understand the meaning of these results it is necessary to appreciate how these locations actually relate to the stimuli that were presented. The screen was thus divided into 192 regions (16 horizontal by 12 vertical) corresponding to squares of just under 1 deg by 1 deg of visual angle. Fixation durations were then summed across subjects within each of these 192 regions to calculate a fixation density plot for each type of stimulus. Figures 2 to 4 show these fixation plots aggregated separately for dangerous and safe times within urban, suburban, and rural films. In these plots the centre of the screen is at 0 deg by 0 deg and is generally extremely close to the focus of expansion of the road scene ahead. Fixations closer to the front of the vehicle will have negative values for vertical location. Fixations towards the right (oncoming traffic on British roads) have positive values for horizontal location. Regions in the largest, outside band of these plots represent locations that received less than 0.3% of the total fixation time of subjects, while regions in the central band will be those that received approximately 10% of the total fixation time. Figure 2a thus shows two separate areas attracting large proportions of gaze time, one just below 0 deg by 0 deg which will generally be very close to the focus of expansion in these scenes, and a second region at approximately 2 deg by -1 deg which is the location of oncoming traffic.

9 Danger and experience 959 inininmininmin^ninininmininm I I I I I I I I ininininininminmininminininm i> vd in ^' en r4 ^H d d '-H H c«s -^ in vd t-^ I I I I I I I I ( a ) Horizontal degrees /u\ Horizontal degrees Figure 2. Fixation density plots for (a) 'danger' windows and (b) 'safe' windows within films showing urban roads. in»ninininininininininininininin inminmininininmininioinininin r^vdin^tro(n'-hdd'-hcnrn^t>nvdt^ h ^ t n x t m t N H o d H H r n ^ i o v d ^ I I I I I I I I I I I I I I I I (a) Horizontal degrees (b) Horizontal degrees Figure 3. Fixation density plots for (a) 'danger' windows and (b) 'safe' windows within films showing suburban roads. minininininininininininmini/")in I I I I I I I I Horizontal degrees inininminminininininininminin h ^ i n t ' r n t N ^ 6 d ^ ( N r n ' t i n \ d ^ I I I I I I I I Horizontal degrees (a) uv^a (b) Figure 4. Fixation density plots for (a) 'danger' windows and (b) 'safe' windows within films showing rural roads. Analysis of mean horizontal gaze location showed significant main effects of both road environment (F = 75.34, p < 0.01) and danger window^ 75 = 67.48,/? < 0.01), and there was an interaction between the two (F 2150 = 12.30, p < 0.01). These means are shown in table 1, but the nature of the interaction can be seen most clearly by

10 960 P R Chapman, G Underwood comparing figure 3a with figure 3b, and figure 4a with figure 4b. In these two cases fixations are reliably shifted to the left in the dangerous portions of the films (p < 0.01). For both types of window mean gaze location is significantly further to the left in suburban films than in the other two types (p < 0.01), while for danger windows mean fixations are significantly further to the left in rural films than in urban ones (/; < 0.05). There were no significant main effects or interactions involving the between-subjects factor. For mean variance in horizontal fixation locations there were significant main effects of both road environment CF 2,i5o 85.39, p < 0.01) and danger window (F ll5 = , p < 0.01), and there was again an interaction between the two, F 2l , p < This effect is also clear in figures 2 to 4, particularly in the lower horizontal spread in the fixation density plots for danger windows compared with their safe counterparts this difference in variance is significant for all three road types (p < 0.01). Variances are significantly higher for urban films than the other two types in both types of window (p < 0.01), and variances are significantly higher for suburban films than rural films in danger windows only (p < 0.05). Mean variances are given in table 1 it should be noted here that variance in fixation locations is calculated for each subject and film window individually and does not necessarily relate to the variance between subjects in mean location measures (given as standard deviations in table 1). Analysis of the mean vertical gaze locations also revealed a significant main effect of road environment (i^ 150 =97.89, p < 0.01) and of danger window (F lf75 = 8.91, p < 0.01), and an interaction between the two (i^150 = 14.83, p < 0.01). In this complex interaction all three road types are significantly different from one another in danger windows (p < 0.01), and suburban roads are significantly different from the other two types in safe windows (p < 0.01). The difference between danger and safe windows is significant for each road type (p < 0.05), although on urban roads the mean gaze location is significantly lower in danger windows than safe ones, while the effect is the opposite way around for the other two road types. There was also a main effect of subject group (i^75 = 4.02, p < 0.05), with older experienced drivers generally looking lower down (closer to the front of the vehicle from which the films were taken) than young novices. These effects are plotted in figure 5. The analysis of mean vertical gaze variance also showed a significant main effect of road environment CF = 4.04, p < 0.05), and of danger window (F ll5 = 33.44, p < 0.01), and an interaction between the two (F = 3.45, p < 0.05). Here danger windows are characterised by lower variance in vertical gaze angle for both urban and rural films (p < 0.01). Suburban films additionally show lower variance than the other two types in safe windows only (p < 0.05). There was also a main effect of 0 T urban suburban rural Road environment a danger (YN) - -m- - danger (OE) -o safe (YN) - -n- - safe (OE) Figure 5. Mean vertical gaze locations for young novice (YN) and older experienced (OE) drivers in 'danger' and 'safe' windows during films showing three different road environments.

