- Original Paper
- Open Access
Driver behaviour in unexpected critical events and in repeated exposures – a comparison
European Transport Research Review volume 6, pages51–60 (2014)
This paper aims to determine how truck driver steering behaviour seen in repeated exposures to acritical event correlates to the behaviour resulting from an unexpected exposure to the same event.
Test subjects were exposed to an unexpected critical event in a high-fidelity driving simulator. Next, a slightly modified version of the scenario was repeated several times for each subject. The driver behaviour was then analysed using standard statistical tests.
It was found that, in general, drivers keep most of their steering behaviour characteristics between test settings (unexpected and repeated). This is particularly interesting sincea similar kind of behaviour preservation is generally not found in the case of braking behaviour. In fact, onlyone significant difference was found between the two test settings, namely regarding time-to-collision at steering initiation.
In experiments involving both an unexpected event and several repeated events one can,at least in some cases, design the repeated event such that behavioural data collected from that setting can beused along with data from the unexpected setting. Using this procedure, one can significantly increase the amount of collected data, something that can strongly benefit, for example, driver modelling.
Road traffic accidents constitute a large problem on a global scale. Apart from great economical andsocial costs, the accidents also cause a significant number of injuries and deaths. The number of worldwide fatalities have recently been estimated at over one million per year , of which 8 % occur in Europe. Driver behaviour is widely considered as a contributing factor in many road traffic accidents. Therefore, research efforts regarding the safety aspects of such behaviour have been intensified during recent years .
As a way to understand and further study the impacts of driver behaviour, models that capture aspects of driverbehaviour are being developed (for a recent review, see ). An important goal of driver behaviour research is to find simple phenomenological relations that explain driver behaviour in certain situations. Examples of such behaviour could, for instance, be braking behaviour as a function of headway to an obstacle , or steering behaviour as a function of perceptual inputs .
Data used in the development of a driver model are typically collected from real-world driving experiments or driving simulator studies. In most experimental arrangements, the intention is to mimic realistic scenarios in order to measurerealistic behavioural responses in drivers. However, fully realistic scenarios cannot easily be accomplishedin a non-naturalistic driving experiment. In particular, in order to obtain sufficient amounts of data forstatistical analysis and development of driver models, it might be necessary to expose each driver to thestudied scenario multiple times. This is so since, for economical and other reasons, one often cannot involve avery large number of drivers in an experiment. Furthermore, for driver model development, in view of theindividual differences between drivers, one generally needs quite many data points from each driver in orderfor meaningful model development to be possible. Before making use of the results obtained from such experiments one must therefore first understand the effects (if any) of repeated exposures to the scenario under study.
Previously, several studies have considered driver reaction times in braking scenarios . It has been shown that,when reacting to expected (e.g. repeated) events, drivers’ brake reaction times are reduced significantly [2, 10]. Furthermore, it has been concluded that test subjects brake earlier and more strongly in repeated events . However, it has also been shown that some behavioural aspects might be preserved in repeatedexposures . In this paper, the effects on driver steering behaviour caused by repetition of a critical event are studied.
The data used for the analysis presented in this paper were collected in a truck simulator study regarding driverbehaviour in connection with an electronic stability control (ESC) system. In the experiment, the truck drivers were askedto drive on a road in a winter environment, where they were exposed to an unexpected critical lead vehicle braking scenario, inducing a rapid steering manoeuvre. In a brief pre-study, the unexpected scenario was not found to be reliable ininducing sufficient ESC-relevant (loss of control) data, and the scenario was therefore repeated several times for each driver.
The analysis presented here will be centred on the effects of steering behaviour caused by the repeated exposures to the scenario rather than the effects on driver behaviour caused by the ESC system. Furthermore, only events in which thedriver successfully evaded a collision with the lead vehicle will be considered. Events involving a collision should be studied separately since the typical driver behaviour appears to be fundamentally different, in particular for the unexpected case.
In order to ascertain the validity of steering data collected from repeated exposures, a statistical comparison between aspects of the unexpected and the repeated data has been carried out. Four different criteria (C1-C4) for measuring the validity of repeated exposure data have been defined as:
Scenario tuning: Can the repeated scenario be tuned so that the steering manoeuvre is initiated under similar conditions in both test settings (unexpected and repeated)? If so, when the truck driver initiates the evasive steering, there are no crucial differences in road position, speed, or headway to the lead vehicle, regardless of the test setting (unexpected or repeated).
Manoeuvre similarity: Is the steering manoeuvre carried out in a similar fashion in both settings, that is without crucial differences in maximum steering wheel angles, steering wheel rates, or steering wheel reversal ratesFootnote 1?
Preservation of individual behaviour: Do the test subjects keep their individual steering behaviour characteristics between the two test settings?
Effects of learning: Are there no crucial effects of learning on how steering avoidance is carried out, meaning that test subjects do not change their steering performance over repetitions?
Data were collected in a high-fidelity moving-base truck simulator at the Swedish National Road and Transport Research Institute (VTI). The simulator uses a moving base platform providing lateral movement as well as roll and pitch rotations. A visual system provides a 105° field of view using forward-facing projectors, and emulated rear mirrors using monitor screens. For the studyconsidered in this paper, a six-wheeled rigid truck with a wheel base of 6.2 m (from first to last axle) wassimulated.
The main scenario used in the experiment was originally proposed in  and involves acritical lead vehicle braking scenario, as illustrated in Fig. 1. The scenario took place on adivided four-lane motorway with two lanes in each direction. The road speed limit for trucks was 80 km/h. While the testsubject was driving at speed in the right lane, a passenger car, referred to as the principal other vehicle (POV), overtook the truck at speed using the left lane, where is a constant. At a time headway with respect to the truck,the POV (at lateral speed )changed from the left to the right lane. Once in the right lane, the POV continued forward, still at speed . Then, without any apparent reason,at time headway , the POV brakedstrongly with the constant deceleration .The deceleration continued until the POV stopped completely (with one exception, describedin the next paragraph). Right before the deceleration, the speed was instantaneously set to in order to ensure that the time headway was always less than, or equal to,.When the critical scenario was initiated, the road friction was lowered from to in order to emulate a slippery road surface.
As seen in Table 1, the scenario was parametrized for three different versions: (i) unexpected avoidance (UA), (ii) repeated avoidance (RA), and (iii) catch trial (CT), i.e. a version inwhich no steering avoidance was needed. In the CT, the POV deceleration ended at speed and was then followed by an acceleration .The CT was included so that the drivers would not be certain regarding the nature of a repeated event, i.e whether it wasan RA scenario or a CT scenario.
The parameters for the RA scenario were chosen so that the test subjects would initiate their steering approximately atthe same time in the repeated events as they did in the unexpected scenario. In order to meet this requirement, a larger deceleration d b was used in the repeated events and in the catch trials so as to compensate for reduced reaction times , and the speed was used so as to control the time at which the test subjects initiated evasive steering. Appropriate parameter values were chosenbased on results from (i) a brief pre-study in a fixed base simulator, and (ii) computer simulations using a simple drivermodel. Both the UA and RA events were tuned such that evasive braking alone was insufficient to avoid a collision.
The experiment consisted of three parts with a preceding training session. The last part, involving a double lane changeon a cone track, will not be discussed in this paper. The training session, inspired by  and , included driving insteady-state traffic as well as a few repetitions of non-critical braking and steering exercises. The total length of thetraining session was about ten minutes and it was carried out on the same simulated road as the critical scenario (described above).
In the first part of the experiment, the test subjects were instructed to drive normally at 80 km/h on the motorwaydescribed above. After approximately four minutes of driving, including four overtaking vehicles and a non-critical doublelane change induced by a road construction site, the UA scenario occurred. Unknown to the test subjects, half of them hadthe ESC system present in their trucks, and half did not.
In the second part of the experiment, the subjects were informed about the ESC system and whether or not it waspresent in their vehicle (explained as an “anti-skid system”). They were also instructed about the RA and CTscenarios (described above), and that both scenarios would be repeated several times in a random order. Thedrivers were then asked to drive at 80 km/h on the same motorway as in the previous part, and to applyevasive steering only when they considered it required in order to avoid a collision (i.e. only in the RA event).