11 Danger and experience T m danger (YN) danger (OE) 0.54 safe (YN) - -a- - safe (OE) 0.4} o 0.3 > urban suburban rural Road environment Figure 6. Mean vertical gaze location variance for young novice (YN) and older experienced (OE) drivers in 'danger' and 'safe' windows during films showing three different road environments. subject group (F ] , p < 0.01), with young novices generally showing greater variance in vertical fixation locations than older experienced drivers. These effects are plotted in figure 6. 4 Discussion One surprising aspect of these results was the lack of a latency difference in button responses between the young novice and older experienced groups. No differences between the groups were found on this measure, nor were there any main effects or interactions involving subject group on the frequency of button pressing. We had anticipated that more experienced drivers might respond faster to the dangerous events or report identifying more such events in the films. However, it is possible that the events depicted in these films were sufficiently well defined to be easily detected by all subjects and that any advantage from increased traffic experience in our older experienced group was compensated for by slowing in reaction time with their greater age compared with the young novice group. Effects of danger window were observed on all of the measures recorded. The fact that subjects pressed the response button much more frequently in such windows and very rarely outside them provides a validation of our definition of the danger windows. The fact that so many eye-movement measures differed as a function of the window has more important implications. In summary, encountering a dangerous event had the same general effects on eye movements for both groups of drivers: fixation durations increased, mean saccade angular distances decreased, and the spread of visual search (measured by variance in fixation locations) decreased in both the horizontal and the vertical dimension. These effects were very strong and appeared for all three types of road environment. These results are much as would have been predicted from theories of attention focusing from the field of eyewitness testimony. There were also differences in the mean fixation locations in both the horizontal and the vertical axis; however, these effects interacted strongly with the type of road environment and their directions and magnitudes will probably depend on the precise locations of particular objects or people visible in the scenes being viewed. Differences in measures as a function of road environment highlight the importance of finding ways to categorise driving stimuli rather than treating all driving situations as comparable. Fixation durations were greatest on the rural roads and shortest for the urban ones. In contrast, saccade angular distances were greatest on the urban roads and shortest on the rural ones. This implies a pattern of short, widely spaced fixations on the urban roads and long, focused fixations for the rural ones. This interpretation is supported by the differences in horizontal and vertical variance between the stimuli. The finding that fixation durations increase with danger but decrease with

12 962 P R Chapman, G Underwood visual complexity is an important and initially counterintuitive one. Measures of both danger and complexity might be expected to be positively correlated with concepts such as mental workload. Thus it is important to note that indices of mental workload may be associated with short fixation durations in some circumstances but longer durations in others. This has clear practical implications for studies involving such measures. Two types of group difference were observed between our drivers, those relating to fixation durations and those relating to vertical search. The differences in fixation duration are in the direction predicted and in accord with the existing literature. Young novice drivers generally had longer fixation durations than older experienced drivers this is what would be expected if fixation durations represent the difficulty in processing the visual region and experience allows faster processing. The fact that the difference is particularly marked in dangerous situations implies that the young novices may be specifically lacking in relevant experience of such situations. The group differences in vertical search are also striking and not what would have been predicted from the literature. Our young novice drivers showed greater variance in vertical fixation locations, and generally fixated further ahead of their own vehicle (higher up the screen). The greater variance in vertical fixation location might have been predicted by the need for novice drivers to fixate the edge of the road near the vehicle to maintain lane position (Land and Horwood 1995). However, subjects in our task had no need to engage in steering so this would have to be a carryover effect from their normal driving strategies. More importantly, this analysis would predict a lower mean fixation location for the young novices than the older experienced drivers we observed just the opposite result. A possible cause for the surprising vertical-search behaviour of our young novice drivers could be that having just completed their driver training they are still actively following the advice of their instructors