During the second part, each subject experienced, in random order, four overtaking vehicles, six occurrences of the RA scenario, and eight occurrences of the CT scenario. Note that, in no case did the CT scenario evoke a critical steeringavoidance behaviour, and the analysis below thus concerns only the RA scenario in relation to the UA scenario.
In total, 24 test subjects participated in the experiment. The drivers were divided into two equally large groups basedon driving experience (low or high). The drivers belonging to the low-experience group were mainly recruited from a localdriving school, where they soon would get their licence for truck-trailer combinations. However, all low-experience driversalready had a driver’s licence for the rigid truck simulated in the experiment. The drivers belonging to the high-experiencegroup were mainly recruited from local hauler companies.
The steering manoeuvre was divided into four time intervals, as illustrated in Fig. 2. The first interval, , starts at time when the overtaking car initiatesbraking, and ends at time when the truck driver initiates evasive braking. The second interval,, starts at time, and ends at time when the truck driver initiatesevasive steering. The interval starts at time ,and ends at time when the truck driver has reached the maximum steering wheel angle to the left. Finally, the interval starts attime , andends at time when the truck driver reaches the largest steering wheel angle to the right. Time isgiven as the first (in time) local minimum at which the steering wheel angle signal reached a value less than-20°, time is given as the time of the largeststeering wheel angle between and , andtime is given as the time of the first steering wheel angle value below 8° when tracking the signal backwards from . Note that, for theschematic illustration in Fig. 2, could have be defined as the global minimum of the steering wheel angle. However, some steering manoeuvres involved acontrol loss, in which case subsequent minima of the steering wheel angle could be deeper than the first minimum, hence the definition of given above. An actual steering response is presented in Fig. 3.
The independent variables considered in this paper are: test setting (UA or RA), and repetition (one throughsix).
A number of dependent variables, presented in Table 2, were extracted from each event in the driving data. Time to collision (TTC) is measured as the time it would take for the truck to cover the distance between the front of the truck and the rear of the POV (assuming current POV and truck speeds, i.e. disregarding any accelerations). The steering wheel reversal rate (SWRR), was defined as the number of steering wheel reversals per minute, larger than a certain minimum angular value .
Two types of tests were used when testing for significant differences in dependent variables between testsettings (i.e. criteria C1 and C2). If the samples of a dependent variable were found to be normally distributed adependent t-test for paired samples was used, otherwise a Wilcoxon signed-rank test was used. Normality was tested usingthe Shapiro-Wilk test, as reported in Table 2. For each test subject and dependent variable, the unexpected event as wellas the mean of all repeated events are used as one pair of values. Furthermore, for each significance test, the Pearson’s correlation coefficientr was calculated as a measure of effect size. The use of the paired samplest-test or the Wilcoxon test implies that no assumption regarding the homogeneity of variances needs to be made .
Pearson’s correlation coefficient is also used for two other purposes, first to determine the paired sample correlation of adependent variable between test settings, in order to see whether driver behaviour is transferred between settings (C3) and second to determine the correlation of a dependent variable between repetitions, in order to study learning effects (C4).
Three of the four criteria involve an absence of a crucial difference in behaviour. Here, a crucial difference is defined as a medium effect size (|r| > 0.3) .
According to the experimental set-up, the maximum amount of data was 24 repetitions of the unexpected event, and144 repetitions of the repeated event. Data from one of the repeated events were lost due to experimental problems. Ineight unexpected and two repeated events, the test subjects did not attempt any evasive steering at all. A successful evasive manoeuvre (i.e. without collision) occurred in 9 of the unexpected events and 88 of the repeated events. As statedin the end of Section 1, all events where a collision occurred were discarded. Four other repeated events were discardedsince the test subject violated the instructions by initiating steering directly when the POV began braking. Finally, data from one unexpected event were discarded since the driver steered much earlier compared to theother test subjects (with a TTC of 7.9 s). In Table 3, Repeated A refers to all repetitions matching the corresponding row criterion, while Repeated B only includes matching repetitions that also share a test subjectrepresented in the left column. In Table 4, the 33 filtered data points in Repeated B are matched to their corresponding test subject. Figure 4 summarizes the TTC at steering initiation for the retained data points after filtering.