to look as far ahead as possible to detect upcoming hazards (eg Miller and Stacey 1995; Zwahlen 1993). Such behaviour, combined with pursuit movements required to identify and evaluate objects approaching them, would be consistent with the higher mean fixation locations, greater variance in vertical gaze location, and longer fixation durations (which will include pursuit tracking). If this were the case we might predict that such behaviour, which was not observed in the older experienced drivers, might rapidly decay as newly qualified drivers develop their own strategies of visual search as a function of increased traffic experience. However, it should be noted that, although significant, the magnitude of these differences are relatively small in practical terms. The evidence for a group difference between novice and experienced drivers on mean duration of fixations supports the tentative findings reported by a number of authors in different contexts. Mourant and Rockwell (1972) thus reported a decrease in the frequency of pursuit movements as a function of experience. We have classified these as fixations and in Mourant and Rockwell's analysis they represented relatively long fixation durations (over 400 ms). Cohen and Studach (1977) similarly found that experienced drivers had significantly shorter fixation durations than novices, but only for curves to the right. Miltenburg and Kuiken (1990) predicted that experienced drivers would have shorter fixation durations than novices, but only found the effect unambiguously for one of their six films. The findings from our study suggest that this is a general pattern of behaviour that can be observed when fixation data are averaged over most sufficiently long events. Miltenburg and Kuiken's interpretation of the decreased fixation durations was that experienced drivers already have relevant schemata to deal with the situation and hence have to spend less time abstracting information from the scene. This approach is very much in line with the work of Theeuwes (eg 1996) and seems to provide a good explanation for the general effect. However, there are clear limitations to this as an explanation for changes in fixation durations in individual subjects, both

13 Danger and experience 963 within an individual film and particularly across films in different traffic environments. For our task the rural, suburban, urban distinction is particularly clear in that it is related directly to mean fixation durations. For both novice and experienced drivers fixation durations are longest on rural roads and shortest on urban ones. This is consistent with the results reported in actual driving by Underwood et al (1997) and with the general finding that increasing the complexity of the visual scene increases the number of eye movements made and decreases the mean fixation durations on individual objects (Erikson and Horberg 1980; Luoma 1986; Miura 1990; Robinson et al 1972; Rutley and Mace 1968). Differences in fixation durations between subjects may be relatively small in comparison with the large differences between driving environment and from moment to moment as a function of perceived danger. Acknowledgements. This research was supported by a grant from the Department of the Environment, Transport and the Regions (DETR) in Great Britain. Any views expressed in this paper are those of the authors and not necessarily those of DETR. The authors wish to thank Dr Steve Marchant for technical assistance, Sharon Wright, Sara Batts, and David Crundall for assistance in data collection, and the Driving Standards Agency for assistance in recruiting newly qualified drivers. References Antes J R, 1974 "The time course of picture viewing" Journal of Experimental Psychology Antes J R, Penland J G, 1981 "Picture context effects on eye movement patterns", in Eye Movements: Cognition and Visual Perception Eds D F Fisher, R A Monty, J W Senders (Hillsdale, NJ: Lawrence Erlbaum Associates ) pp Buswell G T, 1935 How People Look at Pictures (Chicago, IL: University of Chicago Press) Christianson S, 1992 "Emotional stress and eyewitness memory: A critical review" Psychological Bulletin Christianson S, Loftus E F, Hoffman H, Loftus G R, 1991 "Eye fixations and memory for emotional events" Journal of Experimental Psychology: Learning, Memory, and Cognition Cohen A S, 1981 "Car drivers' pattern of eye fixations on the road and in the laboratory" Perceptual and Motor Skills Cohen A S, Studach H, 1977 "Eye movements while driving cars around curves" Perceptual and Motor Skills Easterbrook J A, 1959 "The effect of emotion on cue utilization and the organization of behavior" Psychological Review Elander J, West R, French D, 1993 "Behavioral correlates of individual differences in road-traffic crash risk: An examination of methods and findings" Psychological Bulletin Erikson B, Horberg U, 1980 "Eye movements of drivers in urban traffic" Uppsala psychological reports 283, University of Uppsala, Sweden Evans L, 1991 Traffic Safety and the Driver (New York: Van Nostrand Reinhold) Forsyth E, Maycock G, Sexton B, 1995 "Cohort study of learner and novice drivers: Part 3, Accidents, offences and driving experience in the first three years of driving" Department of Transport TRL project report 111, Transport Research Laboratory, Crowthorne, Berks Friedman A, 1979 "Framing pictures: The role of knowledge in automatized