In the rest of this section, the steering manoeuvre will be analysed by comparing the dependent variable pairs extracted from the unexpected events with the variables from the repeated events. In order to make it possible to use a paired samplet-test, all repeated events from the same test subjects were merged by calculating the mean of each dependent variable. Therefore, 2 × 8 values will be compared in each t-test.
In Figs. 5–9, the left panel compares the unexpected and averaged repeated scenarios, whereas the right panel comparesthe repetitions over repetition number. The left panel includes the eight sample pairs used in the t-tests, and the rightpanel includes the 84 repetitions presented in the middle column of Table 3. Error bars show 95 % confidence intervals,calculated under the assumption of a normal sampling distribution.
Situation at steering initiation
When analysing the situation at steering initiation (t2)between test settings, a significant difference was found in one out of three dependent variables: By using the t-test, the significant difference was found in TTC (see Fig. 5) between UA (M = 2.86, SD = 0 .74) and RA (M = 3.31,);,,. Nosignificant difference was found in lateral position (see Fig. 6) or longitudinal speed.
The steering manoeuvre was analysed by comparing six of the presented (see Table 2) dependent variablesbetween the UA and the RA. The six variables include: (i) maximum steering wheel angle to the left (), (ii) maximum steering wheel angle to the right (), (iii) maximum steering wheel rate while turning left (), (iv) maximum steering wheel rate while turning right (),(v) SWRR (5°) whileturning left (), and(vi) SWRR (5°) whileturning right ().No significant differences were found between the two test settings (UA and RA) for anyof the variables. However, one of the variables, namely steering wheel rate in the interval (see Fig. 8) gave aconsiderably larger effect size (,) compared to theother five variables (,|r| ≤ 0.30). Two of the other variables,steering wheel angle at time (,) andSWRR (5°) inthe interval (,), areexemplified in Figs. 7 and 9 respectively.
Preservation of steering behaviour
By calculating a paired sample correlation, using the same 2 × 8 values asabove, one can study the preservation of steering behaviour between both test settings. When testing all the nine dependent variables listed in Table 2, it was found that four of them had a significant correlation, namely steering wheel angle attimes (,) and(,), and steering wheelrate in the intervals (,) and(,). No other variable correlation showed significance, nor any high values of the Pearson’s correlation coefficient r. In Fig. 10,two examples of paired sample correlation are presented, one with a large r value, and one with a small r value.
Learning effects in steering behaviour
For the analysis of potential learning effects, the data collected in the unexpected scenario were not used. Therefore,since the use of paired data was not required, the 84 data points presented in the middle column of Table 3 could be used.In this case, a strict test of significance would be complicated (but not impossible) since each driver is only represented ina subset of the repetitions, with different subsets for different drivers. Therefore, only the Pearson’s correlation coefficient was determined for each dependent variable, and for which a large (absolute) r value would indicate a learning effect.
By calculating Pearson’s correlation coefficient for each dependent variable, it was found that no largeabsolute r values existed. For three of the values, the modulus exceeded 0.1, namely SWRRs in the intervals(; see Fig. 9)and (), and steeringwheel angle at time (; see Fig. 7).
The discussion is structured as follows: First the findings in Section 3 will be related to the criteria listed at the end of Section 1. Then, aspects of the experimental set-up will be discussed, focusing on repeated critical events for collectingsteering behaviour data. Finally, the general applicability of such data will be discussed in the context of drivermodelling.
The validity of the first criterion (C1) regarding scenario tuning, cannot be entirely confirmed: When drivers initiated evasive steering, (attime ), there was a significantdifference in TTC (,), suchthat initiation occurred, on average, 0.5 s earlier in the repeated events, resulting in a slightly less critical situation.However, even though the difference is significant and with a large effect size, it might be argued that the difference is sufficiently low for C1 to be fulfilled, since drivers appear to behave similarly during the remainder of the UA and RA scenarios (as discussed below).
At the time of steering initiation, there was no significant difference in lateral position or speed between test settings.For lateral position, one could intuitively assume that test subjects would position their trucks further to the left in arepeated event. No such tendencies were seen, however.
The validity of criterion C2, regarding manoeuvre similarity, can be confirmed, as there were nosignificant differences between any analysed steering wheel angles, steering wheel rates, or SWRRs.There is, however, an apparent difference (not significant) regarding steering wheel angle at time (see Fig. 7) in the very first repetition. The same phenomenon can be seen in the steering wheel rate during the interval (see Fig. 8). A possible explanation could be that test subjects overestimate the severity of the situation in the first repetition,and then regain a behaviour similar to the one seen in the unexpected event.
Also criterion C3, regarding preservation of individual behaviour appears to be valid. Using paired sample correlation, itwas shown that test subjects keep their steering characteristics very well (r values close to 1) regarding maximum steering wheel angles and steering wheel rates. They do not, however, preserve their characteristics regarding SWRRs. Arguably, thecriterion can still be considered to be valid, since SWRRs are less important compared to maximum steering wheel angles and steering wheel rates when considering the outcome of the steering behaviour, i.e. avoiding the collision. It should alsobe pointed out that there were, as stated for the previous criterion, no significant differences in SWRRs between testsettings on the population level.
Interestingly, from Fig. 10 one can see that drivers tend to have a lower SWRR value (1 Hz) in the repeated eventscompared to the unexpected events. A possible explanation, if one considers experience to be scenario-specific, could berelated to the fact that inexperienced drivers (in this case meaning drivers that have not yet been exposed to the criticalevent) in general have a larger SWRR compared to those drivers who have experienced the critical event . In order toexplain the observed phenomenon in terms of control, one can hypothesize that an experienced (in the scenario) drivercarries out the steering manoeuvre in an open-loop manner, rather than in a closed-loop manner . However, asmentioned above, even though a difference in SWRR between test settings seems probable, no statistically significant difference was found.
Finally, the validity of criterion C4, regarding (the absence of) learning effects can be confirmed: There were no crucial() effectsof learning for any of the dependent variables. There were some small effects of learning for the SWRRs (, see Fig. 9, and), and the maximum steeringwheel angle to the right ().However, such small effects would hardly have any bearing here, considering the small number of repetitions.
It should also be pointed out that there were no large effects of learning on lateral position (). Again,one would intuitively assume that drivers should tend to gradually keep further to the left over repetitions. However, nosuch behaviour was seen.
In general, and especially when considering critical situations, the amount of available data is limited. Furthermore,when collecting data from critical situations, there are problems involving safety, cost, and test subjectexpectancy.
By using multiple repetitions of a critical event, the number of acquired data points can be stronglyincreased. Furthermore, by using repeated events, one can reduce costs by decreasing the total numberof test subjects as well as increasing the efficiency (event rate) during the experiment. Safety problemscan also be addressed by using repetitions since it allows for test subjects to be well instructed during theexperiment.
However, repetitions typically require some additional work. First, the repeated scenario must be designed tocompensate for known effects on driver behaviour. For instance, it was concluded in  that driver braking behaviour islargely affected by repetitions, in particular by decreased reaction times and stronger braking. In the case considered here,a significant difference between test settings was found in the TTC at steering initiation. In retrospect, this differencecould perhaps have been eliminated by further tuning the parameters for the RA scenario. Secondly, the datacollected from repeated scenarios must be validated for the application at hand. For this purpose, it is highlyrecommended first to collect data from an unexpected version of the scenario under study, as was done in theexperiment considered here, and then to make a comparative analysis between the data collected in theunexpected scenario and the repeated tests. Another reason for including unexpected events is that thebehaviour observed in repeated events is probably only a subset of all possible types of behaviour for the givenscenario.
Applicability in driver modelling
Since all four criteria (C1-C4) were found to be valid, possibly with the partial exception of C1, it hasbeen shown that the steering behaviour collected in repeated events can be used in driver modelling, atleast for the case of successful evasive manoeuvres. As seen in Table 3, only data from 8 out of 24 driverswere used in the comparison. From the remaining 16 drivers: (i) eight did not attempt steering at all, (ii)seven applied steering but still collided with the POV, and (iii) one steered much earlier compared to theother drivers. By using the data analysed in this paper, one can thus only cover a third of the observedbehaviour. All four types of behaviour must be covered when generating a general driver model for thecritical scenario under study. However, it should be pointed out that it is likely to be much easier to modelthose drivers that do not apply any evasive steering (i.e. another third of the drivers considered in thisexperiment).