encoding and memory for gist" Journal of Experimental Psychology: General Helander M, Soderberg S, 1972 "Driver visual behavior and electrodermal response during highway driving" Goteborg Psychological Reports 2(4) (whole issue) Henderson J M, 1992 "Visual attention and eye movement control during reading and picture viewing", in Eye Movements and Visual Cognition: Scene Perception and Reading Ed. K Rayner (New York: Springer) pp Henderson J M, McClure K K, Pierce S, Schrock G, 1997 "Object identification without foveal vision: Evidence from an artificial scotoma paradigm" Perception & Psychophysics Henderson J M, Hollingworth A, 1998 "Eye movements during scene viewing: An overview", in Eye Guidance in Reading and Scene Perception Ed. G Underwood (London: Elsevier) pp Hughes P K, Cole B L, 1986a "What attracts attention when driving?" Ergonomics Hughes P K, Cole B L, 1986b "Can the conspicuity of objects be predicted from laboratory experiments?" Ergonomics Kramer T H, Buckhout R, Eugenio P, 1990 "Weapon focus, arousal, and eyewitness testimony: Attention must be paid" Law and Human Behavior Land M F, Horwood J, 1995 "Which parts of the road guide steering?" Nature (London)

14 964 P R Chapman, G Underwood Land M F, Lee D N, 1994 "Where we look when we steer" Nature (London) Lester J, 1991 "Individual differences in accident liability: A review of the literature" TRRL RR 306, Transport and Road Research Laboratory, Crowthorne, Berks Loftus E F, 1979 Eyewitness Testimony (Cambridge, MA: Harvard University Press) Loftus E F, Loftus G R, Messo J, 1987 "Some facts about weapon focus" Law and Human Behavior Loftus G R, Mackworth N H, 1978 "Cognitive determinants of fixation location during picture viewing" Journal of Experimental Psychology: Human Perception and Performance Luoma J, 1986 "The acquisition of visual information by the driver: Interaction of relevant and irrelevant information" reports from Liikenneturva 32/1986, Central Organization for Traffic Safety, Helsinki, Finland McKenna F P, Crick J L, 1994 "Hazard perception in drivers: A methodology for testing and training" TRL contractor report 313, Transport Research Laboratory, Crowthorne, Berks Mackworth N H, Morandi A J, 1967 "The gaze selects informative details within pictures" Perception & Psychophysics McLean J R, Hoffmann E R, 1971 "Analysis of drivers' control movements" Human Factors McLean J R, Hoffmann E R, 1973 "The effects of restricted preview on driver steering control and performance" Human Factors Miller J, Stacey M, 1995 The Driving Instructor's Handbook (London: Kogan Page) Miltenburg PGM, Kuiken M J, 1990 The Effect of Driving Experience on Visual Search Strategies: Results of a Laboratory Experiment (Haren, The Netherlands: Traffic Research Centre, University of Groningen) Miura T, 1990 "Active function of eye movement and useful field of view in a realistic setting", in From Eye to Mind: Information Acquisition in Perception, Search and Reading Eds R Groner, G d'ydewalle, R Parham (Amsterdam: Elsevier) pp Mourant R R, Rockwell T H, 1970 "Mapping eye-movement patterns to the visual scene in driving: An exploratory study" Human Factors Mourant R R, Rockwell T H, 1972 "Strategies of visual search by novice and experienced drivers" Human Factors Pelz D C, Krupat E, 1974 "Caution profile and driving record of undergraduate males" Accident Analysis and Prevention Quimby A R, Maycock G, Carter I D, Dixon R, Wall J G, 1986 "Perceptual abilities of accident involved drivers" TRRL RR 27, Transport and Road Research Laboratory, Crowthorne, Berks Quimby A R, Watts G R, 1981 "Human factors and driving performance" TRRL supplementary report 718, Transport and Road Research Laboratory, Crowthorne, Berks Riemersma J B J, 1988 "An empirical study of subjective road categorization" Ergonomics Robinson G H, Erickson D J, Thurston G L, Clark R L, 1972 "Visual search by automobile drivers" Human Factors Rutley K S, Mace D G W, 1968 "A preliminary investigation into the frequency of driver motor actions and eye movements" RRL report LR 162, Road Research Laboratory, Crowthorne, Berks Sabey B E, Taylor H, 1980 "The known risks we run: The highway" TRRL supplementary report 567, Transport and Road Research Laboratory, Crowthorne, Berks Shinar D, McDowell E D, Rockwell T H, 1977 "Eye movements in curve negotiation" Human Factors Staplin L, 1995 "Simulator and field measures of driver age differences in left-turn gap judgments" Transportation Research Record Theeuwes J, 1996 "Visual search at intersections: An eye-movement analysis", in Vision in Vehicles volume 5 Eds A G Gale, I D Brown, C M Haslegrave, S P Taylor (Amsterdam: Elsevier) pp Theeuwes J, Hagenzieker M P, 1993 "Visual search of traffic scenes: On the effect of location expectations", in Vision in Vehicles volume 4 Eds A G Gale, I D Brown, C M Haslegrave, H W Kruysse, S P Taylor (Amsterdam: Elsevier) pp Underwood G, Crundall D E, Chapman P R, 1997 "Visual attention while performing driving and driving-related tasks", in Behavioural Research in Road Safety volume 7 Ed. G B Grayson (Crowthorne, Berks: Transport Research Laboratory) pp Yarbus A L, 1967 Eye Movements and Vision (New York: Plenum) Zwahlen H T, 1993 "Eye scanning rules for drivers: How do they compare with actual observed eye scanning behavior?" Transportation Research Record p J ^ 1998 a Pion publication printed in Great Britain