In this paper, it has been shown that the steering behaviour observed in repeated critical scenarios to a large extentpreserves the characteristics of the steering behaviour found in an unexpected critical scenario, making it possible to usedata from repeated scenarios in, for example, driver modelling for evasive steering.
It has been demonstrated that a repeated scenario can be tuned so that evasive steering is initiated under the sameconditions as in an unexpected scenario. Furthermore, it was found that there were no significant differencesin lateral position and speed between test settings. There was, however, a significant difference in TTC(,), but it was argued that the difference did not seem to change the rest of the observed steering behaviour. Nosignificant differences were found in the maximum steering wheel angles or the steering wheel rates during the manoeuvre itself.
It was also found that test subjects keep their steering characteristics between test settings. Very high correlation values (r close to 1) were obtained from paired sample correlation tests of maximum steering wheel angles and maximum steeringwheel rates. Furthermore, no large effects of learning (measured using Pearson’s correlation coefficient) were found in thecollision avoidance manoeuvre.
Using repeated events in an experiment is recommended when large amounts of data are needed, for example whendeveloping driver models. However, it is also recommended to combine the repeated events with an unexpected event inorder to make it possible to validate the data collected in the repeated events.
Steering wheel reversal rates are explained in Section 2.2.2.
Cohen J (1988) Statistical power analysis for the behavioral sciences. Lawrence Erlbaum
Engström J, Aust M, Viström M (2010) Effects of working memory load and repeated scenario exposure on emergency braking performance. Hum Factors: J Hum Factors Ergon Soc 52(5):551–559
Erséus A (2010) Driver-vehicle interaction: identification, characterization and modelling of path tracking skill. PhD thesis, KTH
Field A (2009) Discovering statistics using SPSS. SAGE publications Ltd
Green M (2000) How long does it take to stop? Methodological analysis of driver perception-brake times. Transp Hum Factors 2(3):195–216
Hollnagel E, Woods DD (2005) Joint cognitive systems foundations of cognitive systems engineering. CRC Press / Taylor & Francis
Jamson A, Smith P (2003) Are you used to it yet? Braking performance and adaptation in a fixed-base driving simulator. In: Proceedings of the driving simulator conference North America, Dearborn, Detroit, Michigan. 8th-10th Oct 2003
Jurecki R, Stańczyk T (2009) Driver model for the analysis of pre-accident situations. Veh Syst Dyn 47(5):589– 612
Lee J (2008) Fifty years of driving safety research. Hum Factors: J Hum Factors Ergon Soc 50(3):521–528
Lee J, McGehee D, Brown T, Reyes M (2002) Collision warning timing, driver distraction, and driver response to imminent rear-end collisions in a high-fidelity driving simulator. Hum Factors: J Hum Factors Ergon Soc 44(2):314–334
Markkula G (2006) A steering wheel reversal rate metric for assessing effects of visual and cog. secondary task load. In: Proceedings of the ITS world congress , London
Markkula G, Benderius O, Wolff K, Wahde M (2012) A review of near-collision driver behavior models. Hum Factors: J Hum Factors Ergon Soc 54(6):1117–1143
McGehee D, Lee J, Rizzo M, Dawson J, Bateman K (2004) Quantitative analysis of steering adaptation on a high performance fixed-base driving simulator. Trans Res Part F: Traffic Psychol Behav 7(3):181–196
Wann JP, Wilkie RM (2004) How do we control high speed steering? In: Vaina LM, Beardsley SA, Rushton SK (eds) Optic flow and beyond. Kluwer Academic Publishers, Dordrecht, pp 371–389
World Health Organization (2009) Global status report on road safety: time for action
This work was supported by VINNOVA (the Swedish Governmental Agency for Innovation Systems; grant number 2009-02766) aswell as by the competence centre ViP (Virtual Prototyping and Assessment by Simulation, www.vipsimulation.se, co-financed by theVINNOVA grant 2007-03083 and the competence centre partners).
The authors wish to thank Lars Eriksson, Anne Bolling, Jonas Andersson Hultgren, and Anders Andersson at the SwedishNational Road and Transport Research Institute, for their invaluable support in the preparation, execution, and analysis phases of